Design of intelligent legal text analysis and information retrieval system based on BERT model

Design of intelligent legal text analysis and information retrieval system based on BERT model | 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 Design of intelligent legal text analysis and information retrieval system based on BERT model Bingyao Li, Meng Wang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2994403/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract The amalgamation of information retrieval systems and soft computing techniques establishes a robust framework to confront the challenges and seize the opportunities presented by the vast expanse of big data. As the volume, diversity, and velocity of data continue to proliferate, further advancements in this domain are poised to significantly contribute to various realms, encompassing healthcare, finance, e-commerce, and scientific research, ultimately propelling innovation and facilitating decision-making in the epoch of big data. The application of artificial intelligence (AI) technology to the analysis and comprehension of legal documents holds the potential to expedite the acquisition of case-specific information by legal researchers, thereby expediting their research endeavors. This paper proposes a legal text analysis and retrieval model, rooted in the Bleem model (Bert-based Legal Paper's Element Extracting Model). Initially, our model employs Bert as the coding layer to extract the semantic information embedded within document sentences and element exemplars. Subsequently, we leverage the Attention mechanism to align the semantic essence of element example sentences with document sentences, while simultaneously computing their respective attention weights. Comparative experiments and attention visualization are then employed to validate the efficacy of the Bleem model. The experimental results corroborate the superiority of the Bleem model in terms of accuracy and F1 scores. The visualization of the attention mechanism effectively reveals the inner workings of the Bleem model and unveils its capacity to explore pertinent fragments within document sentences and element examples. Building upon the aforementioned model, we have devised an intelligent legal text analysis and retrieval system, empowering legal researchers to swiftly acquire pivotal data through case briefs. This application has effectively fostered the integration of legal services within the holistic management of public risks. Social public risk Legal intelligence Legal text analysis and retrieval Bleem model Bert model Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 1 Introduction With the advent of the legal intelligence stage, automation and artificial intelligence technologies are widely used in the legal field, and more and more legal documents will be read and analyzed by machines. In order to promote the development of intelligent justice, the research of natural language processing technology in the judicial field has received extensive attention and deep attention. The judicial field enabled by AI has brought profound changes to the comprehensive governance of public risks, and helped modernize the national emergency management system and capabilities. In exploring the practical path of smart justice, procuratorates at all levels in China have carried out a wealth of research (Yu J, 2021) [ 1 ] . The "intelligent assistant case handling system" of Shanghai "206" project can prompt non-standard problems such as evidence flaws in the process of handling cases by learning a large number of cases, and greatly prevent unjust, false and wrong cases (Xinlei C A I, 2020) [ 2 ] ; The legal service robot jointly developed by Jiangsu Provincial People's Procuratorate and Kevos Electric Co., Ltd. integrates procuratorial business with artificial intelligence technologies such as face recognition and voice interaction, and enhances the interactivity and convenience of the judicial information service platform (Zhukova M N, 2019) [ 3 ] . These studies have enhanced the role of legal services in preventing and resolving public risks, further promoted the effective development of social governance, and actively prevented and resolved social public risks. Intelligent legal text analysis and retrieval means that the machine automatically searches and condenses the information of legal text, and then forms shorter text to help users quickly understand the main content of the text (Taran MO, 2020) [ 4 ] . The intelligent analysis and processing of massive legal documents has become an important research content of natural language processing in the judicial field. Getting useful information from a large number of legal documents is of great significance not only to judicial workers, but also to ordinary users. The use of automatic analysis technology of legal texts can enable users to easily consult relevant cases without fully mastering relevant legal knowledge, so that users can feel the convenience and rapidity of justice, reduce grievances and improve their recognition of the judiciary. For judicial workers, judicial abstracts can compress and summarize the content of legal documents, improve the efficiency of judicial work, reduce the workload, promote judicial workers to be familiar with legal provisions, improve legal literacy, and avoid excessive litigation. It can be seen that the research on intelligent legal text analysis and retrieval technology can effectively compensate for the damaged interests of citizens, make citizens more emotionally identify with the power of law, promote legal education for urban and rural residents, and further promote the comprehensive governance of public risks. However, due to the difference between the analysis and retrieval of legal texts and general documents, and the fact that most existing legal text analysis models can only use the fixed semantic representation of words, they cannot learn the context related word representation, which leads to their lack of representation ability. In view of the above problems and the application requirements of intelligent legal text analysis and retrieval in the comprehensive governance of public risks, this research has carried out the following work: First, we built a text implication model for document element extraction—Bleem, which is based on the Bert class model (Uthirapathy S E 2023) [ 5 ] , combined with the element analysis method of law, to solve the problem of element information homogeneity, while improving the attention mechanism algorithm, and exploring the associated fragments of document sentences and element examples. By automatically learning context sensitive word representation, the representation ability of the model has been greatly enhanced; Secondly, based on the Bleem model, we designed an intelligent analysis system for legal documents, including document crawling, automatic reading and analysis tools, to help legal researchers obtain text summary information and enhance the role of legal services in preventing and resolving social governance risks; Finally, we apply the extraction model of legal document elements to the comprehensive governance of public risks. For crimes that are easy to lead to public risks, such as theft, intentional homicide, provocation, etc., our Bleem model can quickly and accurately identify the criminal elements, which is of great significance to the comprehensive governance of public risks. 2 Related works The PASCAL-2005 contest (Hu W, 2014) [ 6 ] , held in 2005, was an early contest to propose the question of text analysis and retrieval. The dataset released by the contest gave birth to many early studies in this field, such as similarity-based methods (Paul C, 2016) [ 7 ] , text alignment based methods (Voeste A, 2021) [ 8 ] , logic calculus based methods (Waszek D, 2021) [ 9 ] and text transformation based methods (Ghadage Y H, 2016) [ 10 ] . In the similarity-based approach, Park designed a word bag-based text implication model (Park B, 2018) [ 11 ] . This model first segments the text, assigns weight to each word in the created word bag according to the word frequency of the reverse document, and then uses the measurement based on the lexical chain in WordNet to calculate the semantic similarity. In the method based on text alignment, Li used the method of contradiction detection to complete the task of text implication (Li L, 2017) [ 12 ] . First, he used text alignment to align the words in conditional sentences and hypothetical sentences, and defined the part that is not aligned as the conflict between them. Then, he extracted the contradictory semantic features of conditions and assumptions, and used logical regression to complete the text implication analysis. In the method based on logic calculus, Upmeier, from the perspective of logic calculus, used retrospective reasoning to calculate the cost of conditional sentence inference to hypothesis sentence, so as to judge the dependency between the two (Upmeier zu Belzen A, 2021) [ 13 ] . However, the early legal text analysis model has serious problems such as sequence repetition and poor readability, which cannot adapt to more complex semantics. With the development of deep nerves, various problems in this field have been gradually improved, which has attracted extensive attention. In 2010, the concept of deep learning was proposed by researchers, and deep learning methods were also applied to the field of text analysis and retrieval by some researchers. Rhanoui introduced the convolutional neural network CNN into the text analysis and retrieval task (Rhanoui M, 2019) [ 14 ] to explore the hierarchical structure of sentences, and tried to obtain potential matching patterns in premise sentences and hypothesis sentences at different levels of abstraction. By superimposing convolutional pooling blocks, feature mapping can become deeper, so as to explore deeper semantic relationships. Dave used the deep learning model to build a hybrid automatic summarization system for feature extraction of legal document sentences (Dave H, 2015) [ 15 ] , including document frequency, legal noun frequency, sentence location, whether to cite legal provisions, whether to elaborate legal provisions, etc., and through these feature construction rules to limit the content extracted from the summary. In recent years, the neural network abstract model of sequence to sequence (Seq2Seq) has gradually attracted the attention of the academic community (Li Z, 2018) [ 16 ] . He proposed an end-to-end algorithm that carefully handles the translation of entities (Xie S, 2022). [ 17 ] . The model used an RNN encoder and a decoder. The former encodes the input sequence into a vector, and the latter generates the target sequence according to the coding representation vector of the sequence. The model structure is shown in Fig. 1 (a). Since the performance of the Seq2Seq model of RNNs decreased rapidly with the increase of the text length, Bahuleyan introduced the attention mechanism to enable the decoder to dynamically focus on the information at different positions of the original sentence at each moment. The Seq2Seq model based on attention is shown in Fig. 1 (b) (Yu T, 2022) [ 18 ] . The text representation layer used in the model based on CNN and RNN can only use the fixed semantic representation of words, and cannot learn the context related word representation, which makes its representation ability still lacking. A research team has proposed a model that uses transformer and BERT language representations to identify Arabic conversation behavior.