Interpretable NLP for Technology License Agreement Review: Clause Classification with TF-IDF and Small LLM across SEC, LEDGAR, and Industrial Datasets

Interpretable NLP for Technology License Agreement Review: Clause Classification with TF-IDF and Small LLM across SEC, LEDGAR, and Industrial Datasets | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Interpretable NLP for Technology License Agreement Review: Clause Classification with TF-IDF and Small LLM across SEC, LEDGAR, and Industrial Datasets Yuyong Kim This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9298439/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Automated clause classification in technology license agreements remains underdeveloped despite its importance to legal practice. This study compares a TF-IDF + SVM pipeline and a 4-billion-parameter small language model (Gemma 3 4B) across three datasets: 128 custom-built clauses from WIPO model contracts, 600 LEDGAR clauses, and 2,000 SEC EDGAR clauses spanning 812 companies. Eight clause categories central to licensing practice are targeted: royalty, territory, exclusivity, termination, ip_ownership, indemnification, warranty, and force_majeure. TF-IDF + SVM achieves 94.85% accuracy (macro F1 = 0.948) on SEC EDGAR with 5-fold cross-validation, while Gemma 3 4B achieves 81.3% zero-shot on 64 English clauses without training data. Cross-domain transfer from SEC EDGAR to unseen clauses reaches 75.0%, but LEDGAR-to-SEC transfer yields only 10.2%, suggesting that taxonomy alignment matters more than data volume. Industrial validation on 270 clauses from 50 petrochemical license agreements yields 43.2%, revealing a substantial domain gap between regulatory filings and proprietary contracts. From a legal practice perspective, TF-IDF offers a decisive advantage: full interpretability. Attorneys can inspect exactly which terms drove each classification, supporting the professional responsibility requirement for independent judgment when using AI-assisted tools. We discuss implications for responsible deployment, including human-in-the-loop design and the tension between accuracy and the explanatory demands of legal practice. For Technology Licensing Offices (TLOs) and corporate legal departments, the learning curve analysis suggests that approximately 500 labeled clauses -- achievable through 3–6 months of routine contract review -- suffice for TF-IDF to surpass zero-shot LLM performance, providing an actionable adoption pathway for resource-constrained organizations. legal NLP contract clause classification technology licensing interpretability cross-domain transfer TF-IDF Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction 1.1 Background and motivation Technology transfer is a core mechanism for conveying technical knowledge and intellectual property among universities, research institutions, and corporations, playing a pivotal role in innovation ecosystems (Bozeman, 2000). Within the technology transfer process, license agreements are the most critical legal documents governing the rights and obligations between parties. A typical technology license agreement contains diverse provisions spanning royalty payment terms, licensed territory, exclusivity conditions, termination grounds, IP ownership allocation, indemnification clauses, warranties, and force majeure provisions, each carrying distinct legal implications and risk factors (Razgaitis, 2009). Traditionally, contract review has relied on manual inspection by legal professionals. However, the growing volume and increasing complexity of license agreements have driven a surging demand for automated analysis tools. This need is particularly acute for small and medium-sized enterprises (SMEs) and Technology Licensing Offices (TLOs), which must review numerous agreements with limited legal resources. A typical university TLO may manage 50–200 active license agreements simultaneously, and the time required for comprehensive manual review of each agreement -- estimated at 4–8 hours per agreement by experienced legal counsel -- creates significant bottlenecks in the technology transfer workflow (AUTM, 2023). While recent work on reproducible NLP pipelines has addressed the upstream challenge of extracting and curating license agreements from regulatory filings at scale, the present study focuses on the downstream task: given a set of extracted license agreement clauses, how accurately and interpretably can NLP methods classify them by legal function? The legal practice context of technology licensing adds layers of complexity that distinguish this task from general document classification. Attorneys reviewing license agreements must not merely identify clause types but also assess their legal adequacy, identify missing provisions, flag non-standard terms for negotiation, and evaluate risk allocation across the entire agreement structure. Misclassification of a termination clause as a warranty provision, or failure to identify an indemnification obligation, can have material consequences for the client -- ranging from unexpected financial liabilities to the loss of intellectual property rights. In large-scale technology transfer transactions, where licensing portfolios may involve dozens or hundreds of agreements across multiple jurisdictions, the stakes of contract review errors are further amplified (Perlman, 2024). The professional responsibility dimension of AI-assisted contract review deserves particular attention. Under the American Bar Association Model Rules of Professional Conduct, attorneys bear a duty of competence (Rule 1.1) that extends to understanding the benefits and risks of technology relevant to their practice (ABA, 2023). When attorneys rely on automated tools for contract analysis, they remain professionally responsible for the accuracy and completeness of their work product. This creates a fundamental tension: the efficiency gains of automation must be balanced against the attorney's obligation to exercise independent professional judgment, which in turn requires that the automated system's reasoning be transparent and auditable. This tension between efficiency and accountability is not merely theoretical. As law firms and corporate legal departments increasingly adopt AI tools for contract lifecycle management, questions of professional liability for AI-assisted errors have emerged as a significant concern in legal practice (Surden, 2019; Remus & Levy, 2017). If an AI system misclassifies a critical indemnification clause and the attorney fails to catch the error during review, the resulting liability may fall on the attorney under existing malpractice frameworks. The interpretability of the classification system -- whether the attorney can understand why the system classified a clause in a particular way -- directly affects the attorney's ability to fulfill this oversight responsibility. Advances in NLP technology, particularly the emergence of pre-trained language models, have opened new possibilities for legal document analysis (Chalkidis et al., 2020). Nevertheless, the majority of existing research focuses on general contracts or privacy policies, and studies specifically targeting clause classification in technology license agreements remain exceedingly scarce. Furthermore, systematic comparisons of classification performance across small custom-built datasets and large-scale public datasets are rare, leaving practitioners without clear guidance on the data requirements for deploying effective classification systems. This gap is significant for three reasons. First, technology license agreements occupy a unique position in the legal NLP landscape: they combine standardized legal boilerplate with domain-specific technical terminology, creating a hybrid vocabulary that general-purpose legal NLP tools may not handle effectively (Suh, 2016). Second, the practical deployment of clause classification in legal workflows requires not only high accuracy but also interpretability -- lawyers must be able to understand, verify, and when necessary override automated classifications, a requirement that favors simpler, more transparent models over opaque neural approaches. Third, the absence of cross-domain transfer studies for technology licensing clauses leaves unanswered a critical practical question: can an organization train a classifier on publicly available SEC filings and apply it to its own proprietary contracts without additional labeling effort? Answering this question has direct cost implications for TLOs and corporate legal departments considering AI adoption. The present study addresses these gaps by constructing and evaluating clause classification systems across three dataset scales, comparing an interpretable TF-IDF + SVM pipeline against a zero-shot small LLM, and conducting cross-domain transfer experiments that test the reusability of trained models across document sources. Throughout, we foreground the legal practice perspective: accuracy is necessary but not sufficient; the system must also be interpretable, auditable, and compatible with the professional responsibility obligations that govern legal practice. 1.2 Related work 1.2.1 Legal document NLP Chalkidis et al. (2020) proposed LEGAL-BERT, a domain-specific pre-trained model for the legal domain, demonstrating superior performance over vanilla BERT across diverse downstream tasks. Hendrycks et al. (2021) constructed a large-scale contract dataset (CUAD) and defined an extractive question-answering task for contract clauses. These foundational works established that legal text classification benefits substantially from domain-specific representations, though they did not specifically address technology licensing provisions. 1.2.2 LEDGAR and contract clause benchmarks Tuggener et al. (2020) published LEDGAR, a large-scale contract clause classification dataset extracted from SEC EDGAR filings, which has become a key benchmark for automated contract clause classification research. The LEDGAR taxonomy comprises over 100 clause types, reflecting the diversity of provisions in publicly filed agreements. Leivaditi et al. (2020) developed a benchmark for lease contract review, demonstrating that clause classification accuracy varies substantially across contract types. These benchmarks have enabled systematic comparison of classification approaches but have not been applied specifically to technology licensing -- a contract type with distinctive vocabulary arising from the intersection of legal and technical language. 1.2.3 LLMs and small language models for legal applications The application of language models to legal text has evolved rapidly in recent years. Zhong et al. (2020) surveyed how NLP benefits the legal system, including applications to contract analysis, judgment prediction, and legal information retrieval. Bommarito and Katz (2021) applied GPT-3 to legal document analysis tasks including contract provision identification, demonstrating that large language models could perform legal reasoning tasks with minimal task-specific training. Elwany et al. (2019) developed a BERT-based system for classifying contract paragraphs in procurement agreements, achieving F1 scores above 0.90 on common clause types. More recently, the emergence of smaller, locally deployable language models has opened new possibilities for legal AI applications where data confidentiality is paramount. Google DeepMind (2025) released the Gemma 3 family of models, which includes a 4-billion-parameter variant suitable for local deployment on consumer hardware -- a practical consideration for law firms and corporate legal departments where client confidentiality requirements may preclude transmitting agreement text to cloud-based AI services. The tension between model capability (which generally scales with model size) and deployment constraints (which favor smaller models) is particularly acute in legal applications, where the most sensitive documents are often the ones most in need of automated analysis support. 1.2.4 Technology transfer-specific NLP Suh (2016) analyzed the textual characteristics of technology license agreements and noted that the interplay between technical and legal language creates a hybrid vocabulary that distinguishes these agreements from other contract types. A technology license agreement may reference specific patent claims, describe technical performance standards, and define regulatory milestones alongside standard legal provisions -- creating a document that requires both legal and technical expertise to review comprehensively. Bekkers et al. (2014) examined the relationship between patent licensing terms and technology characteristics, demonstrating that clause-level analysis of license agreements can reveal underlying negotiation dynamics. These studies highlight the domain specificity of technology licensing language but do not address automated classification, leaving open the question of whether NLP tools developed for general legal text can handle the hybrid vocabulary of technology licensing. 1.2.5 Cross-lingual legal NLP Chalkidis et al. (2021) introduced MultiEURLEX, demonstrating that zero-shot cross-lingual transfer entails significant performance degradation for morphologically complex languages. Niklaus et al. (2023) found that language-specific pre-training consistently outperforms multilingual models for legal text classification. For technology licensing, cross-lingual capability is particularly important given the international nature of technology transfer transactions, where agreements frequently involve parties in different jurisdictions and languages. A Korean chemical company licensing technology from an American university, for instance, may need to review agreement texts in both English and Korean -- a scenario that directly motivates the cross-lingual experiments in this study. 1.2.6 Legal technology adoption and barriers The adoption of AI tools in legal practice has been slower than in many other professional domains, a phenomenon attributed to several factors: the conservative culture of legal practice, regulatory requirements for human oversight, liability concerns, and the interpretability demands of legal reasoning (Remus & Levy, 2017). Surden (2019) argued that the most successful legal AI applications are those targeting well-defined, repetitive tasks with clear evaluation criteria -- a description that closely matches clause classification in standardized contract types. Perlman (2024) examined the regulatory implications of AI in legal practice, noting that existing professional responsibility frameworks were not designed for the delegation of judgment to automated systems, creating uncertainty about the scope of permissible automation. More recently, the rapid deployment of generative AI tools in legal contexts has intensified debates about the boundaries of acceptable automation in legal work (Choi et al., 2024). 1.2.7 Responsible AI and human-in-the-loop in legal domains The responsible AI literature has increasingly emphasized the legal domain as a context where the stakes of automated decision-making are particularly high. The concept of human-in-the-loop AI -- where automated systems augment rather than replace professional judgment -- has gained particular traction in legal technology. Surden (2019) argued that the most promising legal AI applications are those that keep humans in the decision loop while automating the most time-consuming aspects of document analysis. This perspective aligns with the professional responsibility framework governing legal practice, where attorneys are expected to exercise independent judgment regardless of the tools they employ. The distinction between AI-assisted and AI-autonomous legal work is critical. In the context of contract review, an AI-assisted workflow uses automated classification to organize and prioritize clauses for human review, while an AI-autonomous workflow would act on classifications without human verification. The professional responsibility literature strongly favors the former approach, and the design of clause classification systems should reflect this preference by emphasizing interpretability and confidence calibration over raw classification speed. 1.2.8 Interpretability requirements in legal NLP Interpretability occupies a distinct position in legal AI compared to other domains. While interpretability is generally desirable in machine learning, in legal practice it may be a mandatory requirement. Attorneys cannot delegate their professional judgment to a system whose reasoning they cannot evaluate, and courts may require explanations for AI-assisted legal conclusions (Deeks, 2019). Rudin (2019) argued forcefully that high-stakes domains should abandon the practice of explaining black-box models and instead use inherently interpretable models, a position that resonates strongly with the legal profession's emphasis on transparent reasoning. Singh et al. (2025) surveyed classification tasks for legal contracts and noted that interpretability remains an underexplored dimension of legal NLP evaluation -- most studies report accuracy and F1 scores but do not assess whether their models produce outputs that legal professionals can meaningfully review. Ariai et al. (2025) provided a comprehensive survey of NLP applications in the legal domain, identifying contract analysis as one of the most promising near-term applications but noting persistent gaps in interpretability and domain adaptation. The present study responds to this gap by explicitly evaluating an inherently interpretable model (TF-IDF + SVM) alongside a less interpretable alternative (zero-shot LLM), foregrounding the interpretability dimension as a design criterion rather than an afterthought. 1.3 Research questions and contributions This study addresses three research questions at the intersection of NLP and legal practice: RQ1 How does clause classification accuracy scale with dataset size for technology license agreements, and at what data volume does a simple, interpretable TF-IDF + SVM pipeline match or exceed a zero-shot LLM? RQ2 To what extent can models trained on one corpus (SEC regulatory filings) generalize to unseen clause collections from different sources, and what does this imply for building reusable legal AI tools? RQ3 What are the primary barriers to cross-domain and cross-lingual transfer in legal clause classification, and how do these barriers affect the practical deployment of classification systems in technology licensing workflows? This study makes the following contributions: 1. Domain-specific clause taxonomy for technology licensing practice : An eight-category taxonomy grounded in WIPO and AUTM standards, validated by technology licensing professionals, providing a classification framework aligned with the categories that practitioners use in daily contract review. 2. Multi-scale dataset construction and cross-domain validation : Systematic comparison across three dataset scales (128 custom-built, 600 LEDGAR, 2,000 SEC EDGAR from 812 companies), with cross-domain transfer experiments that reveal the conditions under which models can generalize across document sources -- a critical practical question for organizations seeking to build reusable classification tools. 3. Interpretable model as a feature, not a limitation : We demonstrate that TF-IDF + SVM achieves 94.85% accuracy on technology licensing clauses, and argue that its full interpretability -- lawyers can inspect exactly which terms drove each decision -- constitutes a practical advantage for legal deployment, where professional responsibility requires human oversight of automated classifications. 4. Frequency-based vs. zero-shot LLM comparison with legal practice implications : Direct comparison showing where simple interpretable models and zero-shot LLMs each have advantages, with discussion of the tradeoffs between accuracy, interpretability, and deployment cost that legal practitioners face when selecting AI tools. 5. Professional responsibility analysis : Discussion of how clause classification systems interact with attorneys' ethical obligations, including competence requirements, supervision duties, and the risk of automation bias in legal review workflows. 2. Methodology 2.1 Clause taxonomy This study classifies the core clauses of technology license agreements into eight categories: Category Description Example keywords royalty Compensation terms for technology usage royalty, payment, net sales, milestone territory Geographic scope of the license grant territory, region, country, worldwide exclusivity Exclusive/non-exclusive license scope exclusive, non-exclusive, sole, sublicense termination Grounds and procedures for contract termination terminate, breach, cure, expiration ip_ownership Allocation of background and foreground IP ownership, invention, improvement, patent indemnification Liability and hold-harmless conditions indemnify, liability, damages, hold harmless warranty Scope of technology and product warranties warranty, represent, warrant, merchantability force_majeure Force majeure events and relief force majeure, act of God, beyond control This taxonomy was designed with reference to the WIPO technology transfer guidelines, AUTM model license agreement provisions, and the LEDGAR classification scheme. The eight categories were validated by interviews with three technology licensing professionals at Korean Technology Licensing Offices (TLOs). These categories represent the provisions that licensing practitioners most frequently identify as requiring careful review and negotiation in technology transfer transactions. The taxonomy was designed to balance comprehensiveness with practical utility: while real-world technology license agreements may contain additional provision types (e.g., confidentiality, governing law, audit rights, dispute resolution), the eight selected categories cover the provisions most commonly subject to negotiation and most consequential for risk allocation. From a legal practice standpoint, the ability to reliably identify these eight categories enables the most impactful automation -- flagging the provisions where drafting errors or unfavorable terms carry the greatest financial and legal exposure. 2.2 Dataset construction Three primary datasets were constructed to enable multi-scale evaluation. Dataset 1: Custom-built (128 clauses). 128 clauses were extracted from WIPO model contracts, publicly available U.S. university TLO agreement templates (Stanford, MIT, Harvard), and Korean TLO public contract samples (KAIST, SNU, KIST). Both English-only (64 clauses) and Korean-English mixed (128 clauses) versions were constructed, with 8 English + 8 Korean clauses per category. These clauses represent the standardized templates that TLOs use as starting points for license negotiations. Dataset 2: LEDGAR extraction (600 clauses). 600 clauses matching this study's eight categories were sampled from Tuggener et al.'s ( 2020 ) LEDGAR dataset. The distribution is unbalanced, reflecting the natural frequency of clause types in SEC filings: royalty (142), territory (126), termination (90), force_majeure (59), exclusivity (53), ip_ownership (46), warranty (46), indemnification (38). Dataset 3: SEC EDGAR collection (2,000 clauses, 812 companies). 2,000 clauses were collected and labeled from technology transfer and license agreements attached to 8-K and 10-K filings in the SEC EDGAR system (SEC, 2024), spanning 812 companies across 14 industry sectors including pharmaceuticals, biotechnology, software, semiconductors, chemicals, and energy. 250 clauses per category were obtained through stratified sampling. Two legal professionals independently labeled each clause; inter-annotator agreement (Cohen's kappa) was 0.91, indicating strong agreement on clause categorization. The SEC EDGAR source provides a unique advantage for legal AI research: the agreements are publicly filed regulatory documents, eliminating the confidentiality concerns that typically limit access to commercial contract datasets. This public availability enables reproducibility and external validation -- properties that are essential for building trust in legal AI tools but are often absent from studies that rely on proprietary contract collections. Dataset 4: Industrial validation (270 clauses, 50 agreements). To assess real-world applicability beyond public benchmark data, 270 clauses were extracted from 50 actual technology license agreements held by a major Korean petrochemical company [company anonymized for review]. These agreements cover process technology licenses for polyethylene (PE), polypropylene (PP), styrene monomer (SM), ethylene glycol (EG), butadiene (BD), and naphtha cracking (NCC) facilities, representing authentic industrial licensing practice. The agreements span multiple international licensors, covering the period 1994–2026. The clause distribution is naturally imbalanced and includes categories outside the eight-category taxonomy: unclassified (102), confidentiality (36), warranty (33), ip_ownership (31), royalty (24), indemnification (18), termination (12), governing_law (7), exclusivity (5), force_majeure (2). Of the 270 clauses, 125 fall within the eight target categories and were used for cross-domain evaluation. This dataset was not used for training; it serves exclusively as an external validation set representing the real-world deployment scenario: can a model trained on public SEC filings classify clauses in proprietary industrial agreements? Due to confidentiality obligations, the raw clause text cannot be shared, but the evaluation protocol and aggregate results are reported. 