Evolving Compliance Processes through LLM-Supported Transformation of Regulations into SBVR

Evolving Compliance Processes through LLM-Supported Transformation of Regulations into SBVR | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 15 October 2025 V1 Latest version Share on Evolving Compliance Processes through LLM-Supported Transformation of Regulations into SBVR Authors : Anderson dos Santos 0000-0001-9254-7764 [email protected] and Paulo Sérgio Muniz Silva Authors Info & Affiliations https://doi.org/10.22541/au.176051910.04955057/v1 347 views 132 downloads Contents Abstract 1 INTRODUCTION 2 CONCEPTUAL FOUNDATIONS 3 RELATED WORKS 4 A BROAD VIEW OF THE PROPOSED METHOD FOR TRANSFORMATION 4.1 PRELIMINARY EVALUATION 5 CONCLUSION AND FUTURE WORK References Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Compliance processes in software systems face persistent challenges from regulatory texts that are complex, ambiguous, and frequently updated. Traditional Governance, Risk, and Compliance (GRC) approaches often rely on manual analysis or deterministic algorithms, which limit adaptability and compromise traceability between legal requirements and organizational practices. This work proposes an automatic method to transform natural-language regulations into business vocabularies and rules expressed in the Semantics of Business Vocabulary and Business Rules (SBVR) in Structured English (SE). The method combines semantic structuring, taxonomy-based classification, and the use of Large Language Models (LLMs) for disambiguation and rule creation. The fidelity of the transformation is assessed through semantic textual similarity, ensuring that the resulting rules accurately preserve the original provisions’ intent. A preliminary evaluation, limited to the identification of elements, demonstrates feasibility; however, full validation of all steps remains a task for future research. By integrating SBVR with LLM-based interpretation, the method supports the maintainability and evolution of compliance processes while enabling experts to review the fidelity of transformations without requiring technical knowledge of SBVR. This contribution advances regulatory compliance by improving consistency, scalability, and automation in software systems that must adapt to evolving legal frameworks. Evolving Compliance Processes through LLM-Supported Transformation of Regulations into SBVR MSCS Anderson dos Santos * , Dr. Paulo Sérgio Muniz Silva † ABSTRACT Compliance processes in software systems face persistent challenges from regulatory texts that are complex, ambiguous, and frequently updated. Traditional Governance, Risk, and Compliance (GRC) approaches often rely on manual analysis or deterministic algorithms, which limit adaptability and compromise traceability between legal requirements and organizational practices. This work proposes an automatic method to transform natural-language regulations into business vocabularies and rules expressed in the Semantics of Business Vocabulary and Business Rules (SBVR) in Structured English (SE). The method combines semantic structuring, taxonomy-based classification, and the use of Large Language Models (LLMs) for disambiguation and rule creation. The fidelity of the transformation is assessed through semantic textual similarity, ensuring that the resulting rules accurately preserve the original provisions’ intent. A preliminary evaluation, limited to the identification of elements, demonstrates feasibility; however, full validation of all steps remains a task for future research. By integrating SBVR with LLM-based interpretation, the method supports the maintainability and evolution of compliance processes while enabling experts to review the fidelity of transformations without requiring technical knowledge of SBVR. This contribution advances regulatory compliance by improving consistency, scalability, and automation in software systems that must adapt to evolving legal frameworks. KEYWORDS: SBVR, GRC, CFR, FRO, FIBO, Taxonomy, NLP, LLM. 1 INTRODUCTION Financial institutions are subject to increasingly dense, dynamic, and open-to-interpretation regulatory frameworks, which complicate compliance activities. Governance, Risk, and Compliance (GRC) practices continue to rely mainly on reactive strategies, with audits typically occurring only after noncompliance is identified. These practices are still largely manual and prone to errors (Sadiq et al., 2007; Kulkarni et al., 2021; Roychoudhury et al., 2017; Sunkle et al., 2015). In addition, regulatory interpretation often relies on external specialists, which requires significant time and financial resources. Linguistic properties of regulations, such as ambiguity, polysemy, and context dependence, further complicate compliance verification (Hough & Gluck, 2019; Jackson, 2020), increasing the need for precise traceability between legal requirements and organizational processes (Bouzidi et al., 2011). Advances in Natural Language Processing (NLP) have been applied to mitigate these challenges and improve the analysis of regulatory texts (Khurana et al., 2023). For software-intensive organizations, these challenges translate directly into difficulties in maintaining and evolving compliance processes within their systems and services. Manual and deterministic approaches to compliance limit scalability and hinder adaptability when regulations change, leading to high maintenance costs and limited auditability. This connects compliance directly with issues of software process improvement, maintainability, and evolution, where automation and formalization of rules become essential. Controlled Natural Languages (CNLs), such as the Semantics of Business Vocabulary and Business Rules - Structured English (SBVR-SE), enable the representation of rules with semantic clarity while ensuring interoperability across information systems (Gruzitis & Barzdins, 2009; OMG, 2019). Several methods have been proposed for transforming business rules. Abi-Lahoud et al. (2013) suggest a process centered on experts. Joshi and Saha (2020) utilize artificial intelligence to structure the Code of Federal Regulations (CFR) as a knowledge graph. Haj et al. (2021) present a method for automatically converting explicit rules into SBVR using NLP; however, their approach struggles with managing ambiguity and complex sentence structures. In general, existing methods depend on deterministic algorithms, which restrict scalability and adaptability, as discussed in Section 3. This work presents a method that applies NLP techniques supported by Large Language Models (LLMs) to automatically transform and validate regulatory provisions from the CFR, expressed in natural language, into structured vocabularies and rules in Structured English (SE), following the SBVR specification defined by the Object Management Group (OMG). The contribution of this work lies in introducing a methodology that employs SBVR-SE to formalize financial regulations, while allowing Subject Matter Experts (SMEs) to participate in the validation process. By expressing rules in CNL and embedding metadata for validation and traceability, the approach enables experts to confirm the fidelity of transformations without requiring technical knowledge of SBVR. In doing so, it contributes to the evolution of compliance processes by improving consistency, maintainability, and automation in software systems that must adapt to dynamic regulatory environments. At the current stage of this research, the evaluation of the method is limited to a preliminary experiment focused on identifying elements, which provides initial evidence of feasibility but does not yet encompass the entire transformation pipeline. A comprehensive validation across all steps, including classification, transformation, and semantic similarity assessment, is left for ongoing research11Code and other materials could be found at the author’s repository at [Retrieved October 11, 2025] https://github.com/asantos2000/master-degree-santos-anderson. The work is structured as follows. Section 2 discusses SBVR, CNL, taxonomy, ontologies, LLMs, and semantic similarity as conceptual foundations. Section 3 examines several prior approaches to identifying, extracting, and transforming business rules into CNL. Section 4 outlines the proposed method for automating the transformation of business rules into SBVR. Section 5 concludes with reflections and directions for future research. 