(Alaa J, 2021) [ 19 ] .Thanks to the pure attention mechanism of Transformer, Bert model can focus on more important parts of the text context. It is not necessary to mine the context information by recording historical states like LSTM. The word vector output by Bert model really realizes context relevance, which greatly enhances the model's representation ability, He has achieved good performance improvement in 11 NLP tasks such as machine reading comprehension. Research showed that Bert type model is an ideal model for legal reading comprehension (Zhang N N, 2021) [ 20 ] . Because, among the questions of legal reading and understanding, in addition to the fragment questions such as the amount of money involved and the criminal entity, there may also be the right and wrong questions such as "whether there is torture to extract confessions", and the unanswered questions that cannot be answered by the documents. The Bert model can fully learn the context semantics of words. At the same time, Bert also used a multi task learning architecture, which can understand complex legal semantics and handle multiple types of legal issues. The main architecture of Bert model is shown in Fig. 2 . 3 Legal text analysis and retrieval model In order to obtain the information homogenous elements of the document, this study selects element examples as rules to limit the importance of the document sentences. The element examples are given by legal experts, including facts, processes, backgrounds and other document elements. 3.1 Bert model construction Bert model has been used in text analysis tasks such as Multi genre Natural Language Inference (MNLI), Question Answering Natural Language Inference (QNLI), and has achieved excellent performance. Based on Bert model, this study constructs a model for extracting homogeneous elements of legal documents. The main architecture of the model is shown in Fig. 3 . In this model, the word level sequence \(({c}_{1}^{i},{c}_{2}^{i},\dots ,{c}_{n}^{i})\) of the \(itℎ\) sentence in a legal document will be matched with the example sentence \(({s}_{1}^{i},{s}_{2}^{i},\dots ,{s}_{n}^{i})\) of the \(jtℎ\) element. The matched sentence pairs will be connected, and then input to the Bert model to learn the word vector. In the Bert model, the input sentence pair participates in the MLM (Masked Language Model) training task, and the training goal is shown in Formula (1). $$P\left({w}_{i}|{w}_{1},\dots ,{w}_{i-1},{w}_{i+1},\dots ,{w}_{n}\right) \left(1\right)$$ That is, use the context word sequence ( \({w}_{1},\dots ,{w}_{i-1},{w}_{i+1},\dots ,{w}_{n}\) ) of the document, predict that the position \({w}_{i}\) is the probability of a word, and use the generated probability as the word vector of the character. Finally, the word vector sequence of sentence pairs is obtained, as shown in Formula (2). $$(c\_{ℎ}_{1}^{i}, c\_{ℎ}_{2}^{i},\dots ,c\_{ℎ}_{n}^{i},s\_{ℎ}_{1}^{i},s\_{ℎ}_{2}^{i},\dots ,s\_{ℎ}_{m}^{i}) \left(2\right)$$ At the same time, the full connection layer of a single layer is used as the output layer of the model. The full connection layer maps the word vector sequence of the sentence pair to the implication relation probability of the final output, as shown in Formula (3). $${logit}_{i}=sigmoid\left({W}_{l}\right[c\_{ℎ}_{i};s\_{ℎ}_{i}]+{b}_{l}) \left(3\right)$$ 3.2 Construction of Bleem Model for Extracting Elements of Legal Documents In the analysis and reading of legal documents, there are two characteristics: the homogeneity of element information of legal documents and the diversity of problems in legal reading comprehension, which hinder the development of legal intelligence. In order to solve this problem, this research constructs a text implication model for document element extraction: Bleem (Bert based Legal paper's Element Extracting Model). First, match each sentence with the element example sentences set by the legal experts one by one to construct a series of document sentences - element example sentence pairs. Then input these sentence pairs into the Bleem model. The Bleem model calculates the semantic similarity between document sentences and element examples, and outputs the matching scores between text sentences and element examples. Then, with the help of these matching scores, we can screen out the essential sentences in the text, so as to achieve the purpose of analyzing the key content in the document. The architecture of the model is shown in Fig. 4 . Specifically, the \(itℎ\) sentence \(({c}_{1}^{i},{c}_{2}^{i},\dots ,{c}_{n}^{i})\) of the text and the \(jtℎ\) example sentence \(({s}_{1}^{i},{s}_{2}^{i},\dots ,{s}_{n}^{i})\) of the element set form a sentence pair, where the length of the text sentence \({c}_{i}\) is \(n\) , and the length of the element example sentence \({s}_{i}\) is \(m\) . Input into the Bleem model. The Bleem model will use the three-layer neural network to process the sentence pair \(({c}_{i},{s}_{i})\) . The first layer is the Bert layer, which is implemented using the Bert-Chinese. Before inputting this layer, the sentence pairs \(({c}_{i},{s}_{i})\) will be connected together and converted into three sequences. The first sequence is the word ID sequence. Each word in the sentence pair \(({c}_{i},{s}_{i})\) will be mapped to the word ID through the dictionary counted from all corpora. The second sequence is the marker sequence of front and back sentences, in which the sentence \({c}_{i}\) of document text is converted into a full zero sequence with the same length, and the element example sentence s-i is converted into a full one sequence with the same length. The third sequence is a mask sequence, which is equal to the length of the sentence pair \(({c}_{i},{s}_{i})\) and is a full 1 sequence. The second layer is the attention layer. After obtaining the word vector sequence \(c\_{ℎ}_{i}\) of the document sentence and the word vector sequence \(s\_{ℎ}_{i}\) of the element example sentence, connect the two-word vector sequences together, and then input them into a feedforward neural network to learn a hidden layer representation \({z}_{i}^{j}\) . Then input the hidden layer representation \({z}_{i}^{j}\) into softmax, map the value in the \({z}_{i}^{j}\) vector to the (0,1) range, and output it as the attention weight value \({\alpha }_{i}^{j}\) of the document sentence for the element sentence. The perfect results are shown in Formula (4)~(5). $${z}_{i}^{j}=tanℎ\left({W}_{z}\right[c\_{ℎ}_{i};s\_{ℎ}_{i}\left]\right) \left(4\right)$$ $${\alpha }_{i}^{j}=softmax\left({z}_{i}^{j}\right) \left(5\right)$$ Here, \({W}_{z}\) is the weight value of the feedforward neural network, and \([c\_{ℎ}_{i};s\_{ℎ}_{i}]\) represents the connection of word vector sequences of document sentences and element example sentences. Use the learned document sentences to pay attention to the element sentences \({\alpha }_{i}\) . The word vector sequence \(c\_{ℎ}_{1}^{i}\) of the document sentence is weighted to obtain the hidden layer representation \({ℎ}_{i}\) of the rewritten document sentence. $${ℎ}_{i}=\sum _{k=1}^{m}{\alpha }_{jk}^{i}\bullet c\_{ℎ}_{k}^{i} \left(6\right)$$ Here, \(m\) represents the length of the document sentence, \(c\_{ℎ}_{k}^{i}\) represents the \(ktℎ\) word in the document sentence, \({\alpha }_{jk}^{i}\) is the corresponding attention weight. At the same time, the cross-entropy function is used to construct the loss function of the Bleem model. The specific form is shown in Formula (7). $$loss=-\frac{1}{n}\sum _{i=1}^{n}\left[{y}_{i}\text{log}\left({logit}_{i}\right)+\left(1-{y}_{i}\right)\text{log}\left(1-{logit}_{i}\right)\right] \left(7\right)$$ Here, \({y}_{i}\) is the true value, \({logit}_{i}\) is the predicted value, and \(n\) is the number of samples. 4 Experimental results and analysis 4.1 Experimental data The CAIL2020 abstract dataset (Yuan J, 2021) [ 21 ] used in the experiment totaled 9848 civil first instance judgments. The judgment document is divided into several sentences in advance, each sentence is marked with the label of whether it is important, and the corresponding full text of the reference abstract is provided. The types of documents include civil disputes such as tort liability, lease contracts, labor contracts, inheritance contracts, and loan contracts. The documents average 2568 words, with a maximum length of 13060 words. 95% of the documents exceed 4663 words. The summary average 283 words, with a maximum length of 474 words. 95% of the documents exceed 327 words. This data set belongs to the long text summary data set. The labels of the elements of the document and the sentences to be classified of the examples are given by the legal experts. Table 1 shows some of the element examples. Table 1 Elements of divorce proceedings Type Example sentences of elements Type Example sentences of elements 1 Having children after marriage 11 Separation after marriage 2 Personal property before marriage 12 Failure to perform the divorce agreement 3 Child support with limited capacity 13 Sue for divorce again 4 Statutory divorce 14 Damages 5 Having joint property of husband and wife 15 Monthly alimony payment 6 Failure to fulfill family obligations 16 Two years of separation due to emotional discord 7 Pay alimony 17 Divorce granted 8 There are children born out of wedlock 18 Children living with non-custodians 9 Real estate division 19 Joint debt of husband and wife 10 Appropriate help 20 Personal property after marriage From Table 1 , we can see that the original dataset has a small amount of data, but there are many categories classified. At the same time, the sentence of the document does not only contain one element semantics, and there is a case where one sentence corresponds to multiple elements. In order to solve the above problem, this study will convert the multi classification dataset into the text implication dataset. Combine the document sentence with the element example sentence corresponding to the label to build "implication" example and "non implication" example respectively. Use synonymous sentences to enhance data, so that one element can correspond to about three synonymous sentences, which increases the sample size of the dataset. At the same time, the combination of document sentences and element example sentences in pairs can make the document sentences correspond to multiple categories, giving more opportunities to explore the semantics of different elements in the document sentences. Data enhancement scheme based on synonymous semantics is shown in Fig. 5 . After dataset transformation, the sample of feature extraction data includes three fields, including premise, hypothesis and implication label. The premise is the document sentence to be classified in the original data set. Assume that it is an element example sentence corresponding to a category, or a synonymous sentence of an element example sentence. The implication tag is 0 or 1. When it is 0, the document sentence does not contain the semantics of element example sentences. When it is 1, the document sentence contains the semantics of element example sentences. The overall data set size is: the training set contains 52479 samples, the validation set contains 17494 samples, and the test set contains 17494 samples. For the converted data, we calculated the length proportion of document sentences and element sample sentences in all samples, and the final result is shown in Fig. 6 . It can be seen that most of the document sentences are within 100 and most of the element example sentences are within 28. 