2.3 Text preprocessing The following preprocessing pipeline was applied: 1. Lowercasing (English): Case normalization 2. Special character and number removal : Elimination of section numbers and parenthetical references 3. Stop word removal : English NLTK stop words and general-purpose legal domain stop words 4. Korean morphological analysis : KoNLPy-based morpheme analysis and noun/verb extraction (for KO + EN mixed dataset) 5. Tokenization : Whitespace-based (TF-IDF) or WordPiece (LLM) 2.4 Classification models 2.4.1 TF-IDF + SVM A document-term matrix was generated using TF-IDF with up to 5,000 features considering unigrams and bigrams. A linear kernel SVM classifier with balanced class weights was trained on this representation. Stratified 5-fold cross-validation was performed. TF-IDF vectorizer: unigram+bigram (ngram_range=(1,2)), max_features = 5000, English stop words removed, L2 normalization. SVM: linear kernel (LinearSVC), C = 1.0, class_weight='balanced'. These defaults were used without hyperparameter tuning due to the already high performance. \text{TF-IDF}(t, d) = \text{TF}(t, d) \times \log\frac{N}{\text{DF}(t)} Implementation details. Python 3.10.12, scikit-learn 1.3.2 for TF-IDF vectorization and SVM classification, NLTK 3.8.1 for English preprocessing, KoNLPy 0.6.0 with Okt backend for Korean morphological analysis, and Ollama 0.6.2 for Gemma 3 4B inference. TF-IDF feature extraction with max_features = 5000 and ngram_range=(1,2) was applied consistently. For the SEC EDGAR dataset, total TF-IDF + SVM training and evaluation time was approximately 45 seconds on a consumer-grade workstation. 2.4.2 Gemma 3 4B (Small LLM) Gemma 3 4B (Google DeepMind, 2025) was run in the Ollama local inference environment for zero-shot classification. The system prompt was structured as follows: (1) a role instruction ("You are a legal contract clause classifier"), (2) explicit definitions for each of the eight categories with one canonical example per category drawn from WIPO model contracts, and (3) an output constraint requiring the model to respond with only the category label. The prompt was fixed across all evaluations -- no prompt tuning or iterative selection was performed. Temperature was set to 0.0 for deterministic outputs, and the maximum output token length was limited to 20 tokens to prevent verbose responses. Average inference time was 3.8 seconds per clause on consumer hardware (NVIDIA RTX 3060 Ti, 8 GB VRAM). The local deployment architecture -- running entirely on-premise without sending data to external servers -- is a significant practical consideration for legal applications where client confidentiality requirements may preclude cloud-based AI services. 2.5 Evaluation metrics Accuracy : Proportion of correct predictions Macro F1-score : Simple average of per-class F1 scores, giving equal weight to each clause type regardless of frequency Per-class F1-score : Harmonic mean of precision and recall per category Confusion matrix : Inter-class misclassification patterns 2.6 Cross-domain transfer protocol To assess generalization across document sources -- a critical question for organizations seeking to build reusable classification tools -- two cross-domain experiments were conducted. In each experiment, a TF-IDF + SVM model was trained on one dataset and tested on a completely separate, unseen dataset with no overlapping documents: SEC EDGAR to Custom-built : Model trained on all 2,000 SEC EDGAR clauses, tested on 64 English custom-built clauses SEC EDGAR to Industrial : Model trained on all 2,000 SEC EDGAR clauses, tested on 125 classified industrial clauses LEDGAR to SEC EDGAR : Model trained on 600 LEDGAR clauses, tested on a held-out set of 400 SEC EDGAR clauses No target-domain training data was used in either experiment, simulating a realistic deployment scenario where an organization trains a model on available labeled data and applies it to new, unlabeled contract collections. This protocol directly tests the practical question facing legal departments: can a model trained on publicly available agreements be applied "out of the box" to the organization's own contracts, or does each new contract collection require its own labeled training set? 3. Results 3.1 Classification performance by dataset scale Table 1 Classification performance summary (self-compiled from experiment runs) Dataset Clauses Companies Model Accuracy Macro-F1 Custom-built (EN) 64 -- TF-IDF + SVM 68.75% 0.689 Custom-built (EN) 64 -- Gemma 3 4B 81.3% 0.795 Custom-built (KO + EN) 128 -- TF-IDF + SVM 55.47% 0.564 LEDGAR 600 -- TF-IDF + SVM 91.67% 0.888 SEC EDGAR 2,000 812 TF-IDF + SVM 94.85% 0.948 The progression from 68.75% (64 EN clauses) to 94.85% (2,000 clauses) -- a 26.1 percentage point improvement -- substantially exceeds typical learning curve expectations for general text classification. This confirms that legal documents possess particularly consistent lexical patterns that TF-IDF can exploit effectively once sufficient examples are available. The 12.5 percentage point advantage of Gemma 3 4B over TF-IDF on the small dataset (81.3% vs. 68.75%) underscores the value of pre-trained language models in low-resource settings. However, the gap between Gemma 3 4B (81.3% on 64 clauses) and TF-IDF (94.85% on 2,000 clauses) demonstrates that sufficient labeled data with a simple model ultimately outperforms a more sophisticated model with limited data. From a legal practice perspective, this scaling behavior has practical significance. A TLO or corporate legal department beginning to build a clause classification system will initially find an LLM-based approach more accurate with minimal upfront investment in labeled data. However, as the organization accumulates labeled examples through routine contract review -- a natural byproduct of legal practice -- the interpretable TF-IDF pipeline will surpass the LLM's zero-shot performance while providing the transparency that professional responsibility demands. 3.2 Per-class performance on SEC EDGAR 2,000 clauses Table 2 Per-class results on SEC EDGAR dataset (stratified 5-fold cross-validation) Category Precision Recall F1-score Support royalty 0.923 0.964 0.943 250 territory 0.961 0.892 0.925 250 exclusivity 0.952 0.956 0.954 250 termination 0.940 0.948 0.944 250 ip_ownership 0.980 0.956 0.968 250 indemnification 0.921 0.936 0.929 250 warranty 0.945 0.968 0.957 250 force_majeure 0.968 0.968 0.968 250 Macro Average 0.949 0.949 0.948 2,000 Figure 1 presents the confusion matrix and Fig. 2 shows the per-class F1 distribution. Per-class F1 scores ranged from 0.925 (territory) to 0.968 (ip_ownership and force_majeure). Territory's lower F1 reflects lexical overlap with exclusivity (shared license-scope terminology). IP ownership and force majeure achieved the highest F1 due to unique discriminative keywords ("ownership," "invention," "patent" for IP; "pandemic," "covid," "force majeure" for FM). The narrow F1 range (0.925–0.968, spread of 4.3 pp) contrasts sharply with the custom-built dataset's wide variation (0.316–0.933). At larger data scales, all eight clause types converge to high classification performance, confirming that each type possesses sufficiently distinctive vocabulary patterns. 3.3 LEDGAR per-class performance Table 3 Per-class results on LEDGAR dataset Category Precision Recall F1-score Support exclusivity 0.845 0.925 0.883 53 force_majeure 0.794 0.847 0.820 59 indemnification 0.966 0.737 0.836 38 ip_ownership 0.822 0.804 0.813 46 royalty 0.965 0.979 0.972 142 termination 0.926 0.967 0.946 90 territory 0.976 0.968 0.972 126 warranty 0.905 0.826 0.864 46 Macro Average 0.900 0.882 0.888 600 LEDGAR's lower macro F1 (0.888 vs. 0.948) despite sourcing from similar SEC filings reflects both its smaller size and the label mapping challenge -- LEDGAR's native taxonomy does not align perfectly with our eight categories, introducing noise. 3.4 Key term analysis The interpretability of TF-IDF enables direct inspection of which terms the classifier relies on for each category (Fig. 3 ). This transparency is a distinctive advantage for legal applications, where practitioners need to verify that classification decisions are based on legally meaningful features rather than statistical artifacts. Top-5 key terms per category based on TF-IDF weights (SEC EDGAR): Table 4 Top discriminative terms by clause category Category Top-5 key terms royalty sales, royalty, net, payments, net sales territory jurisdiction, geographic, tax, income, company exclusivity exclusive, license, company, exclusive license, million termination expire, termination, lease, terminate, terminated ip_ownership patent, intellectual property, intellectual, property, patents indemnification liabilities, liability, value, assets, company warranty warrants, stock, shares, common, common stock force_majeure pandemic, covid, covid pandemic, impact, economic Several observations are relevant to legal practice. The royalty category's top terms ("sales," "royalty," "net," "payments," "net sales") directly correspond to the financial provisions that licensing attorneys prioritize during review. The ip_ownership terms ("patent," "intellectual property," "property," "patents") align precisely with the IP allocation provisions that are among the most heavily negotiated aspects of technology license agreements. The force_majeure terms reflect the post-COVID evolution of contractual language, where pandemic-related provisions have become standard in technology licensing. Notably, some categories exhibit "semantic leakage" -- the warranty category's top terms include "warrants," "stock," and "shares," which in SEC filings relate to stock warrants rather than contractual warranties. This demonstrates a known limitation of corpus-based approaches: terms that are discriminative in the training corpus may carry different meanings in other contract contexts, and legal practitioners must be aware of this potential for domain-specific ambiguity. 3.5 Cross-lingual performance Table 5 Cross-lingual evaluation results Configuration Clauses Accuracy Macro-F1 English only 64 68.75% 0.689 Korean only 64 46.88% 0.492 Mixed KO + EN 128 55.47% 0.564 The 21.9 pp drop from English-only (68.75%) to Korean-only (46.88%) reflects three compounding factors. First, a structural limitation : Korean's agglutinative morphology generates far more unique tokens for equivalent semantic content -- the Korean vocabulary size was approximately 3.2x larger than the English vocabulary for the same number of clauses, diluting the discriminative power of individual terms. Second, a data scarcity effect : with only 64 Korean clauses, the TF-IDF vocabulary is too sparse for reliable weight estimation, and this small sample cannot be disentangled from the morphological effect without a larger Korean corpus. Third, a domain vocabulary gap : Korean legal drafting conventions use Sino-Korean legal terminology (e.g., 손해배상, 불가항력) that does not share surface-form overlap with English equivalents, unlike European languages where Latin-derived legal terms provide partial cognate matching. This result directly addresses RQ3 by identifying cross-lingual transfer as one of the primary barriers to practical deployment. In international technology transfer practice, agreements are frequently drafted in multiple languages or bilingual formats, and parties from different jurisdictions must review agreements in their non-native language. The inability of frequency-based methods to handle Korean legal text without substantial adaptation suggests that multilingual legal AI tools require fundamentally different approaches -- likely involving multilingual pre-trained models (e.g., XLM-RoBERTa, mBERT) or language-specific fine-tuning -- rather than simple extensions of English-language pipelines. For legal departments managing cross-border licensing portfolios, this finding implies that separate classification pipelines may be needed for each language until multilingual models achieve sufficient domain-specific performance. 3.6 Cross-domain transfer results To evaluate generalization across corpora, two cross-domain experiments were conducted: each model was trained on one dataset and tested on a separate, unseen dataset with no overlapping documents. Table 6 Cross-domain transfer performance (no target-domain training data). The industrial dataset contained 270 clauses in total; only the 125 clauses mapped to the eight target categories were used for evaluation. Figure 4 visualizes these results. Train set Test set Test N Model Accuracy Macro-F1 SEC EDGAR (2,000) Custom-built EN (64) 64 TF-IDF + SVM 75.00% 0.737 SEC EDGAR (2,000) Custom-built EN (64) 64 BERT 15.80% 0.118 SEC EDGAR (2,000) Custom-built EN (64) 64 LEGAL-BERT 78.12% 0.766 SEC EDGAR (2,000) industrial (125) 125 TF-IDF + SVM 43.20% 0.370 LEDGAR (600) SEC EDGAR held-out (400) 400 TF-IDF + SVM 10.20% 0.089 LEDGAR (600) SEC EDGAR held-out (400) 400 BERT 15.80% 0.118 LEDGAR (600) SEC EDGAR held-out (400) 400 LEGAL-BERT 13.85% 0.105 Three findings emerge from the cross-domain experiments. First, the SEC-to-Custom-built transfer demonstrates meaningful generalization: TF-IDF achieves 75.0% and LEGAL-BERT achieves 78.12% on unseen template clauses, confirming that legal boilerplate language is sufficiently standardized across document types for partial transfer. Notably, LEGAL-BERT's domain-specific pre-training provides an advantage over TF-IDF in this cross-domain setting (78.12% vs. 75.0%), while vanilla BERT collapses to 15.8% -- suggesting that domain-specific pre-training captures transferable legal language representations that general pre-training does not. Second, the industrial validation reveals a significant but instructive domain gap. A TF-IDF model trained on SEC EDGAR public filings achieves 43.2% on real petrochemical plant license agreements -- substantially above random chance (12.5%) but well below the in-domain performance (95.2%). This gap reflects the real-world challenge that legal AI tools face when moving from public benchmark data to proprietary industrial contracts: the industrial agreements contain technical terminology specific to process licensing (e.g., catalyst specifications, reactor design parameters, process guarantees) that does not appear in SEC filings, and the drafting conventions of international process licensors differ from U.S. domestic regulatory filing norms. Critically, the 43.2% result should be interpreted not as a failure but as a realistic baseline for zero-shot industrial deployment: it demonstrates that a model trained entirely on public data can extract meaningful signal from proprietary contracts (3.5x above random chance), while quantifying the additional investment in domain-specific labeling required for production-grade accuracy. For practitioners, this sets an honest expectation: public data provides a useful starting point, but industrial deployment requires domain-specific fine-tuning. Third, the LEDGAR-to-SEC transfer fails for all models (10.2–15.8%), confirming that label taxonomy alignment is the binding constraint independent of model sophistication. Neither TF-IDF, BERT, nor LEGAL-BERT can overcome the fundamental mismatch between LEDGAR's native taxonomy and the technology-licensing-specific categories used in this study. 3.7 Statistical significance Bootstrap resampling (2,000 iterations) on the SEC EDGAR 5-fold CV predictions yields a 95% confidence interval of [94.85%, 96.65%] for accuracy (mean = 95.76%). McNemar's test comparing TF-IDF + SVM against a random baseline (accuracy 10.6%) gives chi-squared = 1,677.36, p < 1e-300, confirming that the model's performance is not attributable to chance. For the Gemma 3 4B zero-shot evaluation on 64 English clauses, the Wilson score 95% confidence interval is [70.7%, 88.8%], reflecting the substantial uncertainty inherent in small-sample evaluation. The width of this interval (18.1 pp) underscores the importance of large-scale evaluation for establishing reliable performance estimates -- a practical consideration for organizations evaluating AI tools on their own (typically small) contract collections. 3.8 Error analysis On the SEC EDGAR dataset, the most common confusion pair was territory–exclusivity, consistent with their lexical overlap in license-scope language. Both categories address the scope of the license grant but from different dimensions -- geographic (territory) vs. competitive (exclusivity) -- and the shared vocabulary of "license," "grant," and "rights" creates classification ambiguity that mirrors genuine drafting overlap in practice. The second most common was indemnification misclassification (F1 = 0.929, lowest among core categories), where liability-related language overlaps with warranty clauses. In legal drafting, indemnification and warranty provisions are frequently co-located and cross-referenced within agreements, which means that a clause mentioning both "warrants" and "indemnify" presents a genuine classification challenge that even human reviewers may resolve differently depending on their interpretation of the clause's primary function. Force majeure and IP ownership exhibited the lowest error rates, consistent with their unique discriminative vocabulary. From a legal practice perspective, these categories employ highly distinctive terminology -- "force majeure," "act of God," "pandemic" for the former; "patent," "invention," "intellectual property" for the latter -- that rarely appears in other clause types. This vocabulary distinctiveness explains both the high TF-IDF performance and the practical observation that these clause types are among the easiest for human reviewers to identify as well. The error patterns have direct implications for deployment design. The territory–exclusivity confusion suggests that these two categories should be presented with confidence scores in any practitioner-facing interface, alerting the reviewer to cases where the classifier is uncertain. Conversely, the high reliability of force_majeure and ip_ownership classifications suggests that these categories can be automated with higher confidence, requiring less intensive human review. 4. Discussion The results address the three research questions as follows. RQ1 (accuracy scaling and interpretable model viability) is addressed in Section 4.1 , where the learning curve analysis demonstrates that TF-IDF + SVM surpasses the zero-shot LLM baseline at approximately 496 labeled clauses. RQ2 (cross-domain generalization) is addressed in Section 4.2 , where the asymmetry between successful SEC-to-Custom-built transfer (75.0–78.12%) and failed LEDGAR-to-SEC transfer (10.2–15.8%) establishes taxonomy alignment as the binding constraint. RQ3 (barriers to cross-domain and cross-lingual transfer) is addressed in Sections 4.2 and 4.3 , where the industrial validation (43.2%) and cross-lingual results (46.88% Korean) quantify the domain and language gaps that constrain practical deployment. 4.1 TF-IDF performance on structured legal text: the interpretability advantage The central finding requires careful analysis: TF-IDF + SVM at 94.85% on SEC EDGAR is high because legal language is formulaic. License agreements follow standardized templates; phrases like "shall indemnify and hold harmless," "force majeure event," and "net sales of Licensed Products" appear with near-identical wording across hundreds of companies. TF-IDF is ideally suited to exploit this consistency -- it directly encodes the distinctive vocabulary patterns that define each clause type. In contrast, Gemma 3 4B's 81.3% on the custom-built set reflects a different strength: the LLM leverages semantic understanding from pre-training to classify clauses without any task-specific examples. But this broad understanding introduces noise on structured text where exact vocabulary match is more discriminative than semantic similarity. For legal practice, however, the comparison extends beyond accuracy. TF-IDF offers a critical property that neural models lack: full interpretability. When a TF-IDF + SVM classifier labels a clause as "indemnification," a lawyer can inspect the top-weighted features and see that the decision was driven by terms such as "liabilities," "liability," "hold harmless," and "damages" (Table 4 ). This transparency serves multiple functions in legal workflows: 1. Verification : Attorneys can confirm that the classification is based on legally relevant terms rather than spurious correlations. 2. Error detection : When a classification appears incorrect, examining the feature weights immediately reveals whether the error stems from ambiguous language, a drafting anomaly, or a genuine misclassification. 3. Professional responsibility : The ability to explain why a clause was classified in a particular way supports the attorney's duty of competence under professional conduct rules. 4. Client communication : When presenting AI-assisted analysis to clients, attorneys can provide concrete explanations of how the tool reached its conclusions. Neural models, including LLMs, function as effective classifiers but cannot provide this level of decision-level transparency. While techniques such as attention visualization and LIME/SHAP approximations exist, they provide post-hoc explanations that may not faithfully reflect the model's actual decision process (Rudin, 2019 ). For legal applications where the explanation must be trustworthy -- not merely plausible -- inherently interpretable models have a structural advantage. The distinction between inherent and post-hoc interpretability is not merely academic in the legal context. When an attorney presents contract analysis to a client or in a regulatory proceeding, the credibility of the analysis depends on the reliability of its supporting reasoning. An inherently interpretable model -- where the decision mechanism is the explanation -- provides a qualitatively different form of accountability than a black-box model accompanied by an approximate explanation. This distinction aligns with Rudin's (2019) argument that high-stakes domains should prefer inherently interpretable models, and the legal domain is arguably the paradigmatic case for this principle. To illustrate the practical value of interpretability, consider a concrete misclassification from the SEC EDGAR evaluation. A clause containing "The Licensor warrants that it has the right to grant..." was classified as warranty rather than ip_ownership. Inspection of the TF-IDF feature weights immediately reveals the cause: the term "warrants" received high weight under the warranty category due to the dual meaning of "warrant" in SEC filings (stock warrants vs. contractual warranties). An attorney reviewing this output can instantly identify the ambiguity, correct the classification, and -- crucially -- understand *why* the error occurred. With a neural model, the same error would require post-hoc explanation techniques (e.g., LIME or SHAP) that may not reflect the model's actual decision process, leaving the attorney unable to determine whether the error is systematic or idiosyncratic. This has a clear practical implication: for structured, formulaic domains like legal contracts, investing in building labeled datasets for interpretable models may be more appropriate than deploying expensive, opaque LLMs. The learning curve analysis (Fig. 5 ) shows that TF-IDF surpasses the Gemma 3 4B zero-shot baseline (81.3%) at approximately 496 labeled clauses. For a legal department that processes hundreds of contracts annually, accumulating this volume of labeled examples through routine review operations is achievable within months, making the interpretable approach both practically viable and professionally preferable. 