2 CONCEPTUAL FOUNDATIONS This work is grounded in four conceptual pillars. The first is SBVR and CNLs , which offer an unambiguous way of representing business rules. The second is taxonomies and ontologies , which provide systematic classification and semantic structuring. The third is the use of LLMs , which enables the classification and transformation of natural language text into SBVR through advanced NLP and text generation techniques. The fourth is semantic similarity validation , which helps SMEs evaluate transformations, particularly given the scarcity and high cost of annotated datasets. Together, these foundations enable the transformation of regulatory texts written in natural language into structured vocabularies and business rules that are both interpretable by machines and verifiable by experts. 2.1 SBVR AND CNL The Object Management Group’s (OMG, 2019) SBVR specification provides a formal and linguistic framework for representing vocabularies and business rules, ensuring semantic clarity and interoperability between organizations and software tools. The specification introduces principles such as applicability across business domains, defined compliance levels for software and documents, and the use of the semiotic triangle to distinguish meaning from representation. It classifies meanings into concepts and propositions, and statements into designations and formulations. SBVR also establishes links between concepts, instances, and states of affairs, enabling propositions to describe real or potential situations and allowing business rules to be directly applied to organizational operations. SBVR differentiates the terminological dictionary, which documents and organizes terms with precise definitions and relationships, from the business vocabulary, which gathers the terms effectively used in daily processes and communications, ensuring clarity and consistency. SBVR defines a metamodel that is both human-readable and machine-processable. To achieve this, it adopts a CNL derived from natural language with restricted grammar and vocabulary, with structured English recommended by OMG. Although the examples in the specification are not prescriptive, they illustrate how to build a CNL aligned with SBVR’s semantic structures. A Controlled Natural Language (CNL) is a deliberately limited subset of a natural language that restricts vocabulary, grammar, and style to minimize ambiguity while remaining easy for humans to understand and process by machines (Angelov & Ranta, 2010). These restrictions are customized to specific domain goals, such as clearly defining business rules in governance (Spreeuwenberg & Healy, 2010), verifying mathematical proofs (Cramer et al., 2009), supporting accurate legal contracts (Pace & Rosner, 2010), or organizing clinical guidelines (Shiffman et al., 2009). Across various fields, CNLs strike a balance between readability and formal precision, enabling precise interpretation for automated validation and execution. In business settings, they enhance transparency and compliance, while in knowledge representation, they lower the barriers for non-programmers to contribute to formal specifications (Spreeuwenberg & Healy, 2010). Among these, SBVR-SE stands out by offering a standardized framework that supports both human understanding and automated processing, with potential benefits from reusing vocabularies and rules (Ashfaq & Bajwa, 2021; Roychoudhury et al., 2017; OMG, 2019). Alternatives like Logical English (LE) extend their use to financial contracts (Kowalski & Datoo, 2022), but SBVR-SE remains broader in scope, despite its limited use in commercial GRC tools (Witt, 2012). A persistent challenge of CNLs is the cognitive load they place on authors. Studies indicate that writers face decreased fluency and higher mental effort when adapting to syntactic and lexical rules (Engelbracht et al., 2010; Clark et al., 2010). Similar issues arise across various fields: contract drafting necessitates reformulation (Pace & Rosner, 2010), clinical guidelines encounter syntactic challenges (Shiffman et al., 2009), and legal texts often require expert input (Wyner et al., 2010). These results show that although CNLs improve precision, their restrictive design reduces usability for non-experts. In this gap, LLMs can offer valuable assistance, helping authors strike a balance between natural and controlled writing strategies. 2.2 TAXONOMIES AND ONTOLOGIES Taxonomies provide hierarchical classification systems that organize concepts, reduce ambiguity, and promote consistency in the interpretation of rules. Ross (1997) proposed a taxonomy of modalities that includes obligations, prohibitions, conditional permissions, and recommendations. The SBVR specification (OMG, 2019) distinguishes between definitional rules, which establish necessary truths, and behavioral rules, which guide organizational conduct. Von Halle (2001) introduced structured models for expressing these rules, and Witt (2012) expanded the classification to include definitions, data, activities, and roles. In this work, Witt’s taxonomy is adopted to identify patterns and link them to templates, thereby improving both clarity and consistency in the representation of rules. Ontologies formalize vocabularies by explicitly defining terms and their relationships, enabling the representation of structured and interoperable knowledge (W3C OWL Working Group, 2012). They are applied in regulatory contexts such as structuring knowledge graphs for the U.S. Code of Federal Regulations (Joshi & Saha, 2020), integrating OntoDT with SBVR for compliance checking (Bouzidi et al., 2011), and modeling contractual relationships (Sharifi et al., 2020). In law, the LKIF ontology offers a core model for fundamental legal concepts (Hoekstra et al., 2007), though domain-specific ontologies remain scarce and costly to maintain (Sansone & Sperlí, 2022). In finance, the Financial Regulation Ontology (FRO) extends FIBO and LKIF to support compliance (Jayzed Data Models Inc., 2021). Although SBVR does not define an ontology, it maps to machine-readable models via MOF and XMI (OMG, 2019), which allows its use in knowledge engineering applications. SBVR 1.5 clarifies that its terminological dictionary and rulebook are not intended as direct metamodels for data structures or business information models, but rather as documentation of meanings and propositions (OMG, 2019). A transformation step is necessary to convert SBVR content into data or reasoning models. This work does not introduce a new ontology, but instead recognizes its importance in improving the interpretation of vocabularies and rules when represented as knowledge graphs. Open ontologies such as FIBO (EDMC, 2024), LKIF, and FRO can support SBVR transformation by enabling semantic integration, complex queries, and accurate rule application, despite ongoing challenges in adoption and maintenance. 