4.2 Performance comparison and analysis of different models In order to compare the performance of Bleem model and other models in extracting document elements, this study selected text similarity model (Islamaj R, 2019) [ 22 ] , BI CNN (Yin W, 2015) [ 23 ] , ABCNN (Type 3) (Yin W, 2016) [ 24 ] , match_ Lstm (Wang X, 2021) [ 25 ] , Bert (Uthirapathy S E 2023) [ 5 ] serve as the baseline. For LSTM class model and CNN model, word2vec word vector is used as the embedded layer uniformly, and the length of the embedded vector is 300. At the same time, word level embedded vector is used. For models that use LSTM layers, the hidden layer size of LSTM is set to 128. For the model using the CNN layer, the filter size of CNN is set to 50, and the window size of the filter is 3. At the same time, in order to reduce over fitting, the Dropout policy is used for these two types of models, and the dropout value is set to 0.8. At the same time, set the batch size of network post transmission to 128 and the number of training rounds to 20. The optimizer of LSTM class model and CNN class model uses Adam algorithm, and the initial learning rate is 0.0001. Put the above model and Bleem model together on the document element extraction dataset, and verify these models. The experimental results are shown in Table 2 . Table 2 Results of comparative experiments Model Accuracy F1 Score Training duration Similar 0.4448 0.5074 4s BI-CNN 0.5812 0.525 4m 40s ABCNN 0.6365 0.5221 9m 34s match_lstm 0.6419 0.591 1h 54m Bert 0.9563 0.9565 65m 43s Bleem(Our) 0.9612 0.961 68m 23s In this study, accuracy and F1 score were used as evaluation indicators of the model. Among them, accuracy is used to evaluate the prediction effect of the model, and F1 score is used to evaluate the stability of the model effect from the two dimensions of recall and precision. At the same time, this study recorded the training duration of each model to evaluate the time complexity of the model. From Table 2 , we can see that Similar model has the minimum training time. However, because its exploration of text semantics is too shallow, it has the worst accuracy in predicting text implication relations and F1 scores. Moreover, the accuracy score of the Similar model is less than 0.5, which is lower than the random guess. It can be seen that the Similar model is not suitable for the analysis task of document elements. At the same time, from the perspective of accuracy and F1 score, CNN model and RNN model have the same effect. In the simple model, the F1 scores and accuracy scores of Bi CNN and Bi LSTM are between 0.5 and 0.6, which is better than random guess and Similar model. Based on the word level task, CNN model analyzes the local features of the text with a sliding window with a length of 3 in the document element extraction task, which can combine and filter out the potential 3-gram features to obtain the local semantics of words. At the same time, the LSTM model involves the semantic representation of words before the text sequence in the semantic computation of words at the current position, so that the previous word semantics can affect the semantic representation of words at the current position, which enables LSTM to analyze the long-distance dependence and short-term dependence in the text sequence, and obtain more rich context information. Because of their respective advantages, CNN and LSTM have both been used in text classification tasks, and have achieved similar performance in the task of extracting document elements in this study. The attention mechanism can span the influence of time step interval, directly align two text sequences, calculate their dependency on long distance time steps, and map this dependency to the attention weight value to weight the text sequence, so that the text sequence contains a more long-distance semantic association. After further adding the attention mechanism to the ABCNN and march LSTM models, their performance also shows a synchronous improvement. However, CNN uses the strategy of sparse interaction and weight sharing. For the previous sequence, the convolution kernel only connects with some neurons of the previous layer, rather than the full connection form of the full connection layer and LSTM layer, which greatly reduces the calculation parameters and the time complexity of CNN. At the same time, each convolution kernel only needs to train a set of weights, and then the weights can be shared among each convolution kernel, which further reduces the complexity of the model. In the training process of this study, the training duration of the two types of training models shows their differences in time complexity. CNN type models end their training in 5 minutes and 10 minutes respectively, while LSTM model takes 2 hours. In terms of accuracy and F1 score, Bert model and Bert based Bleem model have greatly improved compared with ABCNN and match LSTM models based on Attention. In order to obtain further effect improvement on the Bert model, this study uses Attention and pooling layer as a modeling layer after the Bert model. Attention is responsible for aligning the word vector sequence of the obtained element example sentences and document sentences, so as to mine the relationship between the two, and convert the relationship into the attention weight value, weighting the original word vector representation. 4.3 Visualization of attention mechanism In order to further analyze the effectiveness of the attention mechanism in the Bleem model, this study visualizes the Attention weight calculated in the Bleem model. In the images shown, there are corresponding examples of three types of cases, namely, "divorce litigation", "loan dispute" and "labor dispute". The specific results are shown in Fig. 7 : the gradual change from purple to yellow indicates the attention weight value in the range of [0,1], and close to yellow indicates the attention weight value is high. In the first example, the element example sentence given is "sue for divorce for the second time". The visualization results of divorce litigation examples is shown in Fig. 8 . The attention weight value learned by the model in the training process focuses on the document sentence fragment "sue, then withdraw the lawsuit, after withdrawing the lawsuit, the original,". It is not difficult to see that the model has fully understood the meaning of "secondary prosecution" and paid attention to the process of "secondary prosecution" corresponding to "litigation" and "withdrawal of litigation". In the second example, the element example given is "signing the agreement on the transfer of rights and interests of creditor's rights assets". The Visualization results of loan dispute samples is shown in Fig. 9 . In this example, the sentence fragment of the document that the model pays attention to is "the creditor's rights have been transferred to the plaintiff Weize Company and have been notified". The keywords "creditor's rights" and "transfer" in the element example sentences also appear in the document sentence fragments concerned by the model, and the model further finds the entities related to transfer. In the third example, the element example given is "salary arrears". The visual Results of Examples of Labor Disputes is shown in Fig. 10 . In this example, the part with high attention weight given by the model is the opening part of the document sentence "The plaintiff's salary in arrears". This segment has a direct semantic relationship with the element example sentence. The model also finds this segment directly and excludes the relationship between other segments and the element example sentence. Based on the above three examples, it is not difficult to find that the proposed Bleem model can well find the parts with semantic associations between document sentences and element examples in the task of document element extraction, and use these fragments to make efficient judgments on the implication of text elements. 5 Design of intelligent legal text analysis and retrieval system This research is based on the extraction model of legal document elements contained in the text - Bleem, which solves the problem of element information homogeneity in the extraction of legal document elements, and realizes the abstract information extraction of legal documents. Based on this model, this study designed an intelligent legal text analysis and retrieval system to help legal researchers obtain the summary and reading results of legal texts and support legal research. The overall architecture of the system is divided into three parts: document crawling module, automatic reading module, and analysis tool module. The overall architecture of the intelligent analysis system for legal documents is shown in Fig. 11 . Among them, the document crawling module mainly helps legal researchers to automatically obtain the required document data from the public legal document website. The automatic reading module is the core part of the whole system, which mainly includes the machine reading understanding part, the document element analysis part, the result fusion part and the result post-processing part. The main purpose of the analysis tool module is to provide legal researchers with data analysis tools, such as simple statistical tools, regression analysis tools, cluster analysis tools, etc. 5.1 Document crawling module The crawler part, as the input part of the system, is mainly composed of three steps. The first step is to execute the crawler. The crawler program will automatically crawl relevant legal documents from the litigation website according to the case keywords entered by legal researchers. The second step is case similarity detection and case clustering. In this step, the system will first use the "Case Similarity Detection" model to calculate the case similarity between legal documents and get the similarity matrix of documents. Then, input the similarity matrix into k-mean clustering algorithm to cluster documents. The third step is the screening of instrument clusters. According to the clustering results, the system will output a representative instrument of each category cluster. After reading the representative instruments of this cluster, legal researchers will determine the relevance of this cluster of instruments to their own research, and then determine whether this cluster of instruments is retained. Thus, all collections of documents to be read can be obtained. The flow chart of the document crawling module is shown in Fig. 12 . 5.2 Automatic reading module The automatic reading module is divided into four parts: machine reading understanding part, document element analysis part, result fusion part and result post-processing part. The flow chart of this module is shown in Fig. 13 . The machine reading and understanding part implements the reading and understanding of legal documents. First of all, legal researchers designated research questions, such as "how much is the amount involved" and "when is the time of crime". Then, according to the collection of documents to be read obtained from the crawler part and the questions given by the researchers, with the help of the pre trained Bleem model, perform sentence level document element analysis, and output document sentences containing element semantics. In the result fusion part, machine reading comprehension and result fusion will be fused. The Bleem model is used to analyze the elements of the document. After the key content of the document is extracted, the results of machine reading comprehension are enhanced by judging the correlation between the elements of the document and the problem. The post-processing part aims to extract the complete answer more accurately. Classify the questions according to some key words in the questions, such as "who" and "when". Then, post process the answers of machine reading comprehension according to the categories of questions. 5.3 Analysis tool module The analysis tool module provides various types of analysis tools for legal researchers. The activity diagram of the analysis tool module is shown in Fig. 14 . The first is simple statistics, such as the average, variance and median of the amount involved, and the proportion of cases of "whether there is torture". The second is regression analysis. For example, for network crimes, you can use regression analysis tools to judge the correlation between the number of information stolen and the sentence. The third is cluster analysis. After the case elements are extracted, all the current cases can be clustered according to the case elements to judge whether there are subdivision cases in the current category of cases. Legal researchers can find more legal phenomena and put forward research ideas better. 