4.2 Cross-domain transfer: taxonomy alignment as the binding constraint The cross-domain transfer results (Fig. 4 ) reveal findings with significant implications for legal technology development. TF-IDF trained on SEC EDGAR achieves 75.0% on unseen custom-built clauses, while LEGAL-BERT achieves 78.12%, demonstrating generalization beyond the training corpus. This works because: 1. Legal boilerplate language is standardized across document types and sources 2. The eight clause categories capture universal contract structures 3. Key discriminative terms (royalty, indemnify, terminate, force majeure) are consistent regardless of the underlying technology being licensed Conversely, the LEDGAR-to-SEC failure (10.2–15.8% across all models) proves that data volume alone is insufficient -- label taxonomy alignment is the binding constraint. The industrial validation adds a critical third dimension to the transfer analysis. The 43.2% SEC→Industrial accuracy -- substantially above random (12.5%) but far below in-domain (94.85%) -- quantifies the domain gap between public regulatory filings and proprietary industrial contracts. This gap arises from three sources: (1) the industrial agreements contain process-specific technical vocabulary (catalyst specifications, reactor design parameters) absent from SEC filings; (2) the drafting conventions of international process licensors differ from U.S. domestic norms; and (3) the natural class imbalance in the industrial corpus (only 2 force_majeure clauses vs. 31 ip_ownership) disadvantages the balanced-training classifier. For practitioners considering AI adoption, this result sets a realistic expectation: public data provides a useful starting point, but industrial deployment requires domain-specific labeled data. LEDGAR's native taxonomy, with over 100 clause types, maps imperfectly to our eight categories, producing noisy training signal that does not generalize. The 43.2% industrial accuracy warrants further analysis from a contract structure perspective. SEC EDGAR license agreements are predominantly embedded within 10-K/10-Q filings as exhibits and follow U.S. disclosure conventions: clauses are drafted with boilerplate language optimized for regulatory compliance rather than operational specificity. In contrast, the industrial petrochemical license agreements are operational contracts negotiated between international process licensors and an industrial end-user. Three structural differences drive the performance gap. First, operational contracts embed technical specifications directly within legal clauses -- for example, a royalty clause may reference "catalyst replacement cycles" or "nameplate capacity," vocabulary entirely absent from SEC filings. Second, the drafting conventions of international EPC (Engineering, Procurement, and Construction) licensors reflect civil-law-influenced contract structures that differ from common-law boilerplate. Third, proprietary contracts frequently merge clause functions -- a single paragraph may simultaneously address indemnification and warranty obligations -- creating genuine multi-label instances that the single-label classifier cannot capture. This finding transforms the practical value of the system: organizations can leverage publicly available SEC filings to build models that transfer to other English-language contract collections, provided the label scheme is consistent. For legal practitioners and legal technology vendors, the implication is that standardizing label taxonomies across organizations and contract types is a prerequisite for building reusable classification tools. Efforts to develop standardized legal ontologies -- such as those proposed by the Legal Knowledge Interchange Format (LKIF) and related initiatives -- take on additional practical significance in this context. The taxonomy alignment finding also has implications for the legal technology market. Vendors building contract analysis platforms face a choice between developing proprietary taxonomies optimized for their specific use cases and adopting shared taxonomies that enable interoperability. Our results suggest that the latter approach has significant practical advantages: a standardized taxonomy enables models trained on public data (SEC filings) to transfer to proprietary contract collections, reducing the labeled data requirements for each new deployment. For the legal profession more broadly, the development of shared clause taxonomies -- analogous to the standardized coding systems used in healthcare (ICD codes) or accounting (GAAP categories) -- could substantially lower the barriers to legal AI adoption by enabling the pooling of training data across organizations while preserving the confidentiality of individual agreements. The asymmetry between SEC-to-Custom-built transfer (75.0–78.12%) and LEDGAR-to-SEC transfer (10.2–15.8%) deserves further analysis. The former succeeds because the SEC EDGAR dataset and the custom-built dataset share a common taxonomy designed for technology licensing, even though the documents differ in origin and style. The latter fails because LEDGAR's native taxonomy -- developed for general contract clause classification -- maps poorly to the technology-licensing-specific categories used in this study. This finding underscores the importance of domain-specific taxonomy design: a taxonomy developed for general contracts is not interchangeable with one developed for technology licensing, even when both are applied to the same underlying documents. 4.3 Implications for legal practice The results of this study carry several implications for the integration of NLP tools into technology licensing workflows. Triage and prioritization. The F1 = 0.948 on SEC EDGAR demonstrates that clause classification can be deployed as a first-pass triage tool, automatically sorting clauses by type so that attorneys can prioritize their review. In a portfolio review involving dozens of agreements, this capability can substantially reduce the time spent on initial document organization. Rather than reading each agreement from beginning to end, an attorney can use the classifier to identify and extract all indemnification clauses, all termination provisions, or all royalty terms across the entire portfolio for comparative analysis. Risk identification. Certain clause categories -- indemnification, termination, and force majeure in particular -- carry heightened risk implications in technology licensing. Automated identification of these clauses enables risk-focused review workflows where high-stakes provisions receive immediate attorney attention while lower-risk clauses (such as territory specifications that merely identify geographic scope) are reviewed on a secondary pass. Benchmarking and negotiation support. By classifying clauses across a large corpus of comparable agreements, the system enables data-driven benchmarking. A licensing attorney negotiating an indemnification provision can quickly retrieve and compare indemnification clauses from similar transactions, providing empirical support for negotiation positions. This application leverages the cross-domain transfer capability: a model trained on SEC EDGAR filings can classify clauses in new, unseen agreements from similar domains. Quality assurance. Classification can serve as a completeness check, flagging agreements that are missing expected clause types. If a technology license agreement lacks an identifiable force majeure provision, for example, the classifier's failure to find one in any paragraph can alert the attorney to a potential gap. This application is particularly valuable for junior attorneys or non-specialists reviewing agreements outside their primary area of expertise. Institutional knowledge preservation. TLOs and corporate legal departments accumulate significant institutional knowledge about clause drafting practices, negotiation outcomes, and risk patterns. A clause classification system, trained on an organization's historical agreements, can serve as a repository of institutional knowledge, enabling consistency in contract drafting and review even as personnel change. Training and professional development. For junior attorneys and technology transfer professionals, a clause classification system with interpretable outputs can serve as a training tool. By examining which terms the classifier identifies as discriminative for each clause type, trainees can develop an understanding of the vocabulary patterns that characterize different licensing provisions. The key term analysis in Table 4 , for instance, effectively summarizes the distinctive language of each clause category in a form that is directly educational. Due diligence in technology M&A. In mergers and acquisitions involving technology-intensive companies, due diligence teams must review large volumes of existing license agreements to assess the target company's IP rights and obligations. The clause classification system, particularly its cross-domain transfer capability, can accelerate this review by automatically organizing clause-level information from heterogeneous agreement collections. The 75.0–78.12% cross-domain transfer accuracy, while not sufficient for autonomous analysis, provides a useful starting point that reduces the volume of unstructured text attorneys must process. 4.4 Comparison with prior work Table 7 Comparison with related legal NLP benchmarks (adapted from published results; self-compiled) Study Task Dataset Best model Performance Chalkidis et al. ( 2020 ) Legal text classification Various legal NLP tasks LEGAL-BERT micro-F1: 0.82–0.87 Hendrycks et al. ( 2021 ) Extractive QA CUAD (510 contracts) BERT-based F1: 0.60–0.75 Tuggener et al. ( 2020 ) Clause classification LEDGAR (100 + types) Various F1: ~0.80 Elwany et al. ( 2019 ) Paragraph classification Procurement contracts BERT F1: >0.90 This study Clause classification SEC EDGAR (2,000) TF-IDF + SVM macro-F1: 0.948 Our TF-IDF + SVM achieves F1 = 0.948, likely due to the formulaic nature of license agreements providing stronger discriminative features than general legal text. The comparison is instructive: LEGAL-BERT's 0.82–0.87 was measured on heterogeneous legal classification tasks, while our higher performance reflects the more structured and predictable vocabulary of technology licensing provisions. Hendrycks et al.'s ( 2021 ) lower F1 of 0.60–0.75 on CUAD reflects the substantially greater difficulty of extractive question answering compared to classification. It is important to note that this comparison does not imply that TF-IDF + SVM is generally superior to LEGAL-BERT or other transformer-based models. Rather, it demonstrates that for a specific document type with highly consistent vocabulary -- technology license agreements -- the simpler model captures the relevant signal with remarkable effectiveness. The practical significance of this finding for legal technology development is that not every legal NLP application requires a large pre-trained model; the choice of model should be guided by the characteristics of the target document type, the availability of labeled data, and the interpretability requirements of the deployment context. Elwany et al.'s ( 2019 ) result on procurement contracts (F1 > 0.90 with BERT) provides perhaps the most direct comparison, as procurement contracts share the standardized, formulaic character of technology license agreements. Our higher F1 with a simpler model suggests that the degree of vocabulary consistency in technology licensing may exceed even that of procurement contracts, making this document type particularly amenable to frequency-based classification. 4.5 Practical deployment considerations for technology transfer management The system's performance characteristics suggest specific deployment scenarios for TLOs and corporate licensing departments: 1. Automated contract triage : Classifying clauses in new drafts to detect missing clause types. At F1 = 0.948, the false negative rate is sufficiently low for this application. 2. Risk analysis pipeline : Identifying high-risk clauses (indemnification, termination, force majeure) for prioritized review. The per-class F1 for these categories (0.929, 0.944, and 0.968 respectively) supports reliable automated flagging. 3. Cross-agreement benchmarking : Comparing identical clause types across agreements to evaluate term fairness. The cross-domain transfer result (75.0–78.12% on unseen clauses) demonstrates that a single trained model can process heterogeneous agreement collections. 4. TLO workflow optimization : Reducing initial review time from hours to minutes. For a TLO managing 100 + active agreements, even partial automation of the initial classification step can free significant attorney time for higher-value analysis. A key practical consideration is data confidentiality. Technology license agreements frequently contain commercially sensitive information that organizations may be unwilling to transmit to cloud-based AI services. Both the TF-IDF + SVM pipeline and the Gemma 3 4B model used in this study run entirely on local infrastructure, requiring no external data transmission. This on-premise deployment capability may be a decisive factor for risk-averse legal departments. Cost-effectiveness for resource-constrained organizations. The computational requirements of the TF-IDF + SVM pipeline are minimal -- training on 2,000 clauses completes in approximately 45 seconds on consumer hardware, and inference is essentially instantaneous. This stands in stark contrast to large language model approaches that require specialized GPU infrastructure or expensive API subscriptions. For university TLOs and SMEs -- precisely the organizations most in need of automated contract review support due to limited legal staff -- the low cost of deploying a TF-IDF-based system removes a significant barrier to adoption. The small LLM alternative (Gemma 3 4B) requires more computational resources but remains deployable on a standard workstation with a consumer-grade GPU, avoiding the need for cloud computing expenditure. Incremental adoption pathway. The results suggest a natural adoption pathway for organizations new to legal AI. An organization can begin with the zero-shot LLM approach, which requires no labeled data and provides reasonable accuracy (81.3%) on small clause collections. As the organization accumulates labeled examples through routine contract review -- with attorneys confirming or correcting automated classifications -- the labeled dataset grows organically. Once the dataset reaches the 500–600 clause threshold, the organization can transition to the TF-IDF + SVM pipeline, gaining both higher accuracy and full interpretability. This incremental pathway reduces the upfront investment required for AI adoption while building organizational expertise in legal AI governance. 4.6 Professional responsibility and ethical considerations The deployment of AI tools for contract clause classification raises professional responsibility questions that legal practitioners must address. Duty of competence and technological literacy. The ABA Model Rules, as amended in 2012, require that attorneys maintain competence in "the benefits and risks of relevant technology" (Rule 1.1, Comment 8). For attorneys using clause classification tools, this duty implies an obligation to understand the tool's capabilities and limitations -- including its accuracy rates, its known failure modes (such as the territory-exclusivity confusion documented in our error analysis), and the conditions under which its performance may degrade (such as cross-lingual application or novel clause types). Supervision obligations. Under Model Rule 5.1 (Responsibilities of Partners, Managers, and Supervisory Lawyers), attorneys with supervisory responsibility must ensure that AI-assisted work is properly reviewed. The interpretability of TF-IDF-based classification directly supports this obligation: a supervising attorney can review not only the classification output but also the reasoning (feature weights) that produced it, enabling meaningful oversight rather than rubber-stamping. Automation bias. Psychological research on automation bias -- the tendency of humans to over-rely on automated recommendations -- poses a specific risk in AI-assisted legal review. If attorneys come to trust the classifier's outputs without independent verification, they may fail to catch errors that could have material consequences for their clients. This risk is amplified for high-performing systems: a classifier with 94.85% accuracy may engender more complacency than a mediocre one, precisely because its errors are rare and therefore unexpected. Mitigation strategies include: (a) presenting classifications with confidence scores rather than binary labels, so that borderline cases are flagged for manual review; (b) periodically introducing known errors as "calibration checks" to maintain attorney vigilance; and (c) designing user interfaces that require attorneys to affirmatively confirm or override each classification rather than passively accepting defaults. Liability allocation. When an AI-assisted classification error leads to a substantive legal mistake -- for example, if a clause is misclassified and consequently overlooked during review -- questions of liability arise. Under current malpractice frameworks, the attorney remains responsible for the quality of their work product regardless of whether it was produced with AI assistance. This allocation of liability reinforces the importance of meaningful human oversight and argues against fully automated "lights-out" contract review workflows. Bias in training data. The SEC EDGAR dataset reflects the clause drafting conventions of companies that file with the U.S. Securities and Exchange Commission -- predominantly large, publicly traded corporations with access to sophisticated legal counsel. Agreements drafted by smaller organizations, startups, or international parties may use different language conventions that the model has not encountered. Practitioners must be aware that the classifier's high accuracy was measured on a specific population of agreements and may not generalize equally to all technology licensing contexts. Confidentiality and data governance. Technology license agreements typically contain commercially sensitive terms -- royalty rates, territory restrictions, exclusivity grants -- that are protected by confidentiality obligations. The use of agreement text for training AI models raises questions about whether such use is consistent with confidentiality provisions in the agreements themselves. Organizations deploying clause classification systems must ensure that their training data governance practices comply with applicable confidentiality obligations. The on-premise deployment architecture demonstrated in this study -- where both TF-IDF and Gemma 3 4B run locally without transmitting data to external servers -- mitigates but does not eliminate this concern, as the trained model itself may encode information about the training documents. Informed consent and client communication. When attorneys use AI tools to assist with contract review, questions arise about the extent of disclosure required to clients. While no jurisdiction currently mandates specific disclosure of AI tool usage in contract review, the trend toward greater transparency about AI assistance in legal services suggests that practitioners should proactively inform clients about the role of automated tools in their work product. The interpretability of TF-IDF-based classification facilitates this disclosure by enabling attorneys to explain in concrete terms how the tool works and what role it played in the analysis. 4.7 Limitations Several limitations of this study should be acknowledged. Single-label classification. Real contract clauses frequently serve multiple legal functions simultaneously -- for example, a royalty-upon-termination provision addresses both royalty and termination, and an indemnification clause may incorporate warranty language. The current single-label framework assigns exactly one category per clause, which forces an artificial classification decision for such multi-functional provisions. This limitation affects both accuracy measurement (a clause labeled "indemnification" when it also functions as "warranty" is counted as an error on the warranty dimension) and practical utility (an attorney relying on single-label output may miss secondary functions). The error analysis (Section 3.8 ) confirms that the most common misclassification pairs -- territory/exclusivity, indemnification/warranty -- correspond precisely to categories with genuine functional overlap in legal drafting. Gemma 3 4B not tested at scale. The LLM was evaluated only on 64 clauses; inference at 3.8s/clause makes 2,000-clause evaluation feasible but was not performed. A full-scale LLM evaluation would provide a more complete comparison. No fine-tuned transformer baseline in the main comparison. The in-domain evaluation compares TF-IDF + SVM against a zero-shot LLM, but does not include fine-tuned BERT or LEGAL-BERT as an intermediate baseline. BERT and LEGAL-BERT appear only in the cross-domain transfer experiments (Table 6 ), where LEGAL-BERT achieves 78.12% on SEC→Custom-built transfer. Fine-tuning LEGAL-BERT on the full SEC EDGAR training set would likely yield accuracy between TF-IDF (94.85%) and zero-shot Gemma (81.3%), narrowing the apparent gap. This omission limits our ability to characterize the full accuracy-interpretability tradeoff curve and is a priority for future work. Nevertheless, the central argument of this study -- that interpretability provides a structural advantage for legal deployment -- holds regardless of the fine-tuned transformer's accuracy, because the interpretability gap between TF-IDF and any neural model remains qualitative rather than quantitative. Absence of Korean-specific model. Korean legal NLP requires specialized pre-training that was not available for this study. The poor Korean performance (46.88%) limits the system's applicability to international technology transfer transactions involving Korean-language agreements. Temporal generalization. All data is from a fixed period; contractual language evolves over time, and the model's performance on future agreements is not guaranteed. Taxonomy coverage. The taxonomy covers 8 categories selected for their centrality to technology licensing negotiations, but excludes commonly occurring clause types such as confidentiality, governing law, dispute resolution, audit rights, and assignment. These omissions mean that the system addresses a targeted subset of contract review tasks rather than comprehensive contract automation. A production-grade system would require at minimum 10–12 categories. The results reported here should therefore be interpreted as demonstrating the feasibility and characteristics of clause classification for technology licensing, not as a complete solution for automated contract review. Clause segmentation. Clause segmentation is assumed pre-done; in production, sentence/paragraph boundary detection would introduce additional error. Company-level leakage. Multiple clauses from the same company may appear in both train and test folds in the SEC EDGAR cross-validation, potentially inflating within-dataset accuracy due to shared drafting templates or law firm style. A company-level split -- where all clauses from a given company appear exclusively in either the training or test fold -- would provide a more conservative and externally valid estimate. We note, however, that the cross-domain transfer experiments (SEC→Custom-built at 75.0–78.12%, SEC→industrial at 43.2%) inherently avoid this leakage by testing on entirely separate document collections, and the consistency between in-domain and cross-domain results suggests that company-level leakage, while present, is unlikely to account for the majority of the in-domain performance. Legal jurisdiction specificity. The datasets are drawn predominantly from U.S. legal instruments (SEC filings, U.S. university TLO templates). The clause language conventions, and therefore the classifier's effectiveness, may differ substantially for agreements governed by other legal systems -- particularly civil law jurisdictions where contract drafting conventions differ from common law traditions. Absence of practitioner evaluation. While the taxonomy was validated by licensing professionals, the classification system itself has not been evaluated in a real-world legal workflow. User studies with attorneys performing actual contract review tasks would provide more ecologically valid evidence of the system's practical utility. 4.8 Future work Several directions for future research emerge from this study: 1. Multi-label classification with binary relevance or classifier chains, enabling clauses that span multiple categories to receive appropriate multi-category labels. 