2.3 LLMS IN THE CONTEXT OF REGULATIONS LLMs, such as OpenAI’s GPT-4 and Google’s Gemini, represent a significant advance in NLP by combining training on massive text corpora with billions of parameters. Min et al. (2023) describe three paradigms for their use: pre-training with fine-tuning, prompt-based learning, and reformulating NLP tasks as text generation problems. Sun et al. (2022) expand this perspective by mapping seven paradigms of NLP, emphasizing the ongoing shift toward more generalist and practical approaches. In the legal domain, Anh et al. (2023) demonstrate that LLMs outperform traditional methods in handling ambiguous language and complex structures. Zhong et al. (2020) demonstrate how NLP techniques have improved tasks such as clause analysis and compliance verification. In contrast, Goebel et al. (2023) highlight the practical success of teams in the Competition on Legal Information Extraction and Entailment (COLIEE), where they employ LLMs to address demanding jurisprudence-related tasks. Despite these advances, Anh et al. (2023) also underscore the importance of addressing key ethical and technical challenges, noting that the application of LLMs requires careful attention to domain adaptation and risk mitigation. In this research, general-purpose LLMs are adopted without domain-specific fine-tuning. Instead, all NLP operations are guided exclusively through the use of prompt engineering. 2.4 SEMANTIC TEXTUAL SIMILARITY Semantic Textual Similarity (STS) measures the degree of semantic equivalence between text segments, typically producing a graded score rather than a binary classification (Chandrasekaran & Mago, 2021). It differs from semantic relatedness, which also includes associative connections (e.g., ”coffee”–”mug”). In the context of business rules, where natural language is often ambiguous or idiomatic, STS can assess how faithfully a Controlled Natural Language (CNL) representation preserves the intended meaning of its natural language source. Traditional metrics, such as BLEU and ROUGE, only measure surface-level similarity and often fail to align with human judgments, particularly when synonyms, paraphrases, or factual correctness are involved (Yang et al., 2018). More recent methods include embedding-based measures like SemScore, which better align with human evaluation (Aynetdinov & Akbik, 2024), and LLM-based techniques that utilize preference labels (Gilardi et al., 2023), benchmark competitions (Wei et al., 2024), or hybrid systems combining automation and human feedback (Shankar et al., 2024). Other strategies, such as composite datasets, teacher-student models, and contrastive reinforcement learning, further improve LLM evaluation capabilities (He et al., 2024). Dong et al. (2024) explore LLM-as-a-judge approaches that adapt to individual user preferences. By incorporating verbalized uncertainty and self-reported confidence levels, their method filters out unreliable responses. It reaches up to 80% agreement on high-confidence samples, often exceeding the accuracy of human evaluators. These advances demonstrate a growing trend toward evaluation metrics that more closely align with human judgment, thereby improving both accuracy and scalability in assessing and adapting LLMs. Beyond metrics, several frameworks for evaluating CNLs have also been proposed. Kuhn (2010) suggests that CNLs should be assessed through empirical comparisons with natural language, focusing on four dimensions: expressivity, naturalness, simplicity, and comprehensibility. Engelbrecht et al. (2010) address usability in their Talking Rabbit study, in which non-experts attempted to produce sentences within a controlled language interface, highlighting the importance of learnability and adaptability. Clark et al. (2010) emphasize the trade-off between naturalness and predictability, arguing that CNL must remain accessible to human authors while retaining sufficient predictability to support computational processing. 3 RELATED WORKS This work introduces a set of patterns for classifying and comparing approaches to business rule transformation. Following the notion of patterns as recurring solutions to established problems (Alexander, 1979), seven patterns were identified in the reviewed literature. (P1) Document processing refers to the decomposition of regulatory texts while preserving their semantic structure. (P2) Rule extraction focuses on identifying textual segments that explicitly convey business rules. (P3) Concept extraction addresses the detection and categorization of entities within texts, while (P4) Concept relationships encompass synonym, specialization, and generalization. (P5) Concept disambiguation is essential for interpreting terms in their correct context. (P6) Rule classification deals with assigning rules to categories based on taxonomies, and (P7) Rule transformation into CNL concerns the generation of structured, unambiguous representations. This taxonomy of patterns provides the basis for critically evaluating existing approaches and serves as the conceptual foundation for the method developed in this work. From a software engineering perspective, these patterns also reveal how compliance processes can be made more maintainable and adaptable over time, since each pattern addresses a step in evolving unstructured regulatory requirements into machine-interpretable rules that can be systematically integrated into software systems. Table 1 - Comparison of approaches for transforming business rules into SBVR Abi-Lahoud et al. (2013) Business vocabulary and rules Rule-based 1-7 Manual Bajwa & Shahzada (2017) Business vocabulary and rules NLP + Rule-based 2, 3, 6, 7 UML class diagrams Semiautomatic Chittimalli et al. (2020) Business vocabulary and rules NLP + Rule-based 2-7 Automatic Haj et al. (2021) Business vocabulary and rules NLP + Rule-based 2-7 Automatic Joshi e Saha (2020) Rule extraction and graph for QA NLP + Rule-based 1-7 Automatic Roychoudhury et al. (2017) Business vocabulary and rules NLP + Rule-based 2-4, 6, 7 Dictionary Semiautomatic Skersys et al. (2022) Business vocabulary and rules Rule-based 1-3, 6, 7 BPMN diagrams Semiautomatic The proposed approach Business vocabulary and rules Probabilistic model for NLP and orchestration 1-7 Automatic Table 1 summarizes the works on transforming business rules into SBVR, identified through searches in the ACM, IEEE Xplore, ScienceDirect, SpringerLink, and Web of Science databases between 2018 and 2024. The search employed the keywords finance, financial, natural language processing, NLP, ontology, and semantics of business rules and vocabulary. In addition to the systematic review, the snowballing technique was applied (Greenhalgh & Peacock, 2005; Wohlin, 2014). While this overview provides a comparative classification of approaches according to the seven transformation patterns, a deeper examination of the individual contributions is necessary to understand their strengths, limitations, and implications for the evolution of compliance process automation. The following paragraphs discuss these works. Abi-Lahoud et al. (2013) propose a protocol based on SBVR for interpreting U.S. regulations. The process, which encompasses all seven transformation patterns, is entirely manual and carried out by SMEs. These experts select and analyze legal texts, identify modalities such as obligations, permissions, and prohibitions, extract and define relevant vocabulary, structure the rules in SBVR-SE, apply SBVR stylization, and formalize the rules to ensure they are machine-readable. The final stage involves validating and refining the SBVR rules to maintain alignment with the original legal text. The objective of their work is to enhance the precision and accessibility of complex regulations, thereby strengthening compliance and enhancing understanding of governance. While comprehensive, reliance on manual execution limits scalability and makes it difficult to adapt the process when regulations evolve, constraining its applicability in dynamic compliance environments. Bajwa and Shahzada (2017) present an approach for the automated generation of Object Constraint Language (OCL) constraints from natural language specifications, using SBVR as an intermediate step. Their method applies