6 Application in comprehensive governance of public risk How to use science and technology to deal with and prevent public risks facing mankind has become a major topic of the modernization of governance systems and capabilities in various countries, and also a topic of great concern for industry and academia. This study applies the extraction model of legal document elements to the comprehensive governance of public risk, randomly samples four types of cases from 200 indictments for analysis, and compares the results of the proposed Bleem model with ABCNN, match-lstm, Bert and other models. It can be seen from the four cases shown in Fig. 15 ~ 18 that the summary generation of the Bleem model proposed in this study is the best, and the generated summary is relatively standard and readable, indicating that the model can quickly learn the summary mode of various criminal cases. In the case of the crime of causing trouble and intentional homicide, Bert did not recognize it well, and even in the case of intentional homicide, there was a lot of repetition. In the case of multiple crimes, other comparative models failed to identify key information such as concealing and concealing crimes, which indicates that these models have poor effect on the generation of multiple crimes summaries. The above analysis shows that the Bleem model proposed by us can improve the rapid and accurate identification of criminal elements in criminal facts and is of great significance to the comprehensive management of public risks for crimes that are easy to lead to public risks, such as theft, intentional homicide and provocation. 7 Conclusions The legal intelligence derived from the combination of AI technology and legal theory has had a profound impact on the field of law. AI technology has emerged in a large number of applications in case analysis, intelligent court platforms, evidence collection, etc. However, the homogeneity of element information in legal documents and the diversity of problems in legal reading hinder the application of AI technology. In order to solve these problems, this research constructed a document element extraction model based on text implication - Bleem. This model used the Bert layer as the coding layer to obtain the context sensitive word vector representation, and then used Attention and pooling layers to model the global features, and used the global features to judge whether the document sentence contains element semantics. In the experimental results shown, Bleem model achieved 96% accuracy on the dataset, which was superior to other models. At the same time, our research used the attention weight visualization to prove that our Bleem model can well focus on the sentence fragments associated with document sentences and element example sentences. Then, this research designed an intelligent analysis system for legal documents, including three parts: document crawling, reading analysis, and analysis tools, to help legal researchers obtain the summary information and reading understanding information of legal documents. Finally, we applied the model of extracting elements of legal documents to the comprehensive governance of public risks. The experimental results show that our model can quickly and accurately identify the criminal elements in the criminal facts for crimes that are easy to lead to public risks, such as theft, intentional killing, provocation, etc. Our research can effectively promote the application of legal services in the comprehensive governance of public risks. From the experimental results, the Bleem model has achieved high accuracy, but further discussion is needed on the interpretability. Next, we will carry out the implementation deployment of the model and verify the effect of the intelligent legal text analysis and retrieval system in practice. Declarations Funding The authors have not disclosed any funding. Data availability Enquiries about data availability should be directed to the authors. Conflict Interests The authors have no conflict of interest. Ethical approval The paper does not deal with any ethical problems. Informed consent We declare that all the authors have informed consent. References Yu J, Xia J. E-justice evaluation factors: The case of Smart Court of China[J]. Information Development, 2021, 37(4): 658-670. Xinlei C A I, Rongfu Q I U, Yanli C U I, et al. Power Grid Auxiliary Control System Based on Big Data Application and Artificial Intelligence Decision[C]// 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE). IEEE, 2020: 154-157. Zhukova M N, Zolotarev V V, Zhukov V G, et al. Service robot security from unauthorized access by connection control[C]// 2019 12th International Conference on Developments in eSystems Engineering (DeSE). IEEE, 2019: 526-529. Taran M O, Revunkov G I, Gapanyuk Y E. The text fragment extraction module of the hybrid intelligent information system for analysis of judicial practice of arbitration courts[C]// International Conference on Neuroinformatics. Springer, Cham, 2020: 242-248. Uthirapathy Samson Ebenezar,Sandanam Domnic. Topic Modelling and Opinion Analysis On Climate Change Twitter Data Using LDA And BERT Model.[J]. Procedia Computer Science, 2023,218. Hu W, Hu R, Xie N, et al. Image classification using multiscale information fusion based on saliency driven nonlinear diffusion filtering[J]. IEEE transactions on image processing, 2014, 23(4): 1513-1526. Paul C, Rettinger A, Mogadala A, et al. Efficient graph-based document similarity[C]// European Semantic Web Conference. Springer, Cham, 2016: 334-349. Voeste Anja. Spelling variation and text alignment[J]. Zeitschrift für Sprachwissenschaft,2021,40(3). Waszek D, Schlimm D. Calculus as method or calculus as rules? Boole and Frege on the aims of a logical calculus[J]. Synthese, 2021, 199(5): 11913-11943. Ghadage Y H, Shelke S D. Speech to text conversion for multilingual languages[C]// 2016 International Conference on Communication and Signal Processing (ICCSP). IEEE, 2016: 0236-0240. Park B, Colaresi M, Greene K. Beyond a bag of words: Using PULSAR to extract judgments on specific human rights at scale[J]. Peace Economics, Peace Science and Public Policy, 2018, 24(4). Li L, Qin B, Liu T. Contradiction detection with contradiction-specific word embedding[J]. Algorithms, 2017, 10(2): 59. Upmeier zu Belzen A, Engelschalt P, Krüger D. Modeling as scientific reasoning—The role of abductive reasoning for Modeling competence[J]. Education Sciences, 2021, 11(9): 495. Rhanoui M, Mikram M, Yousfi S, et al. A CNN-BiLSTM model for document-level sentiment analysis[J]. Machine Learning and Knowledge Extraction, 2019, 1(3): 832-847. Dave H, Jaswal S. Multiple text document summarization system using hybrid summarization technique[C]// 2015 1st International Conference on Next Generation Computing Technologies (NGCT). IEEE, 2015: 804-808. Li Z, Cai J, He S, et al. Seq2seq dependency parsing[C]// Proceedings of the 27th International Conference on Computational Linguistics. 2018: 3203-3214. Xie, Shufang,Xia, Yingce,Wu, Lijun,Huang, Yiqing,Fan, Yang,Qin, Tao. End-to-end entity-aware neural machine translation[J]. Machine Learning, 2022(prepublish). Yu Tianxiang,Xin Yang,Zhu Hongliang,Tang Qifeng,Chen Yuling. Network Penetration Intrusion Prediction Based on Attention Seq2seq Model[J]. Security and Communication Networks, 2022,2022. Alaa Joukhadar,Nada Ghneim,Ghaida Rebdawi. Impact of Using Bidirectional Encoder Representations from Transformers (BERT) Models for Arabic Dialogue Acts Identification[J]. ISI,2021,26(5). Zhang N N, Xing Y. Questions and answers on legal texts based on BERT-BiGRU[C]//Journal of Physics: Conference Series. IOP Publishing, 2021, 1828(1): 012035. Yuan J, Wei Z, Gao Y, et al. Overview of SMP-CAIL2020-Argmine: The Interactive Argument-Pair Extraction in Judgement Document Challenge[J]. Data Intelligence, 2021, 3(2): 287-307. Islamaj R, Wilbur W J, Xie N, et al. PubMed Text Similarity Model and its application to curation efforts in the Conserved Domain Database[J]. Database, 2019, 2019. Yin W, Schütze H. Convolutional neural network for paraphrase identification[C]// Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2015: 901-911. Yin W, Schütze H, Xiang B, et al. Abcnn: Attention-based convolutional neural network for modeling sentence pairs[J]. Transactions of the Association for Computational Linguistics, 2016, 4: 259-272. Wang Xiaomin,Wang Haoriqin,Zhao Guocheng,Liu Zhichao,Wu Huarui. ALBERT over Match-LSTM Network for Intelligent Questions Classification in Chinese[J]. Agronomy, 2021,11(8). Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 06 Jun, 2023 Reviewers invited by journal 06 Jun, 2023 Editor assigned by journal 02 Jun, 2023 First submitted to journal 29 May, 2023 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-2994403","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":207208503,"identity":"cb1c1e55-5994-4e90-a746-875716e03f3b","order_by":0,"name":"Bingyao Li","email":"","orcid":"","institution":"1. Shanghai University of Finance and Economics; 2. Shanghai Business school","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Bingyao","middleName":"","lastName":"Li","suffix":""},{"id":207208504,"identity":"6712174d-bb72-4191-9b43-94714e21537a","order_by":1,"name":"Meng Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABA0lEQVRIiWNgGAWjYDADxvbGxsdQtgFxWph7Djcbk6aFfUZ6mzRRWgyOnz386kbFHbvehsS26sI2O3kG9uZtEgw1d3BrOZOXZp1z5lnyzIaDbbdntiUbNvAcK5NgOPYMpxazAzlmxrlth5MNGxvbbvO2MTM2SOSYSTA2HMat5fwboJZ/h5PtDzO2FfO21ds3yL8hoOVGjvHj3IbDdoxtjG3MvG2HExskePBrsb/xxow559jhBMYexmbpGeeOJ7fxpBVbJBzDrUWyP8f4c07NYXvG+c8ffi4oq7btZz+88caHGtxagIBNAkgkNsC5ICIBnwZgvH8AORC/mlEwCkbBKBjRAACRS1sVDIKHBQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0001-4380-2025","institution":"Beijing University of Aeronautics and Astronautics: Beihang University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Meng","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2023-05-29 08:11:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-2994403/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-2994403/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":38187608,"identity":"aa7779fe-6948-48c0-b3a5-14a98f9bcac0","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":102757,"visible":true,"origin":"","legend":"\u003cp\u003e(a) RNN based Seq2Seq model; (b) attention-based Seq2Seq model\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/7ccf8078099310a5069d0188.png"},{"id":38187730,"identity":"83d64c57-e748-4971-963c-5f3987383de9","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":51278,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of Bert model\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/331c5d1c2e08140493bfae0a.png"},{"id":38187937,"identity":"4ceec870-51cd-4754-9a3f-1a40d255bea8","added_by":"auto","created_at":"2023-06-07 20:36:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":70772,"visible":true,"origin":"","legend":"\u003cp\u003eModel structure of extraction model of homogeneous elements of legal documents\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/79c42c9e87a7a1556744adca.png"},{"id":38187734,"identity":"6d88e083-6171-4eb8-8fbe-a25796bd5cc0","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":77723,"visible":true,"origin":"","legend":"\u003cp\u003eModel Structure of Bleem\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/162a7ce8ea9df3aacc155107.png"},{"id":38187732,"identity":"a52d55d4-6015-49df-92e7-6c0d6bc7b9d3","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":56673,"visible":true,"origin":"","legend":"\u003cp\u003eData enhancement scheme based on synonymous semantics\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/c5864a705a931f94cf6e2917.png"},{"id":38187731,"identity":"2bcf2b03-3bce-4535-8399-b27129a1aea1","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":39973,"visible":true,"origin":"","legend":"\u003cp\u003eStatistics of sentence length of document sentences and example sentences