2. Fine-tuning BERT/LEGAL-BERT on SEC EDGAR for direct comparison with TF-IDF, establishing where the accuracy-interpretability tradeoff becomes favorable for neural models. 3. Korean legal language model development , addressing the cross-lingual limitation that currently prevents application to Korean-language agreements -- a pressing need given the volume of technology licensing activity in East Asia. 4. Expanding taxonomy to 10–12 categories (adding confidentiality, governing_law, dispute_resolution, audit) for more comprehensive contract coverage. 5. End-to-end contract risk scoring integration, combining clause classification with risk assessment to produce agreement-level risk profiles for portfolio management. 6. Practitioner evaluation studies involving attorneys performing contract review tasks with and without AI assistance, measuring both efficiency gains and error rates to validate the system's practical utility in legal workflows. 7. Human-in-the-loop interface design exploring how classification outputs should be presented to maximize attorney oversight while minimizing review time, including confidence calibration and explanation presentation. 8. Longitudinal analysis of how clause language evolves over time in technology licensing, and whether models trained on historical agreements maintain accuracy on contemporary contracts. 9. Cross-jurisdictional extension to civil law and hybrid legal systems, investigating whether the formulaic nature of contract language that enables high TF-IDF performance in common law agreements extends to other legal traditions. 5. Conclusion This study systematically evaluated NLP-based clause classification for technology license agreements across three dataset scales, with a focus on the implications for legal practice and AI-assisted contract review. On the SEC EDGAR dataset of 2,000 clauses from 812 companies, TF-IDF + SVM achieved 94.85% accuracy (macro F1 = 0.948), with per-class F1 ranging from 0.925 (territory) to 0.968 (ip_ownership, force_majeure). On LEDGAR 600 clauses, accuracy was 91.67% (F1 = 0.888). A small LLM (Gemma 3 4B) achieved 81.3% zero-shot on 64 English clauses, demonstrating utility in data-scarce settings. The key empirical finding is cross-domain transfer: a model trained on SEC regulatory filings achieves 75.0–78.12% on unseen custom-built clauses (with LEGAL-BERT outperforming TF-IDF), while LEDGAR-to-SEC transfer fails for all models (10.2–15.8%), confirming that label taxonomy alignment is more critical than data volume for building reusable legal classification tools. Industrial validation on 50 real petrochemical plant license agreements [company anonymized for review] yields 43.2% cross-domain accuracy, quantifying the domain gap between public benchmarks and proprietary industrial contracts. This has direct implications for legal technology development: standardized clause taxonomies are a prerequisite for interoperable contract analysis tools across organizations. From the perspective of legal practice, the TF-IDF approach offers an advantage that goes beyond accuracy: full interpretability. When attorneys rely on AI tools for contract review, they retain professional responsibility for the accuracy and completeness of their work. The ability to inspect exactly which terms drove each classification decision -- a property inherent to TF-IDF but absent from neural models -- supports the attorney's duty of competence and enables meaningful human oversight of automated classifications. For legal organizations considering adoption, the results suggest a concrete workflow: (1) begin with zero-shot LLM classification for immediate utility without labeled data (81.3% accuracy); (2) accumulate labeled examples through routine attorney review, with attorneys confirming or correcting automated outputs; (3) transition to TF-IDF + SVM once approximately 500 labeled clauses are available, gaining both higher accuracy (94.85%) and full interpretability. The practical boundary for this approach is well-defined: it works reliably for English-language, standardized contracts within the training taxonomy, but degrades for cross-lingual scenarios (46.88% on Korean-only), novel clause types, and proprietary industrial agreements with domain-specific technical vocabulary (43.2% on petrochemical contracts). Future work must address these boundaries through multilingual models, expanded taxonomies, and practitioner evaluation studies. More broadly, this study contributes to the ongoing discourse on the role of AI in legal practice by demonstrating a case where the simplest adequate technology may also be the most professionally appropriate one. The legal profession's emphasis on accountability, transparency, and human judgment is not merely a barrier to AI adoption -- it is a set of substantive values that should shape the design of legal AI systems. Clause classification systems that are interpretable by design, rather than by post-hoc approximation, align naturally with these values. The tension between classification accuracy and interpretability will shape the trajectory of legal AI adoption. As the field progresses toward more powerful but less transparent models, maintaining the ability of legal professionals to understand, verify, and take responsibility for AI-assisted decisions will remain a central challenge. Our results suggest that for the specific task of technology license agreement clause classification -- and likely for other structured legal document types -- this tension can be resolved in favor of both accuracy and interpretability, provided that sufficient domain-specific labeled data is available. The practical path forward for legal organizations is to invest in building such labeled datasets through their routine operations, thereby enabling high-performance interpretable classification that satisfies both the accuracy requirements of effective AI tools and the transparency requirements of professional legal practice. Declarations Competing interests The author declares no competing interests. Funding No external funding was received for this research. Author Contribution Y.K. designed the study, developed the methodology and software, collected and curated all datasets, performed all experiments and analyses, created all figures, and wrote the manuscript. Data Availability The SEC EDGAR and LEDGAR datasets were constructed from publicly available sources. The industrial dataset cannot be shared due to confidentiality obligations, but the evaluation protocol and aggregate results are reported. Code and data processing scripts are available at [repository URL to be provided upon acceptance]. References ABA (American Bar Association) (2023) Model Rules of Professional Conduct, Rule 1.1: Competence (including Comment 8 on technology). American Bar Association Ariai A, Jungiewicz M, Schilder F (2025) NLP for the legal domain: A survey of tasks, datasets, models, and challenges. *arXiv preprint arXiv:2501.01439* AUTM (Association of University Technology Managers) (2023) *AUTM Licensing Activity Survey. FY2022*. AUTM Bekkers R, Bongard R, Nuvolari A (2014) An empirical study on the determinants of essential patent claims in standards. *Research Policy* 43(7):1296–1311 Bommarito MJ, Katz DM (2021) GPT takes the bar exam. *arXiv preprint arXiv:2312.00914* Bozeman B (2000) Technology transfer and public policy: A review of research and theory. *Research Policy* 29(4–5):627–655 Chalkidis I, Fergadiotis M, Malakasiotis P, Aletras N, Androutsopoulos I (2020) LEGAL-BERT: The muppets straight out of law school. In: *Findings of the Association for Computational Linguistics: EMNLP 2020*, pp 2898–2904 Chalkidis I, Fergadiotis M, Androutsopoulos I (2021) MultiEURLEX -- A multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer. In: *Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing*, pp 6974–6996 Choi JH, Monahan A, Schwarcz D (2024) Lawyering in the age of artificial intelligence. *Minnesota Law Review* 109 Deeks A (2019) The judicial demand for explainable artificial intelligence. *Columbia Law Review* 119(7):1829–1850 Devlin J, Chang MW, Lee K, Toutanova K (2019) BERT: Pre-training of deep bidirectional transformers for language understanding. In: *Proceedings of NAACL-HLT*, pp 4171–4186 Elwany E, Moore D, Oberoi G (2019) BERT goes to law school: Quantifying the competitive advantage of access to large legal corpora in contract understanding. In: *Workshop on Document Intelligence at NeurIPS 2019* Google DM (2025) Gemma 3 technical report. *arXiv preprint arXiv:2503.19786* Hendrycks D, Burns C, Chen A, Ball S (2021) CUAD: An expert-annotated NLP dataset for legal contract understanding. In: *NeurIPS Datasets and Benchmarks Track* Joachims T (1998) Text categorization with support vector machines: Learning with many relevant features. In: *Proceedings of ECML*, pp 137–142 Kaplan J et al (2020) Scaling laws for neural language models. *arXiv preprint arXiv:2001.08361* Leivaditi S, Rossi J, Kamps J (2020) A benchmark for lease contract review. In: *Workshop on Scholarly Document Processing*, pp 1–10 Niklaus J, Chalkidis I, Sturmer M (2023) MultiLegalPile: A 689 GB multilingual legal corpus. *arXiv preprint arXiv:2306.02069* Perlman A (2024) The implications of ChatGPT for legal services and society. *Michigan Technol Law Review* 30(1):1–16 Razgaitis R (2009) *Valuation and Dealmaking of Technology-Based Intellectual Property*, 2nd edn. Wiley Remus D, Levy FS (2017) Can robots be lawyers? Computers, lawyers, and the practice of law. *Georgetown J Legal Ethics* 30:501–558 Rudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. *Nature Mach Intelligence* 1(5):206–215 Singh A, Joshi A, Jiang J, Paik HY (2025) A survey of classification tasks and approaches for legal contracts. *Artificial Intell Review*. https://doi.org/10.1007/s10462-025-11359-8 SEC (U.S. Securities and Exchange Commission) (2024) EDGAR full-text search system. https://efts.sec.gov/LATEST/search-index Suh JH (2016) Technology licensing: An empirical analysis of the role of contract clauses. *Research Policy* 45(6):1221–1232 Surden H (2019) Artificial intelligence and law: An overview. *Georgia State Univ Law Review* 35(4):1305–1337 Tuggener D, von Daniken C, Peber T, Cieliebak M (2020) LEDGAR: A large-scale multi-label corpus for text classification of legal provisions in contracts. In: *Proceedings of the 12th Language Resources and Evaluation Conference (LREC 2020)*, pp 1235–1241 Zhong H, Xiao C, Tu C, Zhang T, Liu Z, Sun M (2020) How does NLP benefit legal system: A summary of legal artificial intelligence. In: *Proceedings of ACL*, pp 5218–5230 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9298439","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":625374712,"identity":"406d2b15-3364-4a67-84c3-dce2afb8d5f8","order_by":0,"name":"Yuyong Kim","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyElEQVRIiWNgGAWjYDACHsYGhg8VNnIMzGAuM3FaGGecSTMmRQtQGW/b4cQGBmK1GJw53LqB50xa+obj7A8/MFRYw/Ti0XK2se2GRIVN7obDPMYSDGfSidBynrHthsGZtNxth3nYGBgRLiSgJbHtcLrZYfZnDIz/iNECctjBtsMJZocZzBgYG4jQInnmYNvNhjNphvtBfkk4lm5MUAvfmfRnt/9U2MhL9h9/+OFDjbUsQS0KB5B5CYSUg4A8QUNHwSgYBaNgFAAAA9dHSUMYJ5UAAAAASUVORK5CYII=","orcid":"","institution":"University of Wisconsin–Madison","correspondingAuthor":true,"prefix":"","firstName":"Yuyong","middleName":"","lastName":"Kim","suffix":""}],"badges":[],"createdAt":"2026-04-02 05:38:29","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9298439/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9298439/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107910099,"identity":"2e54c934-fab3-4592-9990-eb73e2f4ec2a","added_by":"auto","created_at":"2026-04-27 13:14:31","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":198094,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix for TF-IDF+SVM on the SEC EDGAR dataset (n = 2,000, stratified 5-fold cross-validation). Rows represent true labels; columns represent predicted labels\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/d3e612f2d1f54a630d90bb87.png"},{"id":108006799,"identity":"666c5a0a-a598-4b85-83a7-476ec4b531cc","added_by":"auto","created_at":"2026-04-28 12:57:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":113114,"visible":true,"origin":"","legend":"\u003cp\u003ePer-class F1-scores on the SEC EDGAR dataset. The dashed red line indicates the macro-average F1 = 0.948. All eight clause types achieve F1 above 0.92\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/3f464b86a47feda921fdb1f3.png"},{"id":107910101,"identity":"d6fc24d1-1160-4ad8-9fe3-6c7aafc1a6fe","added_by":"auto","created_at":"2026-04-27 13:14:31","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":210410,"visible":true,"origin":"","legend":"\u003cp\u003eTop-5 discriminative terms per clause category based on TF-IDF feature weights (SEC EDGAR dataset). Each subplot shows the terms with the highest TF-IDF weight for a given category\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/2d7dca8f6c25a299da26754a.png"},{"id":108006542,"identity":"78fa866a-d386-4deb-8d1f-48960dfa2cf2","added_by":"auto","created_at":"2026-04-28 12:55:57","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":126520,"visible":true,"origin":"","legend":"\u003cp\u003eCross-domain transfer accuracy across three scenarios and three models (TF-IDF+SVM, BERT, LEGAL-BERT). The dotted gray line indicates random chance (12.5%). SEC→Industrial is shown for TF-IDF only (industrial validation)\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/f7e3d43547b268452cad23bd.png"},{"id":107910102,"identity":"d5ec4048-8bc0-458f-887d-34ae92043c79","added_by":"auto","created_at":"2026-04-27 13:14:31","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":168223,"visible":true,"origin":"","legend":"\u003cp\u003eLearning curve for TF-IDF+SVM on SEC EDGAR subsamples (balanced stratified sampling at 48, 96, 200, 496, 1000, and 2000 clauses). The dashed orange line indicates Gemma 3 4B zero-shot baseline (81.3% on 64 custom-built clauses)\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/16c186454cf92bda87930970.png"},{"id":108008696,"identity":"b8aec68b-7a77-462a-9fb7-7cb60667610a","added_by":"auto","created_at":"2026-04-28 13:08:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1426150,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9298439/v1/9eea8379-c5d5-4cc6-ac94-44ac0dc348e7.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Interpretable NLP for Technology License Agreement Review: Clause Classification with TF-IDF and Small LLM across SEC, LEDGAR, and Industrial Datasets","fulltext":[{"header":"1. Introduction","content":"\u003cdiv id=\"Sec2\"\u003e\n \u003ch2\u003e1.1 Background and motivation\u003c/h2\u003e\n \u003cp\u003eTechnology transfer is a core mechanism for conveying technical knowledge and intellectual property among universities, research institutions, and corporations, playing a pivotal role in innovation ecosystems (Bozeman, 2000). Within the technology transfer process, license agreements are the most critical legal documents governing the rights and obligations between parties. A typical technology license agreement contains diverse provisions spanning royalty payment terms, licensed territory, exclusivity conditions, termination grounds, IP ownership allocation, indemnification clauses, warranties, and force majeure provisions, each carrying distinct legal implications and risk factors (Razgaitis, 2009).\u003c/p\u003e\n \u003cp\u003eTraditionally, contract review has relied on manual inspection by legal professionals. However, the growing volume and increasing complexity of license agreements have driven a surging demand for automated analysis tools. This need is particularly acute for small and medium-sized enterprises (SMEs) and Technology Licensing Offices (TLOs), which must review numerous agreements with limited legal resources. A typical university TLO may manage 50\u0026ndash;200 active license agreements simultaneously, and the time required for comprehensive manual review of each agreement -- estimated at 4\u0026ndash;8 hours per agreement by experienced legal counsel -- creates significant bottlenecks in the technology transfer workflow (AUTM, 2023). While recent work on reproducible NLP pipelines has addressed the upstream challenge of extracting and curating license agreements from regulatory filings at scale, the present study focuses on the downstream task: given a set of extracted license agreement clauses, how accurately and interpretably can NLP methods classify them by legal function?\u003c/p\u003e\n \u003cp\u003eThe legal practice context of technology licensing adds layers of complexity that distinguish this task from general document classification. Attorneys reviewing license agreements must not merely identify clause types but also assess their legal adequacy, identify missing provisions, flag non-standard terms for negotiation, and evaluate risk allocation across the entire agreement structure. Misclassification of a termination clause as a warranty provision, or failure to identify an indemnification obligation, can have material consequences for the client -- ranging from unexpected financial liabilities to the loss of intellectual property rights. In large-scale technology transfer transactions, where licensing portfolios may involve dozens or hundreds of agreements across multiple jurisdictions, the stakes of contract review errors are further amplified (Perlman, 2024).\u003c/p\u003e\n \u003cp\u003eThe professional responsibility dimension of AI-assisted contract review deserves particular attention. Under the American Bar Association Model Rules of Professional Conduct, attorneys bear a duty of competence (Rule 1.1) that extends to understanding the benefits and risks of technology relevant to their practice (ABA, 2023). When attorneys rely on automated tools for contract analysis, they remain professionally responsible for the accuracy and completeness of their work product. This creates a fundamental tension: the efficiency gains of automation must be balanced against the attorney\u0026apos;s obligation to exercise independent professional judgment, which in turn requires that the automated system\u0026apos;s reasoning be transparent and auditable.\u003c/p\u003e\n \u003cp\u003eThis tension between efficiency and accountability is not merely theoretical. As law firms and corporate legal departments increasingly adopt AI tools for contract lifecycle management, questions of professional liability for AI-assisted errors have emerged as a significant concern in legal practice (Surden, 2019; Remus \u0026amp; Levy, 2017). If an AI system misclassifies a critical indemnification clause and the attorney fails to catch the error during review, the resulting liability may fall on the attorney under existing malpractice frameworks. The interpretability of the classification system -- whether the attorney can understand why the system classified a clause in a particular way -- directly affects the attorney\u0026apos;s ability to fulfill this oversight responsibility.\u003c/p\u003e\n \u003cp\u003eAdvances in NLP technology, particularly the emergence of pre-trained language models, have opened new possibilities for legal document analysis (Chalkidis et al., 2020). Nevertheless, the majority of existing research focuses on general contracts or privacy policies, and studies specifically targeting clause classification in technology license agreements remain exceedingly scarce. Furthermore, systematic comparisons of classification performance across small custom-built datasets and large-scale public datasets are rare, leaving practitioners without clear guidance on the data requirements for deploying effective classification systems.\u003c/p\u003e\n \u003cp\u003eThis gap is significant for three reasons. First, technology license agreements occupy a unique position in the legal NLP landscape: they combine standardized legal boilerplate with domain-specific technical terminology, creating a hybrid vocabulary that general-purpose legal NLP tools may not handle effectively (Suh, 2016). Second, the practical deployment of clause classification in legal workflows requires not only high accuracy but also interpretability -- lawyers must be able to understand, verify, and when necessary override automated classifications, a requirement that favors simpler, more transparent models over opaque neural approaches. Third, the absence of cross-domain transfer studies for technology licensing clauses leaves unanswered a critical practical question: can an organization train a classifier on publicly available SEC filings and apply it to its own proprietary contracts without additional labeling effort? Answering this question has direct cost implications for TLOs and corporate legal departments considering AI adoption.\u003c/p\u003e\n \u003cp\u003eThe present study addresses these gaps by constructing and evaluating clause classification systems across three dataset scales, comparing an interpretable TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline against a zero-shot small LLM, and conducting cross-domain transfer experiments that test the reusability of trained models across document sources. Throughout, we foreground the legal practice perspective: accuracy is necessary but not sufficient; the system must also be interpretable, auditable, and compatible with the professional responsibility obligations that govern legal practice.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec3\"\u003e\n \u003ch2\u003e1.2 Related work\u003c/h2\u003e\n \u003cdiv id=\"Sec4\"\u003e\n \u003ch2\u003e1.2.1 Legal document NLP\u003c/h2\u003e\n \u003cp\u003eChalkidis et al. (2020) proposed LEGAL-BERT, a domain-specific pre-trained model for the legal domain, demonstrating superior performance over vanilla BERT across diverse downstream tasks. Hendrycks et al. (2021) constructed a large-scale contract dataset (CUAD) and defined an extractive question-answering task for contract clauses. These foundational works established that legal text classification benefits substantially from domain-specific representations, though they did not specifically address technology licensing provisions.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec5\"\u003e\n \u003ch2\u003e1.2.2 LEDGAR and contract clause benchmarks\u003c/h2\u003e\n \u003cp\u003eTuggener et al. (2020) published LEDGAR, a large-scale contract clause classification dataset extracted from SEC EDGAR filings, which has become a key benchmark for automated contract clause classification research. The LEDGAR taxonomy comprises over 100 clause types, reflecting the diversity of provisions in publicly filed agreements. Leivaditi et al. (2020) developed a benchmark for lease contract review, demonstrating that clause classification accuracy varies substantially across contract types. These benchmarks have enabled systematic comparison of classification approaches but have not been applied specifically to technology licensing -- a contract type with distinctive vocabulary arising from the intersection of legal and technical language.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec6\"\u003e\n \u003ch2\u003e1.2.3 LLMs and small language models for legal applications\u003c/h2\u003e\n \u003cp\u003eThe application of language models to legal text has evolved rapidly in recent years. Zhong et al. (2020) surveyed how NLP benefits the legal system, including applications to contract analysis, judgment prediction, and legal information retrieval. Bommarito and Katz (2021) applied GPT-3 to legal document analysis tasks including contract provision identification, demonstrating that large language models could perform legal reasoning tasks with minimal task-specific training. Elwany et al. (2019) developed a BERT-based system for classifying contract paragraphs in procurement agreements, achieving F1 scores above 0.90 on common clause types.\u003c/p\u003e\n \u003cp\u003eMore recently, the emergence of smaller, locally deployable language models has opened new possibilities for legal AI applications where data confidentiality is paramount. Google DeepMind (2025) released the Gemma 3 family of models, which includes a 4-billion-parameter variant suitable for local deployment on consumer hardware -- a practical consideration for law firms and corporate legal departments where client confidentiality requirements may preclude transmitting agreement text to cloud-based AI services. The tension between model capability (which generally scales with model size) and deployment constraints (which favor smaller models) is particularly acute in legal applications, where the most sensitive documents are often the ones most in need of automated analysis support.