linguistic analysis of natural language text, including part-of-speech tagging and rule-based parsing, to extract core SBVR elements (P2, P3, P6, and P7), which are subsequently transformed into OCL statements. The approach, however, has limitations. It is restricted to the Unified Modeling Language (UML) class model domain, requires vocabulary terms used in constraints to match those in the UML class model, and supports only a subset of OCL sentence types. Moreover, the authors do not address document representation (P1), concept disambiguation, or generalization (P4 and P5). From a process evolution perspective, the dependence on UML models narrows generalizability and limits its role in adaptable compliance management. Chittimalli et al. (2020) propose an unsupervised method for extracting vocabularies and rules in SBVR format from business documents, designed for automated processing and verification in enterprise systems. Their approach is structured into four main steps: sentence extraction (P2), entity extraction, fact extraction, and rule mining (P3, P6, and P7), employing NLP techniques and dependency parsers. Despite its contributions, the method has some limitations, including the absence of coreference resolution (due to SpaCy’s constraints) and the lack of mechanisms for handling synonyms or generalization relationships (P4 and P5). Unlike other approaches, it uses an N-gram model to classify sentences as either rules or noise. However, it treats each sentence as independent (P1), which restricts its ability to capture broader contextual dependencies. These limitations hinder the long-term maintainability of compliance models, as the approach does not support the evolution of rule dependencies or semantic relationships. Haj et al. (2021) present an automatic method for transforming business rules expressed in natural language into SBVR. Their approach applies lexical, syntactic, and semantic analyses to capture both the structure and meaning of texts (P2). The method emphasizes the construction of a terminological dictionary (P3, P4, and P5) and the identification of business rules (P6), which are then mapped to SBVR elements using pattern-matching techniques that can recognize diverse sentence structures, including conditions, negations, and quantifications (P7). The primary focus is the dictionary, giving priority to rules that impose restrictions or provide definitional information about concepts, while excluding those related to processes or calculations. Although effective for explicit and well-structured rules, the method presents challenges when dealing with complex or implicit rules that require domain-specific knowledge. Like Chittimalli et al. (2020), it treats all clauses as independent, disregarding hierarchical relationships (P1). This reduces its ability to support scalable compliance processes that must evolve in response to increasingly complex regulatory contexts. Joshi and Saha (2020) propose converting regulations from the CFR into a structured, machine-readable format. Their approach applies NLP and deep learning techniques to extract terms, definitions, and rules (P2, P3, P4, P6), which are then organized into a semantically enriched knowledge graph (P1). Applied to Title 48 of the CFR, this graph enables automated reasoning. It supports applications in question answering, compliance verification, and regulatory analysis, although it does not directly address the transformation of rules into SBVR (P7). The authors also tackle terminological ambiguity by creating a contextualized dictionary (P5), which is validated by SMEs. To achieve this, they combine rule-based techniques and regular expressions with models such as Recurrent Neural Networks (RNN), Word2Vec, and LLMs. In their experiments, an LLM was pre-trained on general language tasks and fine-tuned with regulatory data, achieving high performance in classifying clauses into obligations, permissions, and prohibitions. While effective for reasoning tasks, the absence of direct SBVR transformation may reduce its applicability for organizations seeking continuous integration of compliance rules into evolving software processes. Roychoudhury et al. (2017) present a semi-automated approach for transforming legal text into structured English, supporting compliance verification in the financial sector. Their methodology identifies sentences and extracts n-ary propositions (P2, P6) using the ClausIE tool. A subsequent selection step is based on a measure of ”informativeness,” defined as the ratio between domain-recognized mentions and the total number of tokens in a proposition. Concepts, definitions, and synonyms are identified through machine learning and NLP techniques (P3, P4), and the selected propositions are then converted into SE (P7). The transformation from SE into SBVR is automated, although the authors do not provide details on the mapping process. Despite limitations such as dependency on ClausIE and the requirement for experts to be proficient in SE notation, the approach enables regulatory rules to be structured by domain experts without formal knowledge of SBVR. Term disambiguation (P5) and clause dependency (P1) are not addressed. These gaps may limit scalability and increase the effort needed to maintain rule sets as regulatory frameworks evolve. Skersys et al. (2022) propose a methodology for converting BPMN diagrams into SBVR rules, employing deontic logic to incorporate obligations, permissions, and prohibitions. The method identifies Business Process Model and Notation (BPMN) elements such as tasks, events, and gateways, and maps them semantically to equivalent SBVR concepts (P1, P2, P3). This is followed by the application of nine specific transformation rules that automate the conversion into SBVR rules (P7). These transformations account for structures ranging from simple sequence flows to complex interactions among participants, while preserving the semantic integrity of the original model. The approach, however, requires BPMN diagrams to be structured with explicit participant definitions, standardized naming conventions, and precise condition specifications. Although validated with 32 test models, the authors acknowledge limitations, including threats to construct and internal validity, reliance on manual modeling, and the lack of standardized modeling practices, all of which may hinder the generalizability of the results. From a process perspective, the dependence on well-structured BPMN inputs constrains its adaptability in diverse organizational contexts where compliance processes must evolve with heterogeneous modeling practices. The reviewed studies demonstrate progress in formalizing business rules through deterministic or hybrid techniques, yet they also reveal limitations in scalability, flexibility, and traceability. Approaches relying on deterministic algorithms often struggle to manage ambiguity or adapt to evolving regulatory requirements, which undermines their applicability in long-term compliance management. This is particularly relevant for software process compliance, where maintainability and evolution are crucial to ensure that systems remain aligned with the dynamic regulatory environment. Some contributions highlight partial advances. Chittimalli et al. (2020) and Haj et al. (2021), for example, include the automatic extraction of terminological dictionaries directly from text, unlike approaches that rely exclusively on external sources or omit details on how definitions are obtained. The quality of such dictionaries has a direct impact on transformation accuracy and may benefit from domain-specific ontologies, as suggested by Ford et al. (2016). Other methods, while incorporating rule-based NLP techniques to decompose and parameterize textual elements, face challenges of generalization, making their results difficult to replicate in