of elements\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/0ddf04f4edeb1bcb3375c92d.png"},{"id":38187936,"identity":"7d06b18c-f397-46d7-b324-4aeda79a85c5","added_by":"auto","created_at":"2023-06-07 20:36:39","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":57949,"visible":true,"origin":"","legend":"\u003cp\u003eComparison sequence of attention weight\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/15e2d01880d9d342d2b615fd.png"},{"id":38187736,"identity":"2389a4c9-15ec-47d8-8d75-baff0f1dbab3","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":22989,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization results of divorce litigation examples\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/3bdfa4f422632f0bff08fa11.png"},{"id":38187624,"identity":"1f5da9ee-83f1-48a5-bcfb-477806c810b6","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":20846,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization results of loan dispute samples\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/9a4c59be4531c4148a3647fb.png"},{"id":38187610,"identity":"62d20238-c55a-4115-bf72-679670643f81","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":15959,"visible":true,"origin":"","legend":"\u003cp\u003eVisual Results of Examples of Labor Disputes\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/32e40994961edcb1a1bc69ce.png"},{"id":38187739,"identity":"8c40e1c1-70b0-418e-bb09-40102d2a16f1","added_by":"auto","created_at":"2023-06-07 20:28:39","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":173510,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture of legal document oriented intelligent analysis system\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/d63a4388e1e848474fa2c779.png"},{"id":38187614,"identity":"56eef940-568a-49fb-84fb-9ab5af0d74a1","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":85781,"visible":true,"origin":"","legend":"\u003cp\u003eFlow chart of document crawling module\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/c3171103758ea584da050b17.png"},{"id":38187621,"identity":"21a8ba02-2cb0-495e-a91c-998356ebe3a1","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":60550,"visible":true,"origin":"","legend":"\u003cp\u003eFlow Chart of Automatic Reading Module\u003c/p\u003e","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/a7d287ad7bdabcfc2c27fa00.png"},{"id":38187938,"identity":"d26a4d8b-8953-48a8-ba1a-48e8c78ecec0","added_by":"auto","created_at":"2023-06-07 20:36:39","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":54816,"visible":true,"origin":"","legend":"\u003cp\u003eActivity diagram of analysis tool module\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/719011c6523f0f8a808484a9.png"},{"id":38188065,"identity":"3c8abe78-5795-48da-a5b1-50d207e1555c","added_by":"auto","created_at":"2023-06-07 20:44:39","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":46146,"visible":true,"origin":"","legend":"\u003cp\u003eThe summary of indictment of theft charge\u003c/p\u003e","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/650cece2849f76c25af2c531.png"},{"id":38187940,"identity":"c1cf5a86-d19f-417b-a265-e16499e47716","added_by":"auto","created_at":"2023-06-07 20:36:39","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":51031,"visible":true,"origin":"","legend":"\u003cp\u003eThe summary of indictment of intentional homicide charge\u003c/p\u003e","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/7eb98fdff7a454c9deab635b.png"},{"id":38187625,"identity":"86b482f3-3cef-4acd-9e63-dbefcd5d4fc5","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":65946,"visible":true,"origin":"","legend":"\u003cp\u003eThe summary of indictment of intentional homicide charge\u003c/p\u003e","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/5e7976a1b354cc2dddf99ca1.png"},{"id":38187618,"identity":"6008b808-692e-4bf5-8d90-9e04dc18f8e5","added_by":"auto","created_at":"2023-06-07 20:20:39","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":77726,"visible":true,"origin":"","legend":"\u003cp\u003eThe summary of indictment of multiple charges\u003c/p\u003e","description":"","filename":"floatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/781acecee25425b18cc69fb1.png"},{"id":38188067,"identity":"439b7c8e-f477-4348-b7ee-46c6ffb0c43a","added_by":"auto","created_at":"2023-06-07 20:44:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2153860,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-2994403/v1/8f831327-9ca6-4154-a098-bfa15f851519.pdf"}],"financialInterests":"","formattedTitle":"Design of intelligent legal text analysis and information retrieval system based on BERT model","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eWith the advent of the legal intelligence stage, automation and artificial intelligence technologies are widely used in the legal field, and more and more legal documents will be read and analyzed by machines. In order to promote the development of intelligent justice, the research of natural language processing technology in the judicial field has received extensive attention and deep attention. The judicial field enabled by AI has brought profound changes to the comprehensive governance of public risks, and helped modernize the national emergency management system and capabilities.\u003c/p\u003e \u003cp\u003eIn exploring the practical path of smart justice, procuratorates at all levels in China have carried out a wealth of research (Yu J, 2021) \u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e. The \"intelligent assistant case handling system\" of Shanghai \"206\" project can prompt non-standard problems such as evidence flaws in the process of handling cases by learning a large number of cases, and greatly prevent unjust, false and wrong cases (Xinlei C A I, 2020) \u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e; The legal service robot jointly developed by Jiangsu Provincial People's Procuratorate and Kevos Electric Co., Ltd. integrates procuratorial business with artificial intelligence technologies such as face recognition and voice interaction, and enhances the interactivity and convenience of the judicial information service platform (Zhukova M N, 2019) \u003csup\u003e[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. These studies have enhanced the role of legal services in preventing and resolving public risks, further promoted the effective development of social governance, and actively prevented and resolved social public risks.\u003c/p\u003e \u003cp\u003eIntelligent legal text analysis and retrieval means that the machine automatically searches and condenses the information of legal text, and then forms shorter text to help users quickly understand the main content of the text (Taran MO, 2020) \u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. The intelligent analysis and processing of massive legal documents has become an important research content of natural language processing in the judicial field. Getting useful information from a large number of legal documents is of great significance not only to judicial workers, but also to ordinary users. The use of automatic analysis technology of legal texts can enable users to easily consult relevant cases without fully mastering relevant legal knowledge, so that users can feel the convenience and rapidity of justice, reduce grievances and improve their recognition of the judiciary. For judicial workers, judicial abstracts can compress and summarize the content of legal documents, improve the efficiency of judicial work, reduce the workload, promote judicial workers to be familiar with legal provisions, improve legal literacy, and avoid excessive litigation. It can be seen that the research on intelligent legal text analysis and retrieval technology can effectively compensate for the damaged interests of citizens, make citizens more emotionally identify with the power of law, promote legal education for urban and rural residents, and further promote the comprehensive governance of public risks.\u003c/p\u003e \u003cp\u003eHowever, due to the difference between the analysis and retrieval of legal texts and general documents, and the fact that most existing legal text analysis models can only use the fixed semantic representation of words, they cannot learn the context related word representation, which leads to their lack of representation ability. In view of the above problems and the application requirements of intelligent legal text analysis and retrieval in the comprehensive governance of public risks, this research has carried out the following work:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eFirst, we built a text implication model for document element extraction\u0026mdash;Bleem, which is based on the Bert class model (Uthirapathy S E 2023) \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e, combined with the element analysis method of law, to solve the problem of element information homogeneity, while improving the attention mechanism algorithm, and exploring the associated fragments of document sentences and element examples. By automatically learning context sensitive word representation, the representation ability of the model has been greatly enhanced;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSecondly, based on the Bleem model, we designed an intelligent analysis system for legal documents, including document crawling, automatic reading and analysis tools, to help legal researchers obtain text summary information and enhance the role of legal services in preventing and resolving social governance risks;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eFinally, we apply the extraction model of legal document elements to the comprehensive governance of public risks. For crimes that are easy to lead to public risks, such as theft, intentional homicide, provocation, etc., our Bleem model can quickly and accurately identify the criminal elements, which is of great significance to the comprehensive governance of public risks.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e"},{"header":"2 Related works","content":"\u003cp\u003eThe PASCAL-2005 contest (Hu W, 2014) \u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e, held in 2005, was an early contest to propose the question of text analysis and retrieval. The dataset released by the contest gave birth to many early studies in this field, such as similarity-based methods (Paul C, 2016) \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e, text alignment based methods (Voeste A, 2021) \u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e, logic calculus based methods (Waszek D, 2021)\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e and text transformation based methods (Ghadage Y H, 2016)\u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e. In the similarity-based approach, Park designed a word bag-based text implication model (Park B, 2018) \u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. This model first segments the text, assigns weight to each word in the created word bag according to the word frequency of the reverse document, and then uses the measurement based on the lexical chain in WordNet to calculate the semantic similarity. In the method based on text alignment, Li used the method of contradiction detection to complete the task of text implication (Li L, 2017)\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e. First, he used text alignment to align the words in conditional sentences and hypothetical sentences, and defined the part that is not aligned as the conflict between them. Then, he extracted the contradictory semantic features of conditions and assumptions, and used logical regression to complete the text implication analysis. In the method based on logic calculus, Upmeier, from the perspective of logic calculus, used retrospective reasoning to calculate the cost of conditional sentence inference to hypothesis sentence, so as to judge the dependency between the two (Upmeier zu Belzen A, 2021) \u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. However, the early legal text analysis model has serious problems such as sequence repetition and poor readability, which cannot adapt to more complex semantics. With the development of deep nerves, various problems in this field have been gradually improved, which has attracted extensive attention.