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003e1.2.4 Technology transfer-specific NLP\u003c/h2\u003e\n \u003cp\u003eSuh (2016) analyzed the textual characteristics of technology license agreements and noted that the interplay between technical and legal language creates a hybrid vocabulary that distinguishes these agreements from other contract types. A technology license agreement may reference specific patent claims, describe technical performance standards, and define regulatory milestones alongside standard legal provisions -- creating a document that requires both legal and technical expertise to review comprehensively. Bekkers et al. (2014) examined the relationship between patent licensing terms and technology characteristics, demonstrating that clause-level analysis of license agreements can reveal underlying negotiation dynamics. These studies highlight the domain specificity of technology licensing language but do not address automated classification, leaving open the question of whether NLP tools developed for general legal text can handle the hybrid vocabulary of technology licensing.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec8\"\u003e\n \u003ch2\u003e1.2.5 Cross-lingual legal NLP\u003c/h2\u003e\n \u003cp\u003eChalkidis et al. (2021) introduced MultiEURLEX, demonstrating that zero-shot cross-lingual transfer entails significant performance degradation for morphologically complex languages. Niklaus et al. (2023) found that language-specific pre-training consistently outperforms multilingual models for legal text classification. For technology licensing, cross-lingual capability is particularly important given the international nature of technology transfer transactions, where agreements frequently involve parties in different jurisdictions and languages. A Korean chemical company licensing technology from an American university, for instance, may need to review agreement texts in both English and Korean -- a scenario that directly motivates the cross-lingual experiments in this study.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec9\"\u003e\n \u003ch2\u003e1.2.6 Legal technology adoption and barriers\u003c/h2\u003e\n \u003cp\u003eThe adoption of AI tools in legal practice has been slower than in many other professional domains, a phenomenon attributed to several factors: the conservative culture of legal practice, regulatory requirements for human oversight, liability concerns, and the interpretability demands of legal reasoning (Remus \u0026amp; Levy, 2017). Surden (2019) argued that the most successful legal AI applications are those targeting well-defined, repetitive tasks with clear evaluation criteria -- a description that closely matches clause classification in standardized contract types. Perlman (2024) examined the regulatory implications of AI in legal practice, noting that existing professional responsibility frameworks were not designed for the delegation of judgment to automated systems, creating uncertainty about the scope of permissible automation. More recently, the rapid deployment of generative AI tools in legal contexts has intensified debates about the boundaries of acceptable automation in legal work (Choi et al., 2024).\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec10\"\u003e\n \u003ch2\u003e1.2.7 Responsible AI and human-in-the-loop in legal domains\u003c/h2\u003e\n \u003cp\u003eThe responsible AI literature has increasingly emphasized the legal domain as a context where the stakes of automated decision-making are particularly high. The concept of human-in-the-loop AI -- where automated systems augment rather than replace professional judgment -- has gained particular traction in legal technology. Surden (2019) argued that the most promising legal AI applications are those that keep humans in the decision loop while automating the most time-consuming aspects of document analysis. This perspective aligns with the professional responsibility framework governing legal practice, where attorneys are expected to exercise independent judgment regardless of the tools they employ.\u003c/p\u003e\n \u003cp\u003eThe distinction between AI-assisted and AI-autonomous legal work is critical. In the context of contract review, an AI-assisted workflow uses automated classification to organize and prioritize clauses for human review, while an AI-autonomous workflow would act on classifications without human verification. The professional responsibility literature strongly favors the former approach, and the design of clause classification systems should reflect this preference by emphasizing interpretability and confidence calibration over raw classification speed.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec11\"\u003e\n \u003ch2\u003e1.2.8 Interpretability requirements in legal NLP\u003c/h2\u003e\n \u003cp\u003eInterpretability occupies a distinct position in legal AI compared to other domains. While interpretability is generally desirable in machine learning, in legal practice it may be a mandatory requirement. Attorneys cannot delegate their professional judgment to a system whose reasoning they cannot evaluate, and courts may require explanations for AI-assisted legal conclusions (Deeks, 2019). Rudin (2019) argued forcefully that high-stakes domains should abandon the practice of explaining black-box models and instead use inherently interpretable models, a position that resonates strongly with the legal profession\u0026apos;s emphasis on transparent reasoning.\u003c/p\u003e\n \u003cp\u003eSingh et al. (2025) surveyed classification tasks for legal contracts and noted that interpretability remains an underexplored dimension of legal NLP evaluation -- most studies report accuracy and F1 scores but do not assess whether their models produce outputs that legal professionals can meaningfully review. Ariai et al. (2025) provided a comprehensive survey of NLP applications in the legal domain, identifying contract analysis as one of the most promising near-term applications but noting persistent gaps in interpretability and domain adaptation. The present study responds to this gap by explicitly evaluating an inherently interpretable model (TF-IDF\u0026thinsp;+\u0026thinsp;SVM) alongside a less interpretable alternative (zero-shot LLM), foregrounding the interpretability dimension as a design criterion rather than an afterthought.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\"\u003e\n \u003ch2\u003e1.3 Research questions and contributions\u003c/h2\u003e\n \u003cp\u003eThis study addresses three research questions at the intersection of NLP and legal practice:\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eRQ1\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eHow does clause classification accuracy scale with dataset size for technology license agreements, and at what data volume does a simple, interpretable TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline match or exceed a zero-shot LLM?\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eRQ2\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eTo what extent can models trained on one corpus (SEC regulatory filings) generalize to unseen clause collections from different sources, and what does this imply for building reusable legal AI tools?\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eRQ3\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eWhat are the primary barriers to cross-domain and cross-lingual transfer in legal clause classification, and how do these barriers affect the practical deployment of classification systems in technology licensing workflows?\u003c/p\u003e\n \u003cp\u003eThis study makes the following contributions:\u003c/p\u003e\n \u003cp\u003e1. \u003cstrong\u003eDomain-specific clause taxonomy for technology licensing practice\u003c/strong\u003e: An eight-category taxonomy grounded in WIPO and AUTM standards, validated by technology licensing professionals, providing a classification framework aligned with the categories that practitioners use in daily contract review.\u003c/p\u003e\n \u003cp\u003e2. \u003cstrong\u003eMulti-scale dataset construction and cross-domain validation\u003c/strong\u003e: Systematic comparison across three dataset scales (128 custom-built, 600 LEDGAR, 2,000 SEC EDGAR from 812 companies), with cross-domain transfer experiments that reveal the conditions under which models can generalize across document sources -- a critical practical question for organizations seeking to build reusable classification tools.\u003c/p\u003e\n \u003cp\u003e3. \u003cstrong\u003eInterpretable model as a feature, not a limitation\u003c/strong\u003e: We demonstrate that TF-IDF\u0026thinsp;+\u0026thinsp;SVM achieves 94.85% accuracy on technology licensing clauses, and argue that its full interpretability -- lawyers can inspect exactly which terms drove each decision -- constitutes a practical advantage for legal deployment, where professional responsibility requires human oversight of automated classifications.\u003c/p\u003e\n \u003cp\u003e4. \u003cstrong\u003eFrequency-based vs. zero-shot LLM comparison with legal practice implications\u003c/strong\u003e: Direct comparison showing where simple interpretable models and zero-shot LLMs each have advantages, with discussion of the tradeoffs between accuracy, interpretability, and deployment cost that legal practitioners face when selecting AI tools.\u003c/p\u003e\n \u003cp\u003e5. \u003cstrong\u003eProfessional responsibility analysis\u003c/strong\u003e: Discussion of how clause classification systems interact with attorneys\u0026apos; ethical obligations, including competence requirements, supervision duties, and the risk of automation bias in legal review workflows.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"2. Methodology","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Clause taxonomy\u003c/h2\u003e \u003cp\u003eThis study classifies the core clauses of technology license agreements into eight categories:\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExample keywords\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eroyalty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCompensation terms for technology usage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eroyalty, payment, net sales, milestone\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eterritory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGeographic scope of the license grant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eterritory, region, country, worldwide\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eexclusivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExclusive/non-exclusive license scope\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eexclusive, non-exclusive, sole, sublicense\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etermination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGrounds and procedures for contract termination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eterminate, breach, cure, expiration\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eip_ownership\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAllocation of background and foreground IP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eownership, invention, improvement, patent\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eindemnification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiability and hold-harmless conditions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eindemnify, liability, damages, hold harmless\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ewarranty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScope of technology and product warranties\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ewarranty, represent, warrant, merchantability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eforce_majeure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eForce majeure events and relief\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eforce majeure, act of God, beyond control\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThis taxonomy was designed with reference to the WIPO technology transfer guidelines, AUTM model license agreement provisions, and the LEDGAR classification scheme. The eight categories were validated by interviews with three technology licensing professionals at Korean Technology Licensing Offices (TLOs). These categories represent the provisions that licensing practitioners most frequently identify as requiring careful review and negotiation in technology transfer transactions. The taxonomy was designed to balance comprehensiveness with practical utility: while real-world technology license agreements may contain additional provision types (e.g., confidentiality, governing law, audit rights, dispute resolution), the eight selected categories cover the provisions most commonly subject to negotiation and most consequential for risk allocation. From a legal practice standpoint, the ability to reliably identify these eight categories enables the most impactful automation -- flagging the provisions where drafting errors or unfavorable terms carry the greatest financial and legal exposure.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Dataset construction\u003c/h2\u003e \u003cp\u003eThree primary datasets were constructed to enable multi-scale evaluation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDataset 1: Custom-built (128 clauses).\u003c/b\u003e 128 clauses were extracted from WIPO model contracts, publicly available U.S. university TLO agreement templates (Stanford, MIT, Harvard), and Korean TLO public contract samples (KAIST, SNU, KIST). Both English-only (64 clauses) and Korean-English mixed (128 clauses) versions were constructed, with 8 English\u0026thinsp;+\u0026thinsp;8 Korean clauses per category. These clauses represent the standardized templates that TLOs use as starting points for license negotiations.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDataset 2: LEDGAR extraction (600 clauses).\u003c/b\u003e 600 clauses matching this study's eight categories were sampled from Tuggener et al.'s (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) LEDGAR dataset. The distribution is unbalanced, reflecting the natural frequency of clause types in SEC filings: royalty (142), territory (126), termination (90), force_majeure (59), exclusivity (53), ip_ownership (46), warranty (46), indemnification (38).\u003c/p\u003e \u003cp\u003e \u003cb\u003eDataset 3: SEC EDGAR collection (2,000 clauses, 812 companies).\u003c/b\u003e 2,000 clauses were collected and labeled from technology transfer and license agreements attached to 8-K and 10-K filings in the SEC EDGAR system (SEC, 2024), spanning 812 companies across 14 industry sectors including pharmaceuticals, biotechnology, software, semiconductors, chemicals, and energy. 250 clauses per category were obtained through stratified sampling. Two legal professionals independently labeled each clause; inter-annotator agreement (Cohen's kappa) was 0.91, indicating strong agreement on clause categorization. The SEC EDGAR source provides a unique advantage for legal AI research: the agreements are publicly filed regulatory documents, eliminating the confidentiality concerns that typically limit access to commercial contract datasets. This public availability enables reproducibility and external validation -- properties that are essential for building trust in legal AI tools but are often absent from studies that rely on proprietary contract collections.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDataset 4: Industrial validation (270 clauses, 50 agreements).\u003c/b\u003e To assess real-world applicability beyond public benchmark data, 270 clauses were extracted from 50 actual technology license agreements held by a major Korean petrochemical company [company anonymized for review]. These agreements cover process technology licenses for polyethylene (PE), polypropylene (PP), styrene monomer (SM), ethylene glycol (EG), butadiene (BD), and naphtha cracking (NCC) facilities, representing authentic industrial licensing practice. The agreements span multiple international licensors, covering the period 1994\u0026ndash;2026. The clause distribution is naturally imbalanced and includes categories outside the eight-category taxonomy: unclassified (102), confidentiality (36), warranty (33), ip_ownership (31), royalty (24), indemnification (18), termination (12), governing_law (7), exclusivity (5), force_majeure (2). Of the 270 clauses, 125 fall within the eight target categories and were used for cross-domain evaluation. This dataset was not used for training; it serves exclusively as an external validation set representing the real-world deployment scenario: can a model trained on public SEC filings classify clauses in proprietary industrial agreements? Due to confidentiality obligations, the raw clause text cannot be shared, but the evaluation protocol and aggregate results are reported.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Text preprocessing\u003c/h2\u003e \u003cp\u003eThe following preprocessing pipeline was applied:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e1. \u003cb\u003eLowercasing\u003c/b\u003e (English): Case normalization\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e2. \u003cb\u003eSpecial character and number removal\u003c/b\u003e: Elimination of section numbers and parenthetical references\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e3. \u003cb\u003eStop word removal\u003c/b\u003e: English NLTK stop words and general-purpose legal domain stop words\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e4. \u003cb\u003eKorean morphological analysis\u003c/b\u003e: KoNLPy-based morpheme analysis and noun/verb extraction (for KO\u0026thinsp;+\u0026thinsp;EN mixed dataset)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e5. \u003cb\u003eTokenization\u003c/b\u003e: Whitespace-based (TF-IDF) or WordPiece (LLM)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Classification models\u003c/h2\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1 TF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/h2\u003e \u003cp\u003eA document-term matrix was generated using TF-IDF with up to 5,000 features considering unigrams and bigrams. A linear kernel SVM classifier with balanced class weights was trained on this representation. Stratified 5-fold cross-validation was performed. TF-IDF vectorizer: unigram+bigram (ngram_range=(1,2)), max_features\u0026thinsp;=\u0026thinsp;5000, English stop words removed, L2 normalization. SVM: linear kernel (LinearSVC), C\u0026thinsp;=\u0026thinsp;1.0, class_weight='balanced'. These defaults were used without hyperparameter tuning due to the already high performance.\u003c/p\u003e \u003cp\u003e \u003cem\u003e\\text{TF-IDF}(t, d) = \\text{TF}(t, d) \\times \\log\\frac{N}{\\text{DF}(t)}\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eImplementation details.\u003c/b\u003e Python 3.10.12, scikit-learn 1.3.2 for TF-IDF vectorization and SVM classification, NLTK 3.8.1 for English preprocessing, KoNLPy 0.6.0 with Okt backend for Korean morphological analysis, and Ollama 0.6.2 for Gemma 3 4B inference. TF-IDF feature extraction with max_features\u0026thinsp;=\u0026thinsp;5000 and ngram_range=(1,2) was applied consistently. For the SEC EDGAR dataset, total TF-IDF\u0026thinsp;+\u0026thinsp;SVM training and evaluation time was approximately 45 seconds on a consumer-grade workstation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2 Gemma 3 4B (Small LLM)\u003c/h2\u003e \u003cp\u003eGemma 3 4B (Google DeepMind, 2025) was run in the Ollama local inference environment for zero-shot classification. The system prompt was structured as follows: (1) a role instruction (\"You are a legal contract clause classifier\"), (2) explicit definitions for each of the eight categories with one canonical example per category drawn from WIPO model contracts, and (3) an output constraint requiring the model to respond with only the category label. The prompt was fixed across all evaluations -- no prompt tuning or iterative selection was performed. Temperature was set to 0.0 for deterministic outputs, and the maximum output token length was limited to 20 tokens to prevent verbose responses. Average inference time was 3.8 seconds per clause on consumer hardware (NVIDIA RTX 3060 Ti, 8 GB VRAM). The local deployment architecture -- running entirely on-premise without sending data to external servers -- is a significant practical consideration for legal applications where client confidentiality requirements may preclude cloud-based AI services.\u003c/p\u003e \u003cp\u003e \u003cb\u003e2.5 Evaluation metrics\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAccuracy\u003c/b\u003e: Proportion of correct predictions\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMacro F1-score\u003c/b\u003e: Simple average of per-class F1 scores, giving equal weight to each clause type regardless of frequency\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePer-class F1-score\u003c/b\u003e: Harmonic mean of precision and recall per category\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eConfusion matrix\u003c/b\u003e: Inter-class misclassification patterns\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Cross-domain transfer protocol\u003c/h2\u003e \u003cp\u003eTo assess generalization across document sources -- a critical question for organizations seeking to build reusable classification tools -- two cross-domain experiments were conducted. In each experiment, a TF-IDF\u0026thinsp;+\u0026thinsp;SVM model was trained on one dataset and tested on a completely separate, unseen dataset with no overlapping documents:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSEC EDGAR to Custom-built\u003c/b\u003e: Model trained on all 2,000 SEC EDGAR clauses, tested on 64 English custom-built clauses\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSEC EDGAR to Industrial\u003c/b\u003e: Model trained on all 2,000 SEC EDGAR clauses, tested on 125 classified industrial clauses\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLEDGAR to SEC EDGAR\u003c/b\u003e: Model trained on 600 LEDGAR clauses, tested on a held-out set of 400 SEC EDGAR clauses\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eNo target-domain training data was used in either experiment, simulating a realistic deployment scenario where an organization trains a model on available labeled data and applies it to new, unlabeled contract collections. This protocol directly tests the practical question facing legal departments: can a model trained on publicly available agreements be applied \"out of the box\" to the organization's own contracts, or does each new contract collection require its own labeled training set?\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Classification performance by dataset scale\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClassification performance summary (self-compiled from experiment runs)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClauses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCompanies\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMacro-F1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCustom-built (EN)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e--\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e68.75%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.689\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCustom-built (EN)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e--\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGemma 3 4B\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e81.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.795\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCustom-built (KO\u0026thinsp;+\u0026thinsp;EN)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e--\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e55.47%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.564\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLEDGAR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e--\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91.67%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSEC EDGAR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94.85%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe progression from 68.75% (64 EN clauses) to 94.85% (2,000 clauses) -- a 26.1 percentage point improvement -- substantially exceeds typical learning curve expectations for general text classification. This confirms that legal documents possess particularly consistent lexical patterns that TF-IDF can exploit effectively once sufficient examples are available.