different contexts. Furthermore, many treat textual segments as isolated units, which reduces their capacity to capture the structural coherence of legal documents (Agnoloni & Francesconi, 2011). Noteworthy exceptions are the frameworks proposed by Joshi and Saha (2020) and Dragoni et al. (2015), which introduce structural ontologies to improve contextual representation. To the best of our knowledge, the existing research lacks reproducibility, as none of the reviewed studies have made their code, data, or artifacts publicly available. This restricts independent validation and limits broader adoption. These constraints highlight the need for approaches that can manage ambiguity, remain adaptable, and support transparent and reproducible workflows in compliance process automation. The proposed automatic approach addresses these needs through a probabilistic model supported by NLP and orchestration techniques that handle context dependence and ambiguity. Aligning the seven transformation patterns with LLM-driven processes enables automation while allowing compliance processes to evolve in response to both regulatory and organizational changes. In doing so, the method extends prior work by addressing maintainability, adaptability, and continuous improvement. 4 A BROAD VIEW OF THE PROPOSED METHOD FOR TRANSFORMATION The automatic extraction of rules and policies from regulatory documents still faces significant limitations, as traditional NLP techniques do not adequately capture the complex semantic relations and hierarchical structures of legal texts (Louis et al., 2023; Anh et al., 2023; Sun et al., 2022). To address these challenges, Joshi and Saha (2020) argue that combining multiple approaches is necessary to effectively represent legal knowledge. This work builds on the hypothesis that LLMs can support the transformation of regulatory natural language into a CNL, creating a bridge between the intention behind the rule and its interpretation in a form recognizable to SMEs. By reducing the ambiguity inherent in natural language, this process enables machine interpretation when combined with a Model-Driven Architecture (MDA) approach to SBVR. Within this approach, vocabularies and rules are defined as an authoritative, technology-neutral model of business meaning, from which explicit transformations can be specified for different execution platforms. Such structuring improves both traceability and adaptability, which are critical for maintaining compliance processes as regulations evolve. At the same time, rules expressed in CNL remain understandable to humans, allowing experts to verify whether the intended meaning has been preserved. The role of the SME is not to perform the transformation but to validate its adequacy, ensuring that ambiguity is resolved and that the resulting rules remain faithful to the original intent. An approximate analogy can be drawn with image generation. Although a person without artistic training may not be able to produce the image themselves, they can still judge whether the generated result reflects the intended description. Similarly, the rule expressed in CNL functions as an artifact that enables experts to assess the adequacy of the transformation. For the construction of the dataset used in this study, this work adopts the protocol of Abi-Lahoud et al. (2013) for the manual extraction and transformation of rules from natural language into SBVR-SE. The protocol of outlining that involves reading regulatory texts, identifying and stylizing modalities, verbs, and concepts, followed by the semantic enrichment of these elements. Applying this process provides a reliable baseline for validation, ensuring that subsequent automation steps can be systematically compared against an expert-driven reference. Abi-Lahoud’s protocol also serves as an inspiration for the proposed methodology, in which the structured manual steps are reinterpreted and partially automated through probabilistic NLP and LLM-based techniques. Although all seven steps are outlined for completeness, only Step 2 has been partially implemented so far, as reported in Section 4.1. Figure 1 - Example of the transformation of a business rule from natural language into SBVR An example obtained through the manual application of the protocol is presented in Figure 1. In this case, the initial sentence was taken from Title 17, Chapter II, Part 275, Section 275.0-2 of the CFR (paragraph a, subparagraph 3). Following Witt’s taxonomy (2012), this excerpt was classified as an operative business rule of the role type, with the subtype responsibility. Text decorations indicate the identified elements: double underlining marks names, single underlining marks terms, italicized blue text denotes verbal symbols, and plain orange text highlights keywords. Each element, including the statement itself, is then represented in a knowledge graph together with its classification metadata (element type, taxonomy class, and so forth). Beyond identification, GRC requires preserving traceability between the candidate statement (CS) and its transformed statement (TS), while retaining all intermediate results generated during the transformation process. This not only ensures explainability but also strengthens auditability and supports the long-term maintainability of compliance processes, as regulatory changes can be systematically traced through the sequence of transformations rather than reconstructed from raw text. Figure 2 - Method for the Transformation Process from Natural Language to SBVR To obtain the result illustrated in Figure 1, the original protocol was extended with additional steps inspired by the works of Ashfaq and Bajwa (2021), Omrane et al. (2011), and Haj et al. (2021). These contributions introduced semantic annotation and pattern-based techniques for identifying and mapping statements to SBVR, associating metadata to reduce ambiguity, and bridging the gap between natural language and formal representation. They also enhanced interpretability by allowing SMEs to understand better both the identified elements and the transformations applied. Building on these extensions, the proposed method is organized into seven steps, as illustrated in Figure 2. Steps 1 and 2 identify and extract relevant elements from legal texts, including fact types, operative rules, terms, names, and verbal symbols. At this stage, dictionaries are created, and terms or names sharing the same signifier are linked. These dictionaries establish the context of each element, since identical signifiers may appear in more than one dictionary with distinct meanings. By systematically managing these dictionaries, the approach supports scalability and facilitates the integration of new or updated regulations. Steps 3 and 4 establish relationships among elements, including the mapping of terms and names to their counterparts in FIBO. These relationships form a graph that connects internal and external elements, as well as dictionaries. While there is no standard taxonomy for such a graph, one possible representation of CFR regulations could involve the FRO or a more general ontology derived from LKIF. Constructing these relationships as graph structures enhances traceability and creates reusable artifacts, which are crucial for maintaining compliance processes across domains. Step 5 classifies the extracted elements according to Witt’s taxonomy (2012), providing both the intended meaning of the element and the template for its transformation. Step 6 then converts the annotated elements into SBVR-SE, addressing challenges such as long sentences, multiple clauses, and terminological ambiguities (Wyner & Peters, 2011), as well as identifying atomic rules (Witt, 2012; Haarst, 2013). By combining classification with controlled transformation, the method ensures that rules can be represented in a form that is both machine-readable