\u003c/p\u003e \u003cp\u003eIn 2010, the concept of deep learning was proposed by researchers, and deep learning methods were also applied to the field of text analysis and retrieval by some researchers. Rhanoui introduced the convolutional neural network CNN into the text analysis and retrieval task (Rhanoui M, 2019) \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e to explore the hierarchical structure of sentences, and tried to obtain potential matching patterns in premise sentences and hypothesis sentences at different levels of abstraction. By superimposing convolutional pooling blocks, feature mapping can become deeper, so as to explore deeper semantic relationships. Dave used the deep learning model to build a hybrid automatic summarization system for feature extraction of legal document sentences (Dave H, 2015) \u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e, including document frequency, legal noun frequency, sentence location, whether to cite legal provisions, whether to elaborate legal provisions, etc., and through these feature construction rules to limit the content extracted from the summary. In recent years, the neural network abstract model of sequence to sequence (Seq2Seq) has gradually attracted the attention of the academic community (Li Z, 2018) \u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e. He proposed an end-to-end algorithm that carefully handles the translation of entities (Xie S, 2022).\u003csup\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. The model used an RNN encoder and a decoder. The former encodes the input sequence into a vector, and the latter generates the target sequence according to the coding representation vector of the sequence. The model structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (a). Since the performance of the Seq2Seq model of RNNs decreased rapidly with the increase of the text length, Bahuleyan introduced the attention mechanism to enable the decoder to dynamically focus on the information at different positions of the original sentence at each moment. The Seq2Seq model based on attention is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (b) (Yu T, 2022) \u003csup\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe text representation layer used in the model based on CNN and RNN can only use the fixed semantic representation of words, and cannot learn the context related word representation, which makes its representation ability still lacking. A research team has proposed a model that uses transformer and BERT language representations to identify Arabic conversation behavior.(Alaa J, 2021) \u003csup\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e.Thanks to the pure attention mechanism of Transformer, Bert model can focus on more important parts of the text context. It is not necessary to mine the context information by recording historical states like LSTM. The word vector output by Bert model really realizes context relevance, which greatly enhances the model's representation ability, He has achieved good performance improvement in 11 NLP tasks such as machine reading comprehension. Research showed that Bert type model is an ideal model for legal reading comprehension (Zhang N N, 2021) \u003csup\u003e[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e. Because, among the questions of legal reading and understanding, in addition to the fragment questions such as the amount of money involved and the criminal entity, there may also be the right and wrong questions such as \"whether there is torture to extract confessions\", and the unanswered questions that cannot be answered by the documents. The Bert model can fully learn the context semantics of words. At the same time, Bert also used a multi task learning architecture, which can understand complex legal semantics and handle multiple types of legal issues. The main architecture of Bert model is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e "},{"header":"3 Legal text analysis and retrieval model","content":"\u003cp\u003eIn order to obtain the information homogenous elements of the document, this study selects element examples as rules to limit the importance of the document sentences. The element examples are given by legal experts, including facts, processes, backgrounds and other document elements.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Bert model construction\u003c/h2\u003e \u003cp\u003eBert model has been used in text analysis tasks such as Multi genre Natural Language Inference (MNLI), Question Answering Natural Language Inference (QNLI), and has achieved excellent performance. Based on Bert model, this study constructs a model for extracting homogeneous elements of legal documents. The main architecture of the model is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn this model, the word level sequence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{1}^{i},{c}_{2}^{i},\\dots ,{c}_{n}^{i})\\)\u003c/span\u003e\u003c/span\u003e of the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(itℎ\\)\u003c/span\u003e\u003c/span\u003e sentence in a legal document will be matched with the example sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({s}_{1}^{i},{s}_{2}^{i},\\dots ,{s}_{n}^{i})\\)\u003c/span\u003e\u003c/span\u003e of the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(jtℎ\\)\u003c/span\u003e\u003c/span\u003e element. The matched sentence pairs will be connected, and then input to the Bert model to learn the word vector. In the Bert model, the input sentence pair participates in the MLM (Masked Language Model) training task, and the training goal is shown in Formula (1).\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$P\\left({w}_{i}|{w}_{1},\\dots ,{w}_{i-1},{w}_{i+1},\\dots ,{w}_{n}\\right) \\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThat is, use the context word sequence (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({w}_{1},\\dots ,{w}_{i-1},{w}_{i+1},\\dots ,{w}_{n}\\)\u003c/span\u003e\u003c/span\u003e) of the document, predict that the position \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({w}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the probability of a word, and use the generated probability as the word vector of the character. Finally, the word vector sequence of sentence pairs is obtained, as shown in Formula (2).\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$(c\\_{ℎ}_{1}^{i}, c\\_{ℎ}_{2}^{i},\\dots ,c\\_{ℎ}_{n}^{i},s\\_{ℎ}_{1}^{i},s\\_{ℎ}_{2}^{i},\\dots ,s\\_{ℎ}_{m}^{i}) \\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAt the same time, the full connection layer of a single layer is used as the output layer of the model. The full connection layer maps the word vector sequence of the sentence pair to the implication relation probability of the final output, as shown in Formula (3).\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$${logit}_{i}=sigmoid\\left({W}_{l}\\right[c\\_{ℎ}_{i};s\\_{ℎ}_{i}]+{b}_{l}) \\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Construction of Bleem Model for Extracting Elements of Legal Documents\u003c/h2\u003e \u003cp\u003eIn the analysis and reading of legal documents, there are two characteristics: the homogeneity of element information of legal documents and the diversity of problems in legal reading comprehension, which hinder the development of legal intelligence. In order to solve this problem, this research constructs a text implication model for document element extraction: Bleem (Bert based Legal paper's Element Extracting Model).\u003c/p\u003e \u003cp\u003eFirst, match each sentence with the element example sentences set by the legal experts one by one to construct a series of document sentences - element example sentence pairs. Then input these sentence pairs into the Bleem model. The Bleem model calculates the semantic similarity between document sentences and element examples, and outputs the matching scores between text sentences and element examples. Then, with the help of these matching scores, we can screen out the essential sentences in the text, so as to achieve the purpose of analyzing the key content in the document. The architecture of the model is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eSpecifically, the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(itℎ\\)\u003c/span\u003e\u003c/span\u003e sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{1}^{i},{c}_{2}^{i},\\dots ,{c}_{n}^{i})\\)\u003c/span\u003e\u003c/span\u003e of the text and the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(jtℎ\\)\u003c/span\u003e\u003c/span\u003e example sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({s}_{1}^{i},{s}_{2}^{i},\\dots ,{s}_{n}^{i})\\)\u003c/span\u003e\u003c/span\u003e of the element set form a sentence pair, where the length of the text sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({c}_{i}\\)\u003c/span\u003e\u003c/span\u003e is \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(n\\)\u003c/span\u003e\u003c/span\u003e, and the length of the element example sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({s}_{i}\\)\u003c/span\u003e\u003c/span\u003e is \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(m\\)\u003c/span\u003e\u003c/span\u003e. Input into the Bleem model. The Bleem model will use the three-layer neural network to process the sentence pair \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{i},{s}_{i})\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe first layer is the Bert layer, which is implemented using the Bert-Chinese. Before inputting this layer, the sentence pairs \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{i},{s}_{i})\\)\u003c/span\u003e\u003c/span\u003e will be connected together and converted into three sequences. The first sequence is the word ID sequence. Each word in the sentence pair \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{i},{s}_{i})\\)\u003c/span\u003e\u003c/span\u003e will be mapped to the word ID through the dictionary counted from all corpora. The second sequence is the marker sequence of front and back sentences, in which the sentence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({c}_{i}\\)\u003c/span\u003e\u003c/span\u003e of document text is converted into a full zero sequence with the same length, and the element example sentence s-i is converted into a full one sequence with the same length. The third sequence is a mask sequence, which is equal to the length of the sentence pair \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(({c}_{i},{s}_{i})\\)\u003c/span\u003e\u003c/span\u003e and is a full 1 sequence.\u003c/p\u003e \u003cp\u003eThe second layer is the attention layer. After obtaining the word vector sequence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(c\\_{ℎ}_{i}\\)\u003c/span\u003e\u003c/span\u003e of the document sentence and the word vector sequence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(s\\_{ℎ}_{i}\\)\u003c/span\u003e\u003c/span\u003e of the element example sentence, connect the two-word vector sequences together, and then input them into a feedforward neural network to learn a hidden layer representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({z}_{i}^{j}\\)\u003c/span\u003e\u003c/span\u003e. Then input the hidden layer representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({z}_{i}^{j}\\)\u003c/span\u003e\u003c/span\u003e into softmax, map the value in the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({z}_{i}^{j}\\)\u003c/span\u003e\u003c/span\u003e vector to the (0,1) range, and output it as the attention weight value \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\alpha }_{i}^{j}\\)\u003c/span\u003e\u003c/span\u003e of the document sentence for the element sentence. The perfect results are shown in Formula (4)~(5).