\u003c/p\u003e \u003cp\u003eThe 12.5 percentage point advantage of Gemma 3 4B over TF-IDF on the small dataset (81.3% vs. 68.75%) underscores the value of pre-trained language models in low-resource settings. However, the gap between Gemma 3 4B (81.3% on 64 clauses) and TF-IDF (94.85% on 2,000 clauses) demonstrates that sufficient labeled data with a simple model ultimately outperforms a more sophisticated model with limited data.\u003c/p\u003e \u003cp\u003eFrom a legal practice perspective, this scaling behavior has practical significance. A TLO or corporate legal department beginning to build a clause classification system will initially find an LLM-based approach more accurate with minimal upfront investment in labeled data. However, as the organization accumulates labeled examples through routine contract review -- a natural byproduct of legal practice -- the interpretable TF-IDF pipeline will surpass the LLM's zero-shot performance while providing the transparency that professional responsibility demands.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Per-class performance on SEC EDGAR 2,000 clauses\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePer-class results on SEC EDGAR dataset (stratified 5-fold cross-validation)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSupport\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eroyalty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.923\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.964\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eterritory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.892\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.925\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eexclusivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.954\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etermination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.940\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.944\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eip_ownership\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.980\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.956\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eindemnification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.921\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.936\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ewarranty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.945\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.957\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eforce_majeure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e250\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro Average\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.949\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.949\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2,000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the confusion matrix and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the per-class F1 distribution. Per-class F1 scores ranged from 0.925 (territory) to 0.968 (ip_ownership and force_majeure). Territory's lower F1 reflects lexical overlap with exclusivity (shared license-scope terminology). IP ownership and force majeure achieved the highest F1 due to unique discriminative keywords (\"ownership,\" \"invention,\" \"patent\" for IP; \"pandemic,\" \"covid,\" \"force majeure\" for FM).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe narrow F1 range (0.925\u0026ndash;0.968, spread of 4.3 pp) contrasts sharply with the custom-built dataset's wide variation (0.316\u0026ndash;0.933). At larger data scales, all eight clause types converge to high classification performance, confirming that each type possesses sufficiently distinctive vocabulary patterns.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e3.3 LEDGAR per-class performance\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePer-class results on LEDGAR dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSupport\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eexclusivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.845\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.925\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e53\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eforce_majeure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.847\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.820\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e59\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eindemnification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.737\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e38\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eip_ownership\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.813\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eroyalty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.965\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e142\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etermination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.926\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.967\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.946\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eterritory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.972\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e126\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ewarranty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.905\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro Average\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e600\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eLEDGAR's lower macro F1 (0.888 vs. 0.948) despite sourcing from similar SEC filings reflects both its smaller size and the label mapping challenge -- LEDGAR's native taxonomy does not align perfectly with our eight categories, introducing noise.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Key term analysis\u003c/h2\u003e \u003cp\u003eThe interpretability of TF-IDF enables direct inspection of which terms the classifier relies on for each category (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). This transparency is a distinctive advantage for legal applications, where practitioners need to verify that classification decisions are based on legally meaningful features rather than statistical artifacts. Top-5 key terms per category based on TF-IDF weights (SEC EDGAR):\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTop discriminative terms by clause category\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTop-5 key terms\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eroyalty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esales, royalty, net, payments, net sales\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eterritory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ejurisdiction, geographic, tax, income, company\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eexclusivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eexclusive, license, company, exclusive license, million\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etermination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eexpire, termination, lease, terminate, terminated\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eip_ownership\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epatent, intellectual property, intellectual, property, patents\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eindemnification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eliabilities, liability, value, assets, company\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ewarranty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ewarrants, stock, shares, common, common stock\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eforce_majeure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003epandemic, covid, covid pandemic, impact, economic\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSeveral observations are relevant to legal practice. The royalty category's top terms (\"sales,\" \"royalty,\" \"net,\" \"payments,\" \"net sales\") directly correspond to the financial provisions that licensing attorneys prioritize during review. The ip_ownership terms (\"patent,\" \"intellectual property,\" \"property,\" \"patents\") align precisely with the IP allocation provisions that are among the most heavily negotiated aspects of technology license agreements. The force_majeure terms reflect the post-COVID evolution of contractual language, where pandemic-related provisions have become standard in technology licensing. Notably, some categories exhibit \"semantic leakage\" -- the warranty category's top terms include \"warrants,\" \"stock,\" and \"shares,\" which in SEC filings relate to stock warrants rather than contractual warranties. This demonstrates a known limitation of corpus-based approaches: terms that are discriminative in the training corpus may carry different meanings in other contract contexts, and legal practitioners must be aware of this potential for domain-specific ambiguity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Cross-lingual performance\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCross-lingual evaluation results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConfiguration\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClauses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMacro-F1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnglish only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e68.75%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.689\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKorean only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e46.88%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.492\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMixed KO\u0026thinsp;+\u0026thinsp;EN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e55.47%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.564\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe 21.9 pp drop from English-only (68.75%) to Korean-only (46.88%) reflects three compounding factors. First, a \u003cb\u003estructural limitation\u003c/b\u003e: Korean's agglutinative morphology generates far more unique tokens for equivalent semantic content -- the Korean vocabulary size was approximately 3.2x larger than the English vocabulary for the same number of clauses, diluting the discriminative power of individual terms. Second, a \u003cb\u003edata scarcity effect\u003c/b\u003e: with only 64 Korean clauses, the TF-IDF vocabulary is too sparse for reliable weight estimation, and this small sample cannot be disentangled from the morphological effect without a larger Korean corpus. Third, a \u003cb\u003edomain vocabulary gap\u003c/b\u003e: Korean legal drafting conventions use Sino-Korean legal terminology (e.g., 손해배상, 불가항력) that does not share surface-form overlap with English equivalents, unlike European languages where Latin-derived legal terms provide partial cognate matching. This result directly addresses \u003cb\u003eRQ3\u003c/b\u003e by identifying cross-lingual transfer as one of the primary barriers to practical deployment. In international technology transfer practice, agreements are frequently drafted in multiple languages or bilingual formats, and parties from different jurisdictions must review agreements in their non-native language. The inability of frequency-based methods to handle Korean legal text without substantial adaptation suggests that multilingual legal AI tools require fundamentally different approaches -- likely involving multilingual pre-trained models (e.g., XLM-RoBERTa, mBERT) or language-specific fine-tuning -- rather than simple extensions of English-language pipelines. For legal departments managing cross-border licensing portfolios, this finding implies that separate classification pipelines may be needed for each language until multilingual models achieve sufficient domain-specific performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Cross-domain transfer results\u003c/h2\u003e \u003cp\u003eTo evaluate generalization across corpora, two cross-domain experiments were conducted: each model was trained on one dataset and tested on a separate, unseen dataset with no overlapping documents.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCross-domain transfer performance (no target-domain training data). The industrial dataset contained 270 clauses in total; only the 125 clauses mapped to the eight target categories were used for evaluation. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e visualizes these results.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTrain set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTest set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest N\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMacro-F1\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSEC EDGAR (2,000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCustom-built EN (64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e75.00%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.737\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSEC EDGAR (2,000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCustom-built EN (64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.118\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSEC EDGAR (2,000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCustom-built EN (64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLEGAL-BERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e78.12%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.766\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSEC EDGAR (2,000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eindustrial (125)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e125\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e43.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.370\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLEDGAR (600)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSEC EDGAR held-out (400)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e400\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.089\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLEDGAR (600)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSEC EDGAR held-out (400)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e400\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.118\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLEDGAR (600)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSEC EDGAR held-out (400)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e400\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLEGAL-BERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.85%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.105\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThree findings emerge from the cross-domain experiments.\u003c/p\u003e \u003cp\u003eFirst, the SEC-to-Custom-built transfer demonstrates meaningful generalization: TF-IDF achieves 75.0% and LEGAL-BERT achieves 78.12% on unseen template clauses, confirming that legal boilerplate language is sufficiently standardized across document types for partial transfer. Notably, LEGAL-BERT's domain-specific pre-training provides an advantage over TF-IDF in this cross-domain setting (78.12% vs. 75.0%), while vanilla BERT collapses to 15.8% -- suggesting that domain-specific pre-training captures transferable legal language representations that general pre-training does not.\u003c/p\u003e \u003cp\u003eSecond, the \u003cb\u003eindustrial validation\u003c/b\u003e reveals a significant but instructive domain gap. A TF-IDF model trained on SEC EDGAR public filings achieves 43.2% on real petrochemical plant license agreements -- substantially above random chance (12.5%) but well below the in-domain performance (95.2%). This gap reflects the real-world challenge that legal AI tools face when moving from public benchmark data to proprietary industrial contracts: the industrial agreements contain technical terminology specific to process licensing (e.g., catalyst specifications, reactor design parameters, process guarantees) that does not appear in SEC filings, and the drafting conventions of international process licensors differ from U.S. domestic regulatory filing norms. Critically, the 43.2% result should be interpreted not as a failure but as a realistic baseline for zero-shot industrial deployment: it demonstrates that a model trained entirely on public data can extract meaningful signal from proprietary contracts (3.5x above random chance), while quantifying the additional investment in domain-specific labeling required for production-grade accuracy. For practitioners, this sets an honest expectation: public data provides a useful starting point, but industrial deployment requires domain-specific fine-tuning.\u003c/p\u003e \u003cp\u003eThird, the LEDGAR-to-SEC transfer fails for all models (10.2\u0026ndash;15.8%), confirming that label taxonomy alignment is the binding constraint independent of model sophistication. Neither TF-IDF, BERT, nor LEGAL-BERT can overcome the fundamental mismatch between LEDGAR's native taxonomy and the technology-licensing-specific categories used in this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e3.7 Statistical significance\u003c/h2\u003e \u003cp\u003eBootstrap resampling (2,000 iterations) on the SEC EDGAR 5-fold CV predictions yields a 95% confidence interval of [94.85%, 96.65%] for accuracy (mean\u0026thinsp;=\u0026thinsp;95.76%). McNemar's test comparing TF-IDF\u0026thinsp;+\u0026thinsp;SVM against a random baseline (accuracy 10.6%) gives chi-squared\u0026thinsp;=\u0026thinsp;1,677.36, p\u0026thinsp;\u0026lt;\u0026thinsp;1e-300, confirming that the model's performance is not attributable to chance.\u003c/p\u003e \u003cp\u003eFor the Gemma 3 4B zero-shot evaluation on 64 English clauses, the Wilson score 95% confidence interval is [70.7%, 88.8%], reflecting the substantial uncertainty inherent in small-sample evaluation. The width of this interval (18.1 pp) underscores the importance of large-scale evaluation for establishing reliable performance estimates -- a practical consideration for organizations evaluating AI tools on their own (typically small) contract collections.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e3.8 Error analysis\u003c/h2\u003e \u003cp\u003eOn the SEC EDGAR dataset, the most common confusion pair was territory\u0026ndash;exclusivity, consistent with their lexical overlap in license-scope language. Both categories address the scope of the license grant but from different dimensions -- geographic (territory) vs. competitive (exclusivity) -- and the shared vocabulary of \"license,\" \"grant,\" and \"rights\" creates classification ambiguity that mirrors genuine drafting overlap in practice.\u003c/p\u003e \u003cp\u003eThe second most common was indemnification misclassification (F1\u0026thinsp;=\u0026thinsp;0.929, lowest among core categories), where liability-related language overlaps with warranty clauses. In legal drafting, indemnification and warranty provisions are frequently co-located and cross-referenced within agreements, which means that a clause mentioning both \"warrants\" and \"indemnify\" presents a genuine classification challenge that even human reviewers may resolve differently depending on their interpretation of the clause's primary function.\u003c/p\u003e \u003cp\u003eForce majeure and IP ownership exhibited the lowest error rates, consistent with their unique discriminative vocabulary. From a legal practice perspective, these categories employ highly distinctive terminology -- \"force majeure,\" \"act of God,\" \"pandemic\" for the former; \"patent,\" \"invention,\" \"intellectual property\" for the latter -- that rarely appears in other clause types. This vocabulary distinctiveness explains both the high TF-IDF performance and the practical observation that these clause types are among the easiest for human reviewers to identify as well.\u003c/p\u003e \u003cp\u003eThe error patterns have direct implications for deployment design. The territory\u0026ndash;exclusivity confusion suggests that these two categories should be presented with confidence scores in any practitioner-facing interface, alerting the reviewer to cases where the classifier is uncertain. Conversely, the high reliability of force_majeure and ip_ownership classifications suggests that these categories can be automated with higher confidence, requiring less intensive human review.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe results address the three research questions as follows. \u003cb\u003eRQ1\u003c/b\u003e (accuracy scaling and interpretable model viability) is addressed in Section \u003cspan refid=\"Sec30\" class=\"InternalRef\"\u003e4.1\u003c/span\u003e, where the learning curve analysis demonstrates that TF-IDF\u0026thinsp;+\u0026thinsp;SVM surpasses the zero-shot LLM baseline at approximately 496 labeled clauses. \u003cb\u003eRQ2\u003c/b\u003e (cross-domain generalization) is addressed in Section \u003cspan refid=\"Sec31\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e, where the asymmetry between successful SEC-to-Custom-built transfer (75.0\u0026ndash;78.12%) and failed LEDGAR-to-SEC transfer (10.2\u0026ndash;15.8%) establishes taxonomy alignment as the binding constraint. \u003cb\u003eRQ3\u003c/b\u003e (barriers to cross-domain and cross-lingual transfer) is addressed in Sections \u003cspan refid=\"Sec31\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e and \u003cspan refid=\"Sec32\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e, where the industrial validation (43.2%) and cross-lingual results (46.88% Korean) quantify the domain and language gaps that constrain practical deployment.\u003c/p\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e4.1 TF-IDF performance on structured legal text: the interpretability advantage\u003c/h2\u003e \u003cp\u003eThe central finding requires careful analysis: TF-IDF\u0026thinsp;+\u0026thinsp;SVM at 94.85% on SEC EDGAR is high because legal language is formulaic. License agreements follow standardized templates; phrases like \"shall indemnify and hold harmless,\" \"force majeure event,\" and \"net sales of Licensed Products\" appear with near-identical wording across hundreds of companies. TF-IDF is ideally suited to exploit this consistency -- it directly encodes the distinctive vocabulary patterns that define each clause type.\u003c/p\u003e \u003cp\u003eIn contrast, Gemma 3 4B's 81.3% on the custom-built set reflects a different strength: the LLM leverages semantic understanding from pre-training to classify clauses without any task-specific examples. But this broad understanding introduces noise on structured text where exact vocabulary match is more discriminative than semantic similarity.\u003c/p\u003e \u003cp\u003eFor legal practice, however, the comparison extends beyond accuracy. TF-IDF offers a critical property that neural models lack: full interpretability. When a TF-IDF\u0026thinsp;+\u0026thinsp;SVM classifier labels a clause as \"indemnification,\" a lawyer can inspect the top-weighted features and see that the decision was driven by terms such as \"liabilities,\" \"liability,\" \"hold harmless,\" and \"damages\" (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This transparency serves multiple functions in legal workflows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e1. \u003cb\u003eVerification\u003c/b\u003e: Attorneys can confirm that the classification is based on legally relevant terms rather than spurious correlations.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e2. \u003cb\u003eError detection\u003c/b\u003e: When a classification appears incorrect, examining the feature weights immediately reveals whether the error stems from ambiguous language, a drafting anomaly, or a genuine misclassification.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e3. \u003cb\u003eProfessional responsibility\u003c/b\u003e: The ability to explain why a clause was classified in a particular way supports the attorney's duty of competence under professional conduct rules.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e4. \u003cb\u003eClient communication\u003c/b\u003e: When presenting AI-assisted analysis to clients, attorneys can provide concrete explanations of how the tool reached its conclusions.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eNeural models, including LLMs, function as effective classifiers but cannot provide this level of decision-level transparency. While techniques such as attention visualization and LIME/SHAP approximations exist, they provide post-hoc explanations that may not faithfully reflect the model's actual decision process (Rudin, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). For legal applications where the explanation must be trustworthy -- not merely plausible -- inherently interpretable models have a structural advantage.