and interpretable by experts, thereby strengthening adaptability as regulatory frameworks evolve. Figure 3 - Steps of the Transformation The final step of the method is validation ( step 7 ), illustrated in Figure 3. Given the scarcity of datasets for regulatory rule transformation, the validation process employs a combination of complementary techniques to ensure accuracy and reliability. For the identification and classification stages, gold-standard datasets are comparatively less costly to construct, relying mainly on domain expertise to identify rules in legal texts and classify them according to Witt’s taxonomy (2012). For the transformation stage, validation is performed through semantic similarity techniques such as SemScore and LLM-as-a-Judge, which compare the candidate statement with its transformed version and assign a score. Additional characteristics of the transformed statement, including naturalness, predictability, simplicity, and comprehensibility, can also be evaluated by an LLM to assess usability. Validating each stage independently mitigates cascading reliability failures, in which minor and often unpredictable errors in earlier steps propagate throughout the pipeline. This concern arises from the reliability profile of LLMs, which may perform strongly on complex tasks but fail unpredictably on simple ones. Their reliability is further constrained by input sensitivity, context fragility, and stochastic outputs (Reid et al., 2025). Although the objective of this work is to automate the transformation process, the importance of human oversight is recognized (Huang et al., 2024), as also emphasized by Louis et al. (2023), Min et al. (2023), and Longo et al. (2024). To incorporate expert validation and make the process semi-automatic, the validation stage integrates feedback from SMEs alongside that of the AI system, thereby reinforcing reliability and accountability. Whereas previous approaches, such as Haj et al. (2021), rely on deterministic algorithms based on linguistic features (e.g., part-of-speech tagging) to classify and transform elements into CNL-compatible structures, this work adopts a strategy guided by LLMs through prompt engineering. Each stage of validation is executed independently, relying on the classification and text generation capabilities of LLMs together with a shared global context. Dividing the process into coherent steps simplifies instructions, reduces the likelihood of compounding errors, and improves consistency. In doing so, Step 7 integrates the predictability of deterministic methods with the flexibility and adaptability of LLMs (He et al., 2024; Sun et al., 2024), thereby supporting compliance processes that remain auditable and adaptable as regulatory and organizational requirements evolve. 4.1 PRELIMINARY EVALUATION Figure 4-Initial version of the extraction prompt for Step 2 (Identify elements) A preliminary evaluation was carried out focusing on Step 2 of the proposed method (”Identify elements”), using an initial prompt from Figure 4 to extract declarative elements from Title 17, Chapter II, Part 275, Section 275.0-2 of the CFR (paragraph a, subparagraph 3). The system identified seven statements, including Evidence of Service Certification (Figure 1), and extracted terms such as ”Secretary”, ”Commission”, ”Named party”, and ”Process, pleadings, or other papers”, classifying them as common or proper nouns according to SBVR distinctions. Verb symbols like ”certifies”, ”were served”, ”forwarded”, and ”constitute’ were also captured. Although limited to a single step and based on an initial prompt version, these results provide early evidence of the method’s feasibility and lay the groundwork for more comprehensive evaluations that cover the entire pipeline. Unlike most prior approaches that depend on deterministic algorithms (Haj et al., 2021; Roychoudhury et al., 2017), this preliminary evaluation demonstrates that LLM-based extraction can handle ambiguity in regulatory texts by identifying terms and rules without requiring handcrafted rules. 5 CONCLUSION AND FUTURE WORK The motivation behind this work stems from the limitations of Governance, Risk, and Compliance (GRC) processes, especially in the financial sector, regarding the automation and formalization of policies and regulations. Although the Semantics of Business Vocabulary and Business Rules (SBVR) formalism offers notable advantages, few studies have integrated it with the specific needs of GRC, such as traceability and explainability. Combining SBVR with new AI technologies creates opportunities for automation, particularly in the complex tasks of converting financial regulations into formal rules. LLMs can assist in extracting, structuring, and validating these rules through their linguistic interpretation capabilities, but they also have significant limitations, including bias, inaccuracies, and a lack of explainability. These limitations necessitate the use of them within controlled and auditable workflows. Classifying elements according to a taxonomy helps align regulatory intent with functional meaning, thereby standardizing representations and encouraging the reuse and automation of business rules. This work presents a method for developing and validating automated solutions, aiming to modernize previous proposals by adopting a practical approach to utilizing LLMs in rule transformation and validation. The methodology links SBVR with LLM-based NLP, enabling automation while maintaining interpretability for domain experts. It provides a foundation for future implementations. Although the complete evaluation of the proposed method has not yet been conducted, an initial experiment was performed with Step 2 (Identify elements). This preliminary assessment demonstrates the feasibility of automatically extracting terms, names, and fact types from regulatory texts. However, the results are limited to this single step and cannot be generalized to the complete method. A comprehensive validation covering all stages (Steps 1–7) is under development. Future research should expand this work in several directions. Developing tools that implement the proposed method is a priority, alongside experiments that evaluate their effectiveness and investigate human-machine interaction in validation. Another important area is the creation of gold-standard datasets, which are still scarce but vital for benchmarking solutions (Wyner & Peters, 2011). Because building such datasets is expensive and requires specialized expertise in both the domain and SBVR, future efforts should also consider alternatives, such as synthetic data generation (Endres et al., 2022). Promising paths include combining fine-tuning techniques—such as direct Reinforcement Learning from AI Feedback (Lee et al., 2024)—with AI-generated semantic equivalence feedback to create cost-effective, scalable validation resources. Exploring these options can help researchers develop more adaptable and auditable compliance processes that align with the ongoing evolution of software systems and organizational practices. References 1. Abi-Lahoud, E., Butler, T., Chapin, D., & Hall, J. (2013). Interpreting regulations with SBVR . RuleML2013@ Challenge, Human Language Technology and Doctoral Consortium. Agnoloni, T., & Francesconi, E. (2011, June). Modelling semantic profiles in legislative documents for enhanced norm accessibility. In Proceedings of the 13th International Conference on Artificial Intelligence and Law (pp. 111-115). https://doi.org/10.1145/2018358.2018374 Alexander, C. (1979). The timeless way of building . Oxford University Press. Angelov, K., & Ranta, A. (2010). Implementing Controlled Languages in GF. In: Fuchs, N.E. (eds) Controlled Natural Language. CNL 2009. Lecture Notes in Computer Science(), vol 5972. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-14418-9_6 Anh, D. H., Do, D. T., Tran, V., & Le Minh, N. (2023, October). The impact of large language modeling on natural language processing in legal texts: A comprehensive survey. In 2023 15th International Conference on Knowledge and Systems Engineering (KSE) (pp. 1-7). IEEE. https://doi.org/10.1109/KSE59128.2023.10299488 Ashfaq, F., & Bajwa, I. S. (2021). Natural language ambiguity resolution by intelligent semantic annotation of software requirements. Automated Software Engineering, 28 (1), 13. https://doi.org/10.1007/s10515-021-00291-0 Aynetdinov, A., & Akbik, A. (2024). Semscore: Automated evaluation of instruction-tuned LLMs based on semantic textual similarity. arXiv . https://doi.org/10.48550/arXiv.2401.17072 Bajwa, I. S., & Shahzada, M. A. (2017). Automated generation of OCL constraints: NL based approach vs pattern based approach. Mehran University Research Journal of Engineering & Technology, 36(2), 243-254. https://doi.org/10.22581/muet1982.1702.04 Bouzidi, K. R., Faron-Zucker, C., Fies, B., & Le Thanh, N. (2011, August). An ontological approach for modeling technical standards for compliance checking. In International Conference on Web Reasoning and Rule Systems (pp. 244-249). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-23580-1_19 Chandrasekaran, D., & Mago, V. (2021). Evolution of semantic similarity—a survey. Acm Computing Surveys (Csur) , 54 (2), 1-37. https://doi.org/10.1145/3440755 Chen, K., Yang, Y., Chen, B., López, J. A. H., Mussbacher, G., & Varró, D. (2023, October). Automated domain modeling with large language models: A comparative study. In 2023 ACM/IEEE 26th International Conference on Model Driven Engineering Languages and Systems (MODELS) (pp. 162-172). IEEE. https://doi.org/10.1109/MODELS58315.2023.00037 Chittimalli, P. K., Prakash, C., Naik, R., & Bhattacharyya, A. (2020, February). An approach to mine SBVR vocabularies and rules from business documents. In Proceedings of the 13th Innovations in Software Engineering Conference (formerly known as India Software Engineering Conference) (pp. 1-11). ACM. https://doi.org/10.1145/3385032.3385046 Clark, P., Murray, W. R., Harrison, P., & Thompson, J. (2010). Naturalness vs. predictability: A key debate in controlled languages. In N. E. Fuchs (Ed.), Controlled natural language (Lecture Notes in Computer Science, Vol. 5972, pp. 65–81). Springer. https://doi.org/10.1007/978-3-642-14418-9_5 Cramer, M., Fisseni, B., Koepke, P., Kühlwein, D., Schröder, B., & Veldman, J. (2009, June). The naproche project controlled natural language proof checking of mathematical texts. In International Workshop on Controlled Natural Language (pp. 170-186). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-14418-9_11 Enterprise Data Management Council (EDMC). (2024). The financial industry business ontology: The global standard ontology for efficient and unambiguous financial services. https://spec.edmcouncil.org/fibo/ Dragoni, M., Villata, S., Rizzi, W., & Governatori, G. (2015, December). Combining natural language processing approaches for rule extraction from legal documents. In International Workshop on AI Approaches to the Complexity of Legal Systems (pp. 287-300). Cham: Springer International Publishing. https://doi.org/10.1007/978-3-030-00178-0_19 Dong, Y. R., Hu, T., & Collier, N. (2024). Can LLM be a personalized judge? arXiv . https://doi.org/10.48550/arXiv.2406.11657 Endres, M., Mannarapotta Venugopal, A., & Tran, T. S. (2022, August). Synthetic data generation: A comparative study. In Proceedings of the 26th international database engineered applications symposium (pp. 94-102). https://doi.org/10.1145/3548785.3548793 Engelbrecht, P., Hart, G., & Dolbear, C. (2010). Talking Rabbit: A user evaluation of sentence production. In N. E. Fuchs (Ed.), Controlled natural language (Lecture Notes in Computer Science, Vol. 5972, pp. 56–64). Springer. https://doi.org/10.1007/978-3-642-14418-9_4 Ford, R., Denker, G., Elenius, D., Moore, W., & Abi-Lahoud, E. (2016, September). Automating financial regulatory compliance using ontology+ rules and Sunflower. In Proceedings of the 12th International Conference on Semantic Systems (pp. 113-120). https://doi.org/10.1145/2993318.2993329 Gilardi, F., Alizadeh, M., & Kubli, M. (2023). ChatGPT outperforms crowd workers for text-annotation tasks. arXiv . https://doi.org/10.48550/arXiv.2303.15056 Goebel, R., Kano, Y., Kim, M. Y., Rabelo, J., Satoh, K., & Yoshioka, M. (2023, June). Summary of the competition on legal information, extraction/entailment (COLIEE) 2023. In Proceedings of the nineteenth international conference on artificial intelligence and law (pp. 472-480). https://doi.org/10.1145/3594536.3595176 Greenhalgh, T., & Peacock, R. (2005). Effectiveness and efficiency of search methods in systematic reviews of complex evidence: audit of primary sources. Bmj , 331 (7524), 1064-1065. https://doi.org/10.1136/bmj.38636.593461.68 Gruzitis, N., & Barzdins, G. (2009, June). Polysemy in controlled natural language texts. In International Workshop on Controlled Natural Language (pp. 102-120). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-14418-9_7 Haj, A., Jarrar, A., Balouki, Y., & Gadir, T. (2021). The semantic of business vocabulary and business rules: An automatic generation from textual statements. IEEE Access , 9 , 56506-56522. https://doi.org/10.1109/ACCESS.2021.3071623 Haarst, R. V. (2013). SBVR made easy: Business vocabulary and rules as a critical asset . Conceptual Heaven. He, Q., Zeng, J., He, Q., Liang, J., & Xiao, Y. (2024). From complex to simple: Enhancing multi-constraint complex instruction-following ability of large language models. arXiv . https://doi.org/10.48550/arXiv.2404.15846 Hoekstra, R., Breuker, J., Di Bello, M., & Boer, A. (2007). The LKIF core ontology of basic legal concepts. In Proceedings of the Legal Ontologies and Artificial Intelligence Techniques (LOAIT) Workshop (CEUR Workshop Proceedings, Vol. 321, pp. 43–63). CEUR-WS. Hough, A. R., & Gluck, K. A. (2019). The understanding problem in cognitive science. Advances in Cognitive Systems , 8 , 13-32. Huang, H., Bu, X., Zhou, H., Qu, Y., Liu, J., Yang, M., … & Zhao, T. (2024). An empirical study of LLM-as-a-judge for LLM evaluation: Fine-tuned judge model is not a general substitute for GPT-4. arXiv . https://doi.org/10.48550/arXiv.2403.02839 Jackson, P. (2020). Understanding understanding and ambiguity in natural language. Procedia Computer Science , 169 , 209-225. https://doi.org/10.1016/j.procs.2020.02.138 Jayzed Data Models Inc. (2021). Semantic compliance in finance: The financial regulation ontologies (FRO) . https://finregont.com Joshi, K. P., & Saha, S. (2020). A semantically rich framework for knowledge representation of code of federal regulations. Digital Government: Research and Practice , 1 (3), 1-17. https://doi.org/10.1145/3425192 Kowalski, R., & Datoo, A. (2022). Logical English meets legal English for swaps and derivatives. Artificial Intelligence and Law , 30 (2), 163-197. https://doi.org/10.1007/s10506-021-09295-3 Khurana, D., Koli, A., Khatter, K., & Singh, S. (2023). Natural language processing: state of the art, current trends and challenges. Multimedia tools and applications , 82 (3), 3713-3744. https://doi.org/10.1007/s11042-022-13428-4 Kuhn, T. (2010). An evaluation framework for controlled natural languages. In N. E. Fuchs (Ed.), Controlled natural language (Lecture Notes in Computer Science, Vol. 5972, pp. 1–20). Springer. https://doi.org/10.1007/978-3-642-14418-9_1 Kulkarni, V., Sunkle, S., Kholkar, D., Roychoudhury, S., Kumar, R., & Raghunandan, M. (2021). Toward automated regulatory compliance. CSI Transactions on ICT , 9 (2), 95-104. https://doi.org/10.1007/s40012-021-00329-4 Lee, H., Phatale, S., Mansoor, H., Mesnard, T., Ferret, J., Lu, K., Bishop, C., Hall, E., Carbune, V., Rastogi, A., & Prakash, S. (2024). RLAIF vs. RLHF: Scaling reinforcement learning from human feedback with AI feedback. In Proceedings of the 41st International Conference on Machine Learning (ICML ’24, Vol. 235, Article 1071, pp. 26874–26901). JMLR.org. https://dl.acm.org/doi/10.1145/3726302.3730283 Longo, L., Brcic, M., Cabitza, F., Choi, J., Confalonieri, R., Del Ser, J., … & Stumpf, S. (2024). Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions. Information Fusion , 106 , 102301. https://doi.org/10.1016/j.inffus.2024.102301 Louis, A., Van Dijck, G., & Spanakis, G. (2023). Finding the law: Enhancing statutory article retrieval via graph neural networks. arXiv . https://doi.org/10.48550/arXiv.2301.12847 Min, B., Ross, H., Sulem, E., Veyseh, A. P. B., Nguyen, T. H., Sainz, O., … & Roth, D. (2023). Recent advances in NLP via large pre-trained language models: A survey. arXiv . https://doi.org/10.48550/arXiv.2301.12847 Object Management Group (OMG). (2019). Semantics of business vocabulary and business rules (SBVR) (Version 1.5) . https://www.omg.org/spec/SBVR/1.5/PDF Omrane, N., Nazarenko, A., Rosina, P., Szulman, S., & Westphal, C. (2011, November). Lexicalized ontology for a business rules management platform: An automotive use case. In International Workshop on Rules and Rule Markup Languages for the Semantic Web (pp. 179-192). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-24908-2_21 Pace, G. J., & Rosner, M. (2010). A controlled language for the specification of contracts. In N. E. Fuchs (Ed.), Controlled natural language (Lecture Notes in Computer Science, Vol. 5972, pp. 225–246). Springer. https://doi.org/10.1007/978-3-642-14418-9_14 Reid, A., O’Callaghan, S., Carroll, L., & Caetano, T. (2025). Risk analysis techniques for governed LLM-based multi-agent systems. arXiv . https://doi.org/10.48550/arXiv.2508.05687 Roychoudhury, S., Sunkle, S., Kholkar, D., & Kulkarni, V. (2017, October). From natural language to SBVR model authoring using structured english for compliance checking. In 2017 IEEE 21st International Enterprise Distributed Object Computing Conference (EDOC) (pp. 73-78). IEEE Computer Society. https://doi.org/10.1109/EDOC.2017.19 Ross, R. G. (1997). The business rule book: Classifying, defining, and modeling rules: Ross method (Version 4.0). Database Research Group. Sadiq, S., Governatori, G., Namiri, K. (2007). Modeling Control Objectives for Business Process Compliance. In: Alonso, G., Dadam, P., Rosemann, M. (eds) Business Process Management. BPM 2007. Lecture Notes in Computer Science, vol 4714. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-540-75183-0_12 Sansone, C., & Sperlí, G. (2022). Legal information retrieval systems: state-of-the-art and open issues. Information Systems , 106 , 101967. https://doi.org/10.1016/j.is.2021.101967 Shankar, S., Zamfirescu-Pereira, J. D., Hartmann, B., Parameswaran, A., & Arawjo, I. (2024, October). Who validates the validators? Aligning LLM-assisted evaluation of LLM outputs with human preferences. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology (pp. 1-14). https://doi.org/10.1145/3654777.3676450 Sharifi, S., Parvizimosaed, A., Amyot, D., Logrippo, L., & Mylopoulos, J. (2020). Symboleo: Towards a specification language for legal contracts. In 2020 IEEE 28th International Requirements Engineering Conference (RE) (pp. 364–369). IEEE. https://doi.org/10.1109/RE48521.2020.00049 Shiffman, R. N., Michel, G., Krauthammer, M., Fuchs, N. E., Kaljurand, K., & Kuhn, T. (2009, June). Writing clinical practice guidelines in controlled natural language. In International Workshop on Controlled Natural Language (pp. 265-280). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-14418-9_16 Skersys, T., Danenas, P., Mickeviciute, E., & Butleris, R. (2022). Transforming BPMN Processes to SBVR Process Rules with Deontic Modalities. Applied Sciences , 12 (18), 8976. https://doi.org/10.3390/app12188976 Spreeuwenberg, S., & Healy, K. A. (2010). SBVR’s approach to controlled natural language. In N. E. Fuchs (Ed.), Controlled natural language (Lecture Notes in Computer Science, Vol. 5972, pp. 193–203). Springer. https://doi.org/10.1007/978-3-642-14418-9_10 Sun, H., Liu, L., Li, J., Wang, F., Dong, B., Lin, R., & Huang, R. (2024). Conifer: Improving complex constrained instruction-following ability of large language models. arXiv . https://doi.org/10.48550/arXiv.2404.02823 Sun, T. X., Liu, X. Y., Qiu, X. P., & Huang, X. J. (2022). Paradigm shift in natural language processing. Machine Intelligence Research , 19 (3), 169-183. https://doi.org/10.1007/s11633-022-1331-6 Sunkle, S., Kholkar, D., & Kulkarni, V. (2015). Toward better mapping between regulations and operations of enterprises using vocabularies and semantic similarity. Complex Systems Informatics and Modeling Quarterly , (5), 39-60. http://dx.doi.org/10.7250/csimq.2015-5.04 Von Halle, B. (2001). Business rules applied: building better systems using the business rules approach . Wiley Publishing. W3C OWL Working Group. (2012, December 11). OWL 2 Web Ontology Language Document Overview (W3C Recommendation). World Wide Web Consortium. Retrieved August 23, 2025, from http://www.w3.org/TR/owl2-overview Wei, F., Chen, X., & Luo, L. (2024). Rethinking generative large language model evaluation for semantic comprehension. In Proceedings of the 41st International Conference on Machine Learning (ICML ’24, Vol. 235, Article 2153, pp. 52525–52558). JMLR.org. https://dl.acm.org/doi/10.5555/3692070.3694223 Witt, G. (2012). Writing effective business rules . Elsevier. Wohlin, C. (2014, May). Guidelines for snowballing in systematic literature studies and a replication in software engineering. In Proceedings of the 18th international conference on evaluation and assessment in software engineering (pp. 1-10). https://doi.org/10.1145/2601248.2601268 Wyner, A., Angelov, K., Barzdins, G., Damljanovic, D., Davis, B., Fuchs, N., Hoefler, S., Jones, K., Kaljurand, K., Kuhn, T., Luts, M., Pool, J., Rosner, M., Schwitter, R., & Sowa, J. (2010). On controlled natural languages: Properties and prospects. In N. E. Fuchs (Ed.), Controlled natural language (pp. 281–289). Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-14418-9_17 Wyner, A., & Peters, W. (2011). On rule extraction from regulations. In Legal knowledge and information systems (pp. 113-122). IOS Press. https://doi.org/10.3233/978-1-60750-981-3-113 Yang, A., Liu, K., Liu, J., Lyu, Y., & Li, S. (2018). Adaptations of ROUGE and BLEU to better evaluate machine reading comprehension task. arXiv preprint arXiv:1806.03578 . https://doi.org/10.48550/arXiv.1806.03578 Yu, F., Quartey, L., & Schilder, F. (2023, July). Exploring the effectiveness of prompt engineering for legal reasoning tasks. In Findings of the association for computational linguistics: ACL 2023 (pp. 13582-13596). https://doi.org/10.18653/v1/2023.findings-acl.858 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. arXiv preprint arXiv:2004.12158 . https://doi.org/10.48550/arXiv.2004.12158 Crossref Google Scholar Information & Authors Information Version history V1 Version 1 15 October 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords cfr grc llm nlp sbvr taxonomy Authors Affiliations Anderson dos Santos 0000-0001-9254-7764 [email protected] Instituto de Pesquisas Tecnologicas View all articles by this author Paulo Sérgio Muniz Silva Instituto de Pesquisas Tecnologicas View all articles by this author Metrics & Citations Metrics Article Usage 347 views 132 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Anderson dos Santos, Paulo Sérgio Muniz Silva. Evolving Compliance Processes through LLM-Supported Transformation of Regulations into SBVR. Authorea . 15 October 2025. DOI: https://doi.org/10.22541/au.176051910.04955057/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176051910.04955057/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fea05adc8bdad07',t:'MTc3OTI2NjI3NQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();