\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$${z}_{i}^{j}=tanℎ\\left({W}_{z}\\right[c\\_{ℎ}_{i};s\\_{ℎ}_{i}\\left]\\right) \\left(4\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$${\\alpha }_{i}^{j}=softmax\\left({z}_{i}^{j}\\right) \\left(5\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({W}_{z}\\)\u003c/span\u003e\u003c/span\u003e is the weight value of the feedforward neural network, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\([c\\_{ℎ}_{i};s\\_{ℎ}_{i}]\\)\u003c/span\u003e\u003c/span\u003e represents the connection of word vector sequences of document sentences and element example sentences. Use the learned document sentences to pay attention to the element sentences \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\alpha }_{i}\\)\u003c/span\u003e\u003c/span\u003e. The word vector sequence \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(c\\_{ℎ}_{1}^{i}\\)\u003c/span\u003e\u003c/span\u003e of the document sentence is weighted to obtain the hidden layer representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({ℎ}_{i}\\)\u003c/span\u003e\u003c/span\u003e of the rewritten document sentence.\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$${ℎ}_{i}=\\sum _{k=1}^{m}{\\alpha }_{jk}^{i}\\bullet c\\_{ℎ}_{k}^{i} \\left(6\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(m\\)\u003c/span\u003e\u003c/span\u003e represents the length of the document sentence, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(c\\_{ℎ}_{k}^{i}\\)\u003c/span\u003e\u003c/span\u003e represents the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(ktℎ\\)\u003c/span\u003e\u003c/span\u003e word in the document sentence, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\alpha }_{jk}^{i}\\)\u003c/span\u003e\u003c/span\u003e is the corresponding attention weight.\u003c/p\u003e \u003cp\u003eAt the same time, the cross-entropy function is used to construct the loss function of the Bleem model. The specific form is shown in Formula (7).\u003cdiv id=\"Equg\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equg\" name=\"EquationSource\"\u003e\n$$loss=-\\frac{1}{n}\\sum _{i=1}^{n}\\left[{y}_{i}\\text{log}\\left({logit}_{i}\\right)+\\left(1-{y}_{i}\\right)\\text{log}\\left(1-{logit}_{i}\\right)\\right] \\left(7\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({y}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the true value, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({logit}_{i}\\)\u003c/span\u003e\u003c/span\u003e is the predicted value, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(n\\)\u003c/span\u003e\u003c/span\u003e is the number of samples.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Experimental results and analysis","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Experimental data\u003c/h2\u003e \u003cp\u003eThe CAIL2020 abstract dataset (Yuan J, 2021) \u003csup\u003e[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e used in the experiment totaled 9848 civil first instance judgments. The judgment document is divided into several sentences in advance, each sentence is marked with the label of whether it is important, and the corresponding full text of the reference abstract is provided. The types of documents include civil disputes such as tort liability, lease contracts, labor contracts, inheritance contracts, and loan contracts. The documents average 2568 words, with a maximum length of 13060 words. 95% of the documents exceed 4663 words. The summary average 283 words, with a maximum length of 474 words. 95% of the documents exceed 327 words. This data set belongs to the long text summary data set. The labels of the elements of the document and the sentences to be classified of the examples are given by the legal experts. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows some of the element examples.\u003c/p\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\u003eElements of divorce proceedings\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=\"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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExample sentences of elements\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eType\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExample sentences of elements\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHaving children after marriage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSeparation after marriage\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePersonal property before marriage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFailure to perform the divorce agreement\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChild support with limited capacity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSue for divorce again\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStatutory divorce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDamages\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHaving joint property of husband and wife\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMonthly alimony payment\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFailure to fulfill family obligations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTwo years of separation due to emotional discord\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePay alimony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDivorce granted\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThere are children born out of wedlock\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eChildren living with non-custodians\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal estate division\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eJoint debt of husband and wife\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAppropriate help\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePersonal property after marriage\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\u003eFrom Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, we can see that the original dataset has a small amount of data, but there are many categories classified. At the same time, the sentence of the document does not only contain one element semantics, and there is a case where one sentence corresponds to multiple elements. In order to solve the above problem, this study will convert the multi classification dataset into the text implication dataset. Combine the document sentence with the element example sentence corresponding to the label to build \"implication\" example and \"non implication\" example respectively. Use synonymous sentences to enhance data, so that one element can correspond to about three synonymous sentences, which increases the sample size of the dataset. At the same time, the combination of document sentences and element example sentences in pairs can make the document sentences correspond to multiple categories, giving more opportunities to explore the semantics of different elements in the document sentences. Data enhancement scheme based on synonymous semantics is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eAfter dataset transformation, the sample of feature extraction data includes three fields, including premise, hypothesis and implication label. The premise is the document sentence to be classified in the original data set. Assume that it is an element example sentence corresponding to a category, or a synonymous sentence of an element example sentence. The implication tag is 0 or 1. When it is 0, the document sentence does not contain the semantics of element example sentences. When it is 1, the document sentence contains the semantics of element example sentences. The overall data set size is: the training set contains 52479 samples, the validation set contains 17494 samples, and the test set contains 17494 samples.\u003c/p\u003e \u003cp\u003eFor the converted data, we calculated the length proportion of document sentences and element sample sentences in all samples, and the final result is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. It can be seen that most of the document sentences are within 100 and most of the element example sentences are within 28.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Performance comparison and analysis of different models\u003c/h2\u003e \u003cp\u003eIn order to compare the performance of Bleem model and other models in extracting document elements, this study selected text similarity model (Islamaj R, 2019) \u003csup\u003e[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e, BI CNN (Yin W, 2015) \u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e, ABCNN (Type 3) (Yin W, 2016) \u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e, match_ Lstm (Wang X, 2021) \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e, Bert (Uthirapathy S E 2023) \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e serve as the baseline. For LSTM class model and CNN model, word2vec word vector is used as the embedded layer uniformly, and the length of the embedded vector is 300. At the same time, word level embedded vector is used. For models that use LSTM layers, the hidden layer size of LSTM is set to 128. For the model using the CNN layer, the filter size of CNN is set to 50, and the window size of the filter is 3. At the same time, in order to reduce over fitting, the Dropout policy is used for these two types of models, and the dropout value is set to 0.8. At the same time, set the batch size of network post transmission to 128 and the number of training rounds to 20. The optimizer of LSTM class model and CNN class model uses Adam algorithm, and the initial learning rate is 0.0001.\u003c/p\u003e \u003cp\u003ePut the above model and Bleem model together on the document element extraction dataset, and verify these models. The experimental results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\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\u003eResults of comparative experiments\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=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTraining duration\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSimilar\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.4448\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5074\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBI-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.525\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4m 40s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eABCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.6365\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5221\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9m 34s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ematch_lstm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.6419\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.591\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1h 54m\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBert\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9563\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9565\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e65m 43s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBleem(Our)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9612\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e68m 23s\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\u003eIn this study, accuracy and F1 score were used as evaluation indicators of the model. Among them, accuracy is used to evaluate the prediction effect of the model, and F1 score is used to evaluate the stability of the model effect from the two dimensions of recall and precision. At the same time, this study recorded the training duration of each model to evaluate the time complexity of the model.\u003c/p\u003e \u003cp\u003eFrom Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, we can see that Similar model has the minimum training time. However, because its exploration of text semantics is too shallow, it has the worst accuracy in predicting text implication relations and F1 scores. Moreover, the accuracy score of the Similar model is less than 0.5, which is lower than the random guess. It can be seen that the Similar model is not suitable for the analysis task of document elements.\u003c/p\u003e \u003cp\u003eAt the same time, from the perspective of accuracy and F1 score, CNN model and RNN model have the same effect. In the simple model, the F1 scores and accuracy scores of Bi CNN and Bi LSTM are between 0.5 and 0.6, which is better than random guess and Similar model. Based on the word level task, CNN model analyzes the local features of the text with a sliding window with a length of 3 in the document element extraction task, which can combine and filter out the potential 3-gram features to obtain the local semantics of words. At the same time, the LSTM model involves the semantic representation of words before the text sequence in the semantic computation of words at the current position, so that the previous word semantics can affect the semantic representation of words at the current position, which enables LSTM to analyze the long-distance dependence and short-term dependence in the text sequence, and obtain more rich context information. Because of their respective advantages, CNN and LSTM have both been used in text classification tasks, and have achieved similar performance in the task of extracting document elements in this study.\u003c/p\u003e \u003cp\u003eThe attention mechanism can span the influence of time step interval, directly align two text sequences, calculate their dependency on long distance time steps, and map this dependency to the attention weight value to weight the text sequence, so that the text sequence contains a more long-distance semantic association. After further adding the attention mechanism to the ABCNN and march LSTM models, their performance also shows a synchronous improvement. However, CNN uses the strategy of sparse interaction and weight sharing. For the previous sequence, the convolution kernel only connects with some neurons of the previous layer, rather than the full connection form of the full connection layer and LSTM layer, which greatly reduces the calculation parameters and the time complexity of CNN. At the same time, each convolution kernel only needs to train a set of weights, and then the weights can be shared among each convolution kernel, which further reduces the complexity of the model.\u003c/p\u003e \u003cp\u003eIn the training process of this study, the training duration of the two types of training models shows their differences in time complexity. CNN type models end their training in 5 minutes and 10 minutes respectively, while LSTM model takes 2 hours. In terms of accuracy and F1 score, Bert model and Bert based Bleem model have greatly improved compared with ABCNN and match LSTM models based on Attention. In order to obtain further effect improvement on the Bert model, this study uses Attention and pooling layer as a modeling layer after the Bert model. Attention is responsible for aligning the word vector sequence of the obtained element example sentences and document sentences, so as to mine the relationship between the two, and convert the relationship into the attention weight value, weighting the original word vector representation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Visualization of attention mechanism\u003c/h2\u003e \u003cp\u003eIn order to further analyze the effectiveness of the attention mechanism in the Bleem model, this study visualizes the Attention weight calculated in the Bleem model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the images shown, there are corresponding examples of three types of cases, namely, \"divorce litigation\", \"loan dispute\" and \"labor dispute\". The specific results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e: the gradual change from purple to yellow indicates the attention weight value in the range of [0,1], and close to yellow indicates the attention weight value is high.\u003c/p\u003e \u003cp\u003eIn the first example, the element example sentence given is \"sue for divorce for the second time\". The visualization results of divorce litigation examples is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. The attention weight value learned by the model in the training process focuses on the document sentence fragment \"sue, then withdraw the lawsuit, after withdrawing the lawsuit, the original,\". It is not difficult to see that the model has fully understood the meaning of \"secondary prosecution\" and paid attention to the process of \"secondary prosecution\" corresponding to \"litigation\" and \"withdrawal of litigation\".\u003c/p\u003e \u003cp\u003eIn the second example, the element example given is \"signing the agreement on the transfer of rights and interests of creditor's rights assets\". The Visualization results of loan dispute samples is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e. In this example, the sentence fragment of the document that the model pays attention to is \"the creditor's rights have been transferred to the plaintiff Weize Company and have been notified\". The keywords \"creditor's rights\" and \"transfer\" in the element example sentences also appear in the document sentence fragments concerned by the model, and the model further finds the entities related to transfer.\u003c/p\u003e\u003cp\u003eIn the third example, the element example given is \"salary arrears\". The visual Results of Examples of Labor Disputes is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. In this example, the part with high attention weight given by the model is the opening part of the document sentence \"The plaintiff's salary in arrears\". This segment has a direct semantic relationship with the element example sentence. The model also finds this segment directly and excludes the relationship between other segments and the element example sentence.\u003c/p\u003e \u003cp\u003eBased on the above three examples, it is not difficult to find that the proposed Bleem model can well find the parts with semantic associations between document sentences and element examples in the task of document element extraction, and use these fragments to make efficient judgments on the implication of text elements.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Design of intelligent legal text analysis and retrieval system","content":"\u003cp\u003eThis research is based on the extraction model of legal document elements contained in the text - Bleem, which solves the problem of element information homogeneity in the extraction of legal document elements, and realizes the abstract information extraction of legal documents. Based on this model, this study designed an intelligent legal text analysis and retrieval system to help legal researchers obtain the summary and reading results of legal texts and support legal research.\u003c/p\u003e \u003cp\u003eThe overall architecture of the system is divided into three parts: document crawling module, automatic reading module, and analysis tool module. The overall architecture of the intelligent analysis system for legal documents is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e. Among them, the document crawling module mainly helps legal researchers to automatically obtain the required document data from the public legal document website. The automatic reading module is the core part of the whole system, which mainly includes the machine reading understanding part, the document element analysis part, the result fusion part and the result post-processing part. The main purpose of the analysis tool module is to provide legal researchers with data analysis tools, such as simple statistical tools, regression analysis tools, cluster analysis tools, etc.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Document crawling module\u003c/h2\u003e \u003cp\u003eThe crawler part, as the input part of the system, is mainly composed of three steps. The first step is to execute the crawler. The crawler program will automatically crawl relevant legal documents from the litigation website according to the case keywords entered by legal researchers. The second step is case similarity detection and case clustering. In this step, the system will first use the \"Case Similarity Detection\" model to calculate the case similarity between legal documents and get the similarity matrix of documents. Then, input the similarity matrix into k-mean clustering algorithm to cluster documents. The third step is the screening of instrument clusters. According to the clustering results, the system will output a representative instrument of each category cluster. After reading the representative instruments of this cluster, legal researchers will determine the relevance of this cluster of instruments to their own research, and then determine whether this cluster of instruments is retained. Thus, all collections of documents to be read can be obtained. The flow chart of the document crawling module is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Automatic reading module\u003c/h2\u003e \u003cp\u003eThe automatic reading module is divided into four parts: machine reading understanding part, document element analysis part, result fusion part and result post-processing part. The flow chart of this module is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe machine reading and understanding part implements the reading and understanding of legal documents. First of all, legal researchers designated research questions, such as \"how much is the amount involved\" and \"when is the time of crime\". Then, according to the collection of documents to be read obtained from the crawler part and the questions given by the researchers, with the help of the pre trained Bleem model, perform sentence level document element analysis, and output document sentences containing element semantics. In the result fusion part, machine reading comprehension and result fusion will be fused. The Bleem model is used to analyze the elements of the document. After the key content of the document is extracted, the results of machine reading comprehension are enhanced by judging the correlation between the elements of the document and the problem. The post-processing part aims to extract the complete answer more accurately. Classify the questions according to some key words in the questions, such as \"who\" and \"when\". Then, post process the answers of machine reading comprehension according to the categories of questions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Analysis tool module\u003c/h2\u003e \u003cp\u003eThe analysis tool module provides various types of analysis tools for legal researchers. The activity diagram of the analysis tool module is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eThe first is simple statistics, such as the average, variance and median of the amount involved, and the proportion of cases of \"whether there is torture\". The second is regression analysis. For example, for network crimes, you can use regression analysis tools to judge the correlation between the number of information stolen and the sentence. The third is cluster analysis. After the case elements are extracted, all the current cases can be clustered according to the case elements to judge whether there are subdivision cases in the current category of cases. Legal researchers can find more legal phenomena and put forward research ideas better.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Application in comprehensive governance of public risk","content":"\u003cp\u