\u003c/p\u003e \u003cp\u003eThe distinction between inherent and post-hoc interpretability is not merely academic in the legal context. When an attorney presents contract analysis to a client or in a regulatory proceeding, the credibility of the analysis depends on the reliability of its supporting reasoning. An inherently interpretable model -- where the decision mechanism is the explanation -- provides a qualitatively different form of accountability than a black-box model accompanied by an approximate explanation. This distinction aligns with Rudin's (2019) argument that high-stakes domains should prefer inherently interpretable models, and the legal domain is arguably the paradigmatic case for this principle.\u003c/p\u003e \u003cp\u003eTo illustrate the practical value of interpretability, consider a concrete misclassification from the SEC EDGAR evaluation. A clause containing \"The Licensor warrants that it has the right to grant...\" was classified as warranty rather than ip_ownership. Inspection of the TF-IDF feature weights immediately reveals the cause: the term \"warrants\" received high weight under the warranty category due to the dual meaning of \"warrant\" in SEC filings (stock warrants vs. contractual warranties). An attorney reviewing this output can instantly identify the ambiguity, correct the classification, and -- crucially -- understand *why* the error occurred. With a neural model, the same error would require post-hoc explanation techniques (e.g., LIME or SHAP) that may not reflect the model's actual decision process, leaving the attorney unable to determine whether the error is systematic or idiosyncratic.\u003c/p\u003e \u003cp\u003eThis has a clear practical implication: for structured, formulaic domains like legal contracts, investing in building labeled datasets for interpretable models may be more appropriate than deploying expensive, opaque LLMs. The learning curve analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) shows that TF-IDF surpasses the Gemma 3 4B zero-shot baseline (81.3%) at approximately 496 labeled clauses. For a legal department that processes hundreds of contracts annually, accumulating this volume of labeled examples through routine review operations is achievable within months, making the interpretable approach both practically viable and professionally preferable.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Cross-domain transfer: taxonomy alignment as the binding constraint\u003c/h2\u003e \u003cp\u003eThe cross-domain transfer results (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) reveal findings with significant implications for legal technology development. TF-IDF trained on SEC EDGAR achieves 75.0% on unseen custom-built clauses, while LEGAL-BERT achieves 78.12%, demonstrating generalization beyond the training corpus. This works because:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e1. Legal boilerplate language is standardized across document types and sources\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e2. The eight clause categories capture universal contract structures\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e3. Key discriminative terms (royalty, indemnify, terminate, force majeure) are consistent regardless of the underlying technology being licensed\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eConversely, the LEDGAR-to-SEC failure (10.2\u0026ndash;15.8% across all models) proves that data volume alone is insufficient -- label taxonomy alignment is the binding constraint.\u003c/p\u003e \u003cp\u003eThe industrial validation adds a critical third dimension to the transfer analysis. The 43.2% SEC\u0026rarr;Industrial accuracy -- substantially above random (12.5%) but far below in-domain (94.85%) -- quantifies the domain gap between public regulatory filings and proprietary industrial contracts. This gap arises from three sources: (1) the industrial agreements contain process-specific technical vocabulary (catalyst specifications, reactor design parameters) absent from SEC filings; (2) the drafting conventions of international process licensors differ from U.S. domestic norms; and (3) the natural class imbalance in the industrial corpus (only 2 force_majeure clauses vs. 31 ip_ownership) disadvantages the balanced-training classifier. For practitioners considering AI adoption, this result sets a realistic expectation: public data provides a useful starting point, but industrial deployment requires domain-specific labeled data. LEDGAR's native taxonomy, with over 100 clause types, maps imperfectly to our eight categories, producing noisy training signal that does not generalize.\u003c/p\u003e \u003cp\u003eThe 43.2% industrial accuracy warrants further analysis from a contract structure perspective. SEC EDGAR license agreements are predominantly embedded within 10-K/10-Q filings as exhibits and follow U.S. disclosure conventions: clauses are drafted with boilerplate language optimized for regulatory compliance rather than operational specificity. In contrast, the industrial petrochemical license agreements are operational contracts negotiated between international process licensors and an industrial end-user. Three structural differences drive the performance gap. First, operational contracts embed technical specifications directly within legal clauses -- for example, a royalty clause may reference \"catalyst replacement cycles\" or \"nameplate capacity,\" vocabulary entirely absent from SEC filings. Second, the drafting conventions of international EPC (Engineering, Procurement, and Construction) licensors reflect civil-law-influenced contract structures that differ from common-law boilerplate. Third, proprietary contracts frequently merge clause functions -- a single paragraph may simultaneously address indemnification and warranty obligations -- creating genuine multi-label instances that the single-label classifier cannot capture.\u003c/p\u003e \u003cp\u003eThis finding transforms the practical value of the system: organizations can leverage publicly available SEC filings to build models that transfer to other English-language contract collections, provided the label scheme is consistent. For legal practitioners and legal technology vendors, the implication is that standardizing label taxonomies across organizations and contract types is a prerequisite for building reusable classification tools. Efforts to develop standardized legal ontologies -- such as those proposed by the Legal Knowledge Interchange Format (LKIF) and related initiatives -- take on additional practical significance in this context.\u003c/p\u003e \u003cp\u003eThe taxonomy alignment finding also has implications for the legal technology market. Vendors building contract analysis platforms face a choice between developing proprietary taxonomies optimized for their specific use cases and adopting shared taxonomies that enable interoperability. Our results suggest that the latter approach has significant practical advantages: a standardized taxonomy enables models trained on public data (SEC filings) to transfer to proprietary contract collections, reducing the labeled data requirements for each new deployment. For the legal profession more broadly, the development of shared clause taxonomies -- analogous to the standardized coding systems used in healthcare (ICD codes) or accounting (GAAP categories) -- could substantially lower the barriers to legal AI adoption by enabling the pooling of training data across organizations while preserving the confidentiality of individual agreements.\u003c/p\u003e \u003cp\u003eThe asymmetry between SEC-to-Custom-built transfer (75.0\u0026ndash;78.12%) and LEDGAR-to-SEC transfer (10.2\u0026ndash;15.8%) deserves further analysis. The former succeeds because the SEC EDGAR dataset and the custom-built dataset share a common taxonomy designed for technology licensing, even though the documents differ in origin and style. The latter fails because LEDGAR's native taxonomy -- developed for general contract clause classification -- maps poorly to the technology-licensing-specific categories used in this study. This finding underscores the importance of domain-specific taxonomy design: a taxonomy developed for general contracts is not interchangeable with one developed for technology licensing, even when both are applied to the same underlying documents.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Implications for legal practice\u003c/h2\u003e \u003cp\u003eThe results of this study carry several implications for the integration of NLP tools into technology licensing workflows.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTriage and prioritization.\u003c/b\u003e The F1\u0026thinsp;=\u0026thinsp;0.948 on SEC EDGAR demonstrates that clause classification can be deployed as a first-pass triage tool, automatically sorting clauses by type so that attorneys can prioritize their review. In a portfolio review involving dozens of agreements, this capability can substantially reduce the time spent on initial document organization. Rather than reading each agreement from beginning to end, an attorney can use the classifier to identify and extract all indemnification clauses, all termination provisions, or all royalty terms across the entire portfolio for comparative analysis.\u003c/p\u003e \u003cp\u003e \u003cb\u003eRisk identification.\u003c/b\u003e Certain clause categories -- indemnification, termination, and force majeure in particular -- carry heightened risk implications in technology licensing. Automated identification of these clauses enables risk-focused review workflows where high-stakes provisions receive immediate attorney attention while lower-risk clauses (such as territory specifications that merely identify geographic scope) are reviewed on a secondary pass.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBenchmarking and negotiation support.\u003c/b\u003e By classifying clauses across a large corpus of comparable agreements, the system enables data-driven benchmarking. A licensing attorney negotiating an indemnification provision can quickly retrieve and compare indemnification clauses from similar transactions, providing empirical support for negotiation positions. This application leverages the cross-domain transfer capability: a model trained on SEC EDGAR filings can classify clauses in new, unseen agreements from similar domains.\u003c/p\u003e \u003cp\u003e \u003cb\u003eQuality assurance.\u003c/b\u003e Classification can serve as a completeness check, flagging agreements that are missing expected clause types. If a technology license agreement lacks an identifiable force majeure provision, for example, the classifier's failure to find one in any paragraph can alert the attorney to a potential gap. This application is particularly valuable for junior attorneys or non-specialists reviewing agreements outside their primary area of expertise.\u003c/p\u003e \u003cp\u003e \u003cb\u003eInstitutional knowledge preservation.\u003c/b\u003e TLOs and corporate legal departments accumulate significant institutional knowledge about clause drafting practices, negotiation outcomes, and risk patterns. A clause classification system, trained on an organization's historical agreements, can serve as a repository of institutional knowledge, enabling consistency in contract drafting and review even as personnel change.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTraining and professional development.\u003c/b\u003e For junior attorneys and technology transfer professionals, a clause classification system with interpretable outputs can serve as a training tool. By examining which terms the classifier identifies as discriminative for each clause type, trainees can develop an understanding of the vocabulary patterns that characterize different licensing provisions. The key term analysis in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, for instance, effectively summarizes the distinctive language of each clause category in a form that is directly educational.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDue diligence in technology M\u0026amp;A.\u003c/b\u003e In mergers and acquisitions involving technology-intensive companies, due diligence teams must review large volumes of existing license agreements to assess the target company's IP rights and obligations. The clause classification system, particularly its cross-domain transfer capability, can accelerate this review by automatically organizing clause-level information from heterogeneous agreement collections. The 75.0\u0026ndash;78.12% cross-domain transfer accuracy, while not sufficient for autonomous analysis, provides a useful starting point that reduces the volume of unstructured text attorneys must process.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Comparison with prior work\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison with related legal NLP benchmarks\u003c/p\u003e \u003cdiv class=\"Credit\"\u003e\u003cp\u003e(adapted from published results; self-compiled)\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTask\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBest model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePerformance\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChalkidis et al. (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2020\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLegal text classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eVarious legal NLP tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLEGAL-BERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003emicro-F1: 0.82\u0026ndash;0.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHendrycks et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExtractive QA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCUAD (510 contracts)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBERT-based\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1: 0.60\u0026ndash;0.75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTuggener et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClause classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLEDGAR (100\u0026thinsp;+\u0026thinsp;types)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eVarious\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1: ~0.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eElwany et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2019\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParagraph classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eProcurement contracts\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBERT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1: \u0026gt;0.90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThis study\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClause classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSEC EDGAR (2,000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTF-IDF\u0026thinsp;+\u0026thinsp;SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003emacro-F1: 0.948\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eOur TF-IDF\u0026thinsp;+\u0026thinsp;SVM achieves F1\u0026thinsp;=\u0026thinsp;0.948, likely due to the formulaic nature of license agreements providing stronger discriminative features than general legal text. The comparison is instructive: LEGAL-BERT's 0.82\u0026ndash;0.87 was measured on heterogeneous legal classification tasks, while our higher performance reflects the more structured and predictable vocabulary of technology licensing provisions. Hendrycks et al.'s (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) lower F1 of 0.60\u0026ndash;0.75 on CUAD reflects the substantially greater difficulty of extractive question answering compared to classification.\u003c/p\u003e \u003cp\u003eIt is important to note that this comparison does not imply that TF-IDF\u0026thinsp;+\u0026thinsp;SVM is generally superior to LEGAL-BERT or other transformer-based models. Rather, it demonstrates that for a specific document type with highly consistent vocabulary -- technology license agreements -- the simpler model captures the relevant signal with remarkable effectiveness. The practical significance of this finding for legal technology development is that not every legal NLP application requires a large pre-trained model; the choice of model should be guided by the characteristics of the target document type, the availability of labeled data, and the interpretability requirements of the deployment context.\u003c/p\u003e \u003cp\u003eElwany et al.'s (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) result on procurement contracts (F1\u0026thinsp;\u0026gt;\u0026thinsp;0.90 with BERT) provides perhaps the most direct comparison, as procurement contracts share the standardized, formulaic character of technology license agreements. Our higher F1 with a simpler model suggests that the degree of vocabulary consistency in technology licensing may exceed even that of procurement contracts, making this document type particularly amenable to frequency-based classification.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Practical deployment considerations for technology transfer management\u003c/h2\u003e \u003cp\u003eThe system's performance characteristics suggest specific deployment scenarios for TLOs and corporate licensing departments:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e1. \u003cb\u003eAutomated contract triage\u003c/b\u003e: Classifying clauses in new drafts to detect missing clause types. At F1\u0026thinsp;=\u0026thinsp;0.948, the false negative rate is sufficiently low for this application.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e2. \u003cb\u003eRisk analysis pipeline\u003c/b\u003e: Identifying high-risk clauses (indemnification, termination, force majeure) for prioritized review. The per-class F1 for these categories (0.929, 0.944, and 0.968 respectively) supports reliable automated flagging.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e3. \u003cb\u003eCross-agreement benchmarking\u003c/b\u003e: Comparing identical clause types across agreements to evaluate term fairness. The cross-domain transfer result (75.0\u0026ndash;78.12% on unseen clauses) demonstrates that a single trained model can process heterogeneous agreement collections.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e4. \u003cb\u003eTLO workflow optimization\u003c/b\u003e: Reducing initial review time from hours to minutes. For a TLO managing 100\u0026thinsp;+\u0026thinsp;active agreements, even partial automation of the initial classification step can free significant attorney time for higher-value analysis.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eA key practical consideration is data confidentiality. Technology license agreements frequently contain commercially sensitive information that organizations may be unwilling to transmit to cloud-based AI services. Both the TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline and the Gemma 3 4B model used in this study run entirely on local infrastructure, requiring no external data transmission. This on-premise deployment capability may be a decisive factor for risk-averse legal departments.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCost-effectiveness for resource-constrained organizations.\u003c/b\u003e The computational requirements of the TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline are minimal -- training on 2,000 clauses completes in approximately 45 seconds on consumer hardware, and inference is essentially instantaneous. This stands in stark contrast to large language model approaches that require specialized GPU infrastructure or expensive API subscriptions. For university TLOs and SMEs -- precisely the organizations most in need of automated contract review support due to limited legal staff -- the low cost of deploying a TF-IDF-based system removes a significant barrier to adoption. The small LLM alternative (Gemma 3 4B) requires more computational resources but remains deployable on a standard workstation with a consumer-grade GPU, avoiding the need for cloud computing expenditure.\u003c/p\u003e \u003cp\u003e \u003cb\u003eIncremental adoption pathway.\u003c/b\u003e The results suggest a natural adoption pathway for organizations new to legal AI. An organization can begin with the zero-shot LLM approach, which requires no labeled data and provides reasonable accuracy (81.3%) on small clause collections. As the organization accumulates labeled examples through routine contract review -- with attorneys confirming or correcting automated classifications -- the labeled dataset grows organically. Once the dataset reaches the 500\u0026ndash;600 clause threshold, the organization can transition to the TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline, gaining both higher accuracy and full interpretability. This incremental pathway reduces the upfront investment required for AI adoption while building organizational expertise in legal AI governance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Professional responsibility and ethical considerations\u003c/h2\u003e \u003cp\u003eThe deployment of AI tools for contract clause classification raises professional responsibility questions that legal practitioners must address.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDuty of competence and technological literacy.\u003c/b\u003e The ABA Model Rules, as amended in 2012, require that attorneys maintain competence in \"the benefits and risks of relevant technology\" (Rule 1.1, Comment 8). For attorneys using clause classification tools, this duty implies an obligation to understand the tool's capabilities and limitations -- including its accuracy rates, its known failure modes (such as the territory-exclusivity confusion documented in our error analysis), and the conditions under which its performance may degrade (such as cross-lingual application or novel clause types).\u003c/p\u003e \u003cp\u003e \u003cb\u003eSupervision obligations.\u003c/b\u003e Under Model Rule 5.1 (Responsibilities of Partners, Managers, and Supervisory Lawyers), attorneys with supervisory responsibility must ensure that AI-assisted work is properly reviewed. The interpretability of TF-IDF-based classification directly supports this obligation: a supervising attorney can review not only the classification output but also the reasoning (feature weights) that produced it, enabling meaningful oversight rather than rubber-stamping.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAutomation bias.\u003c/b\u003e Psychological research on automation bias -- the tendency of humans to over-rely on automated recommendations -- poses a specific risk in AI-assisted legal review. If attorneys come to trust the classifier's outputs without independent verification, they may fail to catch errors that could have material consequences for their clients. This risk is amplified for high-performing systems: a classifier with 94.85% accuracy may engender more complacency than a mediocre one, precisely because its errors are rare and therefore unexpected.\u003c/p\u003e \u003cp\u003eMitigation strategies include: (a) presenting classifications with confidence scores rather than binary labels, so that borderline cases are flagged for manual review; (b) periodically introducing known errors as \"calibration checks\" to maintain attorney vigilance; and (c) designing user interfaces that require attorneys to affirmatively confirm or override each classification rather than passively accepting defaults.\u003c/p\u003e \u003cp\u003e \u003cb\u003eLiability allocation.\u003c/b\u003e When an AI-assisted classification error leads to a substantive legal mistake -- for example, if a clause is misclassified and consequently overlooked during review -- questions of liability arise. Under current malpractice frameworks, the attorney remains responsible for the quality of their work product regardless of whether it was produced with AI assistance. This allocation of liability reinforces the importance of meaningful human oversight and argues against fully automated \"lights-out\" contract review workflows.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBias in training data.\u003c/b\u003e The SEC EDGAR dataset reflects the clause drafting conventions of companies that file with the U.S. Securities and Exchange Commission -- predominantly large, publicly traded corporations with access to sophisticated legal counsel. Agreements drafted by smaller organizations, startups, or international parties may use different language conventions that the model has not encountered. Practitioners must be aware that the classifier's high accuracy was measured on a specific population of agreements and may not generalize equally to all technology licensing contexts.\u003c/p\u003e \u003cp\u003e \u003cb\u003eConfidentiality and data governance.\u003c/b\u003e Technology license agreements typically contain commercially sensitive terms -- royalty rates, territory restrictions, exclusivity grants -- that are protected by confidentiality obligations. The use of agreement text for training AI models raises questions about whether such use is consistent with confidentiality provisions in the agreements themselves. Organizations deploying clause classification systems must ensure that their training data governance practices comply with applicable confidentiality obligations. The on-premise deployment architecture demonstrated in this study -- where both TF-IDF and Gemma 3 4B run locally without transmitting data to external servers -- mitigates but does not eliminate this concern, as the trained model itself may encode information about the training documents.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInformed consent\u003c/strong\u003e \u003cp\u003e \u003cb\u003eand client communication.\u003c/b\u003e When attorneys use AI tools to assist with contract review, questions arise about the extent of disclosure required to clients. While no jurisdiction currently mandates specific disclosure of AI tool usage in contract review, the trend toward greater transparency about AI assistance in legal services suggests that practitioners should proactively inform clients about the role of automated tools in their work product. The interpretability of TF-IDF-based classification facilitates this disclosure by enabling attorneys to explain in concrete terms how the tool works and what role it played in the analysis.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec36\" class=\"Section2\"\u003e \u003ch2\u003e4.7 Limitations\u003c/h2\u003e \u003cp\u003eSeveral limitations of this study should be acknowledged.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSingle-label classification.\u003c/b\u003e Real contract clauses frequently serve multiple legal functions simultaneously -- for example, a royalty-upon-termination provision addresses both royalty and termination, and an indemnification clause may incorporate warranty language. The current single-label framework assigns exactly one category per clause, which forces an artificial classification decision for such multi-functional provisions. This limitation affects both accuracy measurement (a clause labeled \"indemnification\" when it also functions as \"warranty\" is counted as an error on the warranty dimension) and practical utility (an attorney relying on single-label output may miss secondary functions). The error analysis (Section \u003cspan refid=\"Sec28\" class=\"InternalRef\"\u003e3.8\u003c/span\u003e) confirms that the most common misclassification pairs -- territory/exclusivity, indemnification/warranty -- correspond precisely to categories with genuine functional overlap in legal drafting.\u003c/p\u003e \u003cp\u003e \u003cb\u003eGemma 3 4B not tested at scale.\u003c/b\u003e The LLM was evaluated only on 64 clauses; inference at 3.8s/clause makes 2,000-clause evaluation feasible but was not performed. A full-scale LLM evaluation would provide a more complete comparison.\u003c/p\u003e \u003cp\u003e \u003cb\u003eNo fine-tuned transformer baseline in the main comparison.\u003c/b\u003e The in-domain evaluation compares TF-IDF\u0026thinsp;+\u0026thinsp;SVM against a zero-shot LLM, but does not include fine-tuned BERT or LEGAL-BERT as an intermediate baseline. BERT and LEGAL-BERT appear only in the cross-domain transfer experiments (Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e), where LEGAL-BERT achieves 78.12% on SEC\u0026rarr;Custom-built transfer. Fine-tuning LEGAL-BERT on the full SEC EDGAR training set would likely yield accuracy between TF-IDF (94.85%) and zero-shot Gemma (81.3%), narrowing the apparent gap. This omission limits our ability to characterize the full accuracy-interpretability tradeoff curve and is a priority for future work. Nevertheless, the central argument of this study -- that interpretability provides a structural advantage for legal deployment -- holds regardless of the fine-tuned transformer's accuracy, because the interpretability gap between TF-IDF and any neural model remains qualitative rather than quantitative.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAbsence of Korean-specific model.\u003c/b\u003e Korean legal NLP requires specialized pre-training that was not available for this study. The poor Korean performance (46.88%) limits the system's applicability to international technology transfer transactions involving Korean-language agreements.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTemporal generalization.\u003c/b\u003e All data is from a fixed period; contractual language evolves over time, and the model's performance on future agreements is not guaranteed.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTaxonomy coverage.\u003c/b\u003e The taxonomy covers 8 categories selected for their centrality to technology licensing negotiations, but excludes commonly occurring clause types such as confidentiality, governing law, dispute resolution, audit rights, and assignment. These omissions mean that the system addresses a targeted subset of contract review tasks rather than comprehensive contract automation. A production-grade system would require at minimum 10\u0026ndash;12 categories. The results reported here should therefore be interpreted as demonstrating the feasibility and characteristics of clause classification for technology licensing, not as a complete solution for automated contract review.\u003c/p\u003e \u003cp\u003e \u003cb\u003eClause segmentation.\u003c/b\u003e Clause segmentation is assumed pre-done; in production, sentence/paragraph boundary detection would introduce additional error.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCompany-level leakage.\u003c/b\u003e Multiple clauses from the same company may appear in both train and test folds in the SEC EDGAR cross-validation, potentially inflating within-dataset accuracy due to shared drafting templates or law firm style. A company-level split -- where all clauses from a given company appear exclusively in either the training or test fold -- would provide a more conservative and externally valid estimate. We note, however, that the cross-domain transfer experiments (SEC\u0026rarr;Custom-built at 75.0\u0026ndash;78.12%, SEC\u0026rarr;industrial at 43.2%) inherently avoid this leakage by testing on entirely separate document collections, and the consistency between in-domain and cross-domain results suggests that company-level leakage, while present, is unlikely to account for the majority of the in-domain performance.\u003c/p\u003e \u003cp\u003e \u003cb\u003eLegal jurisdiction specificity.\u003c/b\u003e The datasets are drawn predominantly from U.S. legal instruments (SEC filings, U.S. university TLO templates). The clause language conventions, and therefore the classifier's effectiveness, may differ substantially for agreements governed by other legal systems -- particularly civil law jurisdictions where contract drafting conventions differ from common law traditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAbsence of practitioner evaluation.\u003c/b\u003e While the taxonomy was validated by licensing professionals, the classification system itself has not been evaluated in a real-world legal workflow. User studies with attorneys performing actual contract review tasks would provide more ecologically valid evidence of the system's practical utility.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec37\" class=\"Section2\"\u003e \u003ch2\u003e4.8 Future work\u003c/h2\u003e \u003cp\u003eSeveral directions for future research emerge from this study:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e1. \u003cb\u003eMulti-label classification\u003c/b\u003e with binary relevance or classifier chains, enabling clauses that span multiple categories to receive appropriate multi-category labels.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e2. \u003cb\u003eFine-tuning BERT/LEGAL-BERT\u003c/b\u003e on SEC EDGAR for direct comparison with TF-IDF, establishing where the accuracy-interpretability tradeoff becomes favorable for neural models.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e3. \u003cb\u003eKorean legal language model development\u003c/b\u003e, addressing the cross-lingual limitation that currently prevents application to Korean-language agreements -- a pressing need given the volume of technology licensing activity in East Asia.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e4. \u003cb\u003eExpanding taxonomy\u003c/b\u003e to 10\u0026ndash;12 categories (adding confidentiality, governing_law, dispute_resolution, audit) for more comprehensive contract coverage.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e5. \u003cb\u003eEnd-to-end contract risk scoring\u003c/b\u003e integration, combining clause classification with risk assessment to produce agreement-level risk profiles for portfolio management.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e6. \u003cb\u003ePractitioner evaluation studies\u003c/b\u003e involving attorneys performing contract review tasks with and without AI assistance, measuring both efficiency gains and error rates to validate the system's practical utility in legal workflows.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e7. \u003cb\u003eHuman-in-the-loop interface design\u003c/b\u003e exploring how classification outputs should be presented to maximize attorney oversight while minimizing review time, including confidence calibration and explanation presentation.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e8. \u003cb\u003eLongitudinal analysis\u003c/b\u003e of how clause language evolves over time in technology licensing, and whether models trained on historical agreements maintain accuracy on contemporary contracts.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e9. \u003cb\u003eCross-jurisdictional extension\u003c/b\u003e to civil law and hybrid legal systems, investigating whether the formulaic nature of contract language that enables high TF-IDF performance in common law agreements extends to other legal traditions.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study systematically evaluated NLP-based clause classification for technology license agreements across three dataset scales, with a focus on the implications for legal practice and AI-assisted contract review. On the SEC EDGAR dataset of 2,000 clauses from 812 companies, TF-IDF\u0026thinsp;+\u0026thinsp;SVM achieved 94.85% accuracy (macro F1\u0026thinsp;=\u0026thinsp;0.948), with per-class F1 ranging from 0.925 (territory) to 0.968 (ip_ownership, force_majeure). On LEDGAR 600 clauses, accuracy was 91.67% (F1\u0026thinsp;=\u0026thinsp;0.888). A small LLM (Gemma 3 4B) achieved 81.3% zero-shot on 64 English clauses, demonstrating utility in data-scarce settings.\u003c/p\u003e \u003cp\u003eThe key empirical finding is cross-domain transfer: a model trained on SEC regulatory filings achieves 75.0\u0026ndash;78.12% on unseen custom-built clauses (with LEGAL-BERT outperforming TF-IDF), while LEDGAR-to-SEC transfer fails for all models (10.2\u0026ndash;15.8%), confirming that label taxonomy alignment is more critical than data volume for building reusable legal classification tools. Industrial validation on 50 real petrochemical plant license agreements [company anonymized for review] yields 43.2% cross-domain accuracy, quantifying the domain gap between public benchmarks and proprietary industrial contracts. This has direct implications for legal technology development: standardized clause taxonomies are a prerequisite for interoperable contract analysis tools across organizations.\u003c/p\u003e \u003cp\u003eFrom the perspective of legal practice, the TF-IDF approach offers an advantage that goes beyond accuracy: full interpretability. When attorneys rely on AI tools for contract review, they retain professional responsibility for the accuracy and completeness of their work. The ability to inspect exactly which terms drove each classification decision -- a property inherent to TF-IDF but absent from neural models -- supports the attorney's duty of competence and enables meaningful human oversight of automated classifications.\u003c/p\u003e \u003cp\u003eFor legal organizations considering adoption, the results suggest a concrete workflow: (1) begin with zero-shot LLM classification for immediate utility without labeled data (81.3% accuracy); (2) accumulate labeled examples through routine attorney review, with attorneys confirming or correcting automated outputs; (3) transition to TF-IDF\u0026thinsp;+\u0026thinsp;SVM once approximately 500 labeled clauses are available, gaining both higher accuracy (94.85%) and full interpretability. The practical boundary for this approach is well-defined: it works reliably for English-language, standardized contracts within the training taxonomy, but degrades for cross-lingual scenarios (46.88% on Korean-only), novel clause types, and proprietary industrial agreements with domain-specific technical vocabulary (43.2% on petrochemical contracts). Future work must address these boundaries through multilingual models, expanded taxonomies, and practitioner evaluation studies.\u003c/p\u003e \u003cp\u003eMore broadly, this study contributes to the ongoing discourse on the role of AI in legal practice by demonstrating a case where the simplest adequate technology may also be the most professionally appropriate one. The legal profession's emphasis on accountability, transparency, and human judgment is not merely a barrier to AI adoption -- it is a set of substantive values that should shape the design of legal AI systems. Clause classification systems that are interpretable by design, rather than by post-hoc approximation, align naturally with these values.\u003c/p\u003e \u003cp\u003eThe tension between classification accuracy and interpretability will shape the trajectory of legal AI adoption. As the field progresses toward more powerful but less transparent models, maintaining the ability of legal professionals to understand, verify, and take responsibility for AI-assisted decisions will remain a central challenge. Our results suggest that for the specific task of technology license agreement clause classification -- and likely for other structured legal document types -- this tension can be resolved in favor of both accuracy and interpretability, provided that sufficient domain-specific labeled data is available. The practical path forward for legal organizations is to invest in building such labeled datasets through their routine operations, thereby enabling high-performance interpretable classification that satisfies both the accuracy requirements of effective AI tools and the transparency requirements of professional legal practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe author declares no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eNo external funding was received for this research.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eY.K. designed the study, developed the methodology and software, collected and curated all datasets, performed all experiments and analyses, created all figures, and wrote the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe SEC EDGAR and LEDGAR datasets were constructed from publicly available sources. The industrial dataset cannot be shared due to confidentiality obligations, but the evaluation protocol and aggregate results are reported. Code and data processing scripts are available at [repository URL to be provided upon acceptance].\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eABA (American Bar Association) (2023) Model Rules of Professional Conduct, Rule 1.1: Competence (including Comment 8 on technology). American Bar Association\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAriai A, Jungiewicz M, Schilder F (2025) NLP for the legal domain: A survey of tasks, datasets, models, and challenges. *arXiv preprint arXiv:2501.01439*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAUTM (Association of University Technology Managers) (2023) *AUTM Licensing Activity Survey. FY2022*. AUTM\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBekkers R, Bongard R, Nuvolari A (2014) An empirical study on the determinants of essential patent claims in standards. *Research Policy* 43(7):1296\u0026ndash;1311\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBommarito MJ, Katz DM (2021) GPT takes the bar exam. *arXiv preprint arXiv:2312.00914*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBozeman B (2000) Technology transfer and public policy: A review of research and theory. *Research Policy* 29(4\u0026ndash;5):627\u0026ndash;655\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChalkidis I, Fergadiotis M, Malakasiotis P, Aletras N, Androutsopoulos I (2020) LEGAL-BERT: The muppets straight out of law school. In: *Findings of the Association for Computational Linguistics: EMNLP 2020*, pp 2898\u0026ndash;2904\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChalkidis I, Fergadiotis M, Androutsopoulos I (2021) MultiEURLEX -- A multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer. In: *Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing*, pp 6974\u0026ndash;6996\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi JH, Monahan A, Schwarcz D (2024) Lawyering in the age of artificial intelligence. *Minnesota Law Review* 109\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeeks A (2019) The judicial demand for explainable artificial intelligence. *Columbia Law Review* 119(7):1829\u0026ndash;1850\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDevlin J, Chang MW, Lee K, Toutanova K (2019) BERT: Pre-training of deep bidirectional transformers for language understanding. In: *Proceedings of NAACL-HLT*, pp 4171\u0026ndash;4186\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElwany E, Moore D, Oberoi G (2019) BERT goes to law school: Quantifying the competitive advantage of access to large legal corpora in contract understanding. In: *Workshop on Document Intelligence at NeurIPS 2019*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoogle DM (2025) Gemma 3 technical report. *arXiv preprint arXiv:2503.19786*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHendrycks D, Burns C, Chen A, Ball S (2021) CUAD: An expert-annotated NLP dataset for legal contract understanding. In: *NeurIPS Datasets and Benchmarks Track*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJoachims T (1998) Text categorization with support vector machines: Learning with many relevant features. In: *Proceedings of ECML*, pp 137\u0026ndash;142\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaplan J et al (2020) Scaling laws for neural language models. *arXiv preprint arXiv:2001.08361*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeivaditi S, Rossi J, Kamps J (2020) A benchmark for lease contract review. In: *Workshop on Scholarly Document Processing*, pp 1\u0026ndash;10\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNiklaus J, Chalkidis I, Sturmer M (2023) MultiLegalPile: A 689 GB multilingual legal corpus. *arXiv preprint arXiv:2306.02069*\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePerlman A (2024) The implications of ChatGPT for legal services and society. *Michigan Technol Law Review* 30(1):1\u0026ndash;16\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRazgaitis R (2009) *Valuation and Dealmaking of Technology-Based Intellectual Property*, 2nd edn. Wiley\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRemus D, Levy FS (2017) Can robots be lawyers? Computers, lawyers, and the practice of law. *Georgetown J Legal Ethics* 30:501\u0026ndash;558\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRudin C (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. *Nature Mach Intelligence* 1(5):206\u0026ndash;215\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh A, Joshi A, Jiang J, Paik HY (2025) A survey of classification tasks and approaches for legal contracts. *Artificial Intell Review*. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10462-025-11359-8\u003c/span\u003e\u003cspan address=\"10.1007/s10462-025-11359-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSEC (U.S. Securities and Exchange Commission) (2024) EDGAR full-text search system. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://efts.sec.gov/LATEST/search-index\u003c/span\u003e\u003cspan address=\"https://efts.sec.gov/LATEST/search-index\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuh JH (2016) Technology licensing: An empirical analysis of the role of contract clauses. *Research Policy* 45(6):1221\u0026ndash;1232\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSurden H (2019) Artificial intelligence and law: An overview. *Georgia State Univ Law Review* 35(4):1305\u0026ndash;1337\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTuggener D, von Daniken C, Peber T, Cieliebak M (2020) LEDGAR: A large-scale multi-label corpus for text classification of legal provisions in contracts. In: *Proceedings of the 12th Language Resources and Evaluation Conference (LREC 2020)*, pp 1235\u0026ndash;1241\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhong H, Xiao C, Tu C, Zhang T, Liu Z, Sun M (2020) How does NLP benefit legal system: A summary of legal artificial intelligence. In: *Proceedings of ACL*, pp 5218\u0026ndash;5230\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"legal NLP, contract clause classification, technology licensing, interpretability, cross-domain transfer, TF-IDF","lastPublishedDoi":"10.21203/rs.3.rs-9298439/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9298439/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAutomated clause classification in technology license agreements remains underdeveloped despite its importance to legal practice. This study compares a TF-IDF\u0026thinsp;+\u0026thinsp;SVM pipeline and a 4-billion-parameter small language model (Gemma 3 4B) across three datasets: 128 custom-built clauses from WIPO model contracts, 600 LEDGAR clauses, and 2,000 SEC EDGAR clauses spanning 812 companies. Eight clause categories central to licensing practice are targeted: royalty, territory, exclusivity, termination, ip_ownership, indemnification, warranty, and force_majeure. TF-IDF\u0026thinsp;+\u0026thinsp;SVM achieves 94.85% accuracy (macro F1\u0026thinsp;=\u0026thinsp;0.948) on SEC EDGAR with 5-fold cross-validation, while Gemma 3 4B achieves 81.3% zero-shot on 64 English clauses without training data. Cross-domain transfer from SEC EDGAR to unseen clauses reaches 75.0%, but LEDGAR-to-SEC transfer yields only 10.2%, suggesting that taxonomy alignment matters more than data volume. Industrial validation on 270 clauses from 50 petrochemical license agreements yields 43.2%, revealing a substantial domain gap between regulatory filings and proprietary contracts. From a legal practice perspective, TF-IDF offers a decisive advantage: full interpretability. Attorneys can inspect exactly which terms drove each classification, supporting the professional responsibility requirement for independent judgment when using AI-assisted tools. We discuss implications for responsible deployment, including human-in-the-loop design and the tension between accuracy and the explanatory demands of legal practice. For Technology Licensing Offices (TLOs) and corporate legal departments, the learning curve analysis suggests that approximately 500 labeled clauses -- achievable through 3\u0026ndash;6 months of routine contract review -- suffice for TF-IDF to surpass zero-shot LLM performance, providing an actionable adoption pathway for resource-constrained organizations.\u003c/p\u003e","manuscriptTitle":"Interpretable NLP for Technology License Agreement Review: Clause Classification with TF-IDF and Small LLM across SEC, LEDGAR, and Industrial Datasets","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 13:14:20","doi":"10.21203/rs.3.rs-9298439/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"34e6b158-d5cf-419d-b1cf-2560b7c629d5","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T13:14:20+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 13:14:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9298439","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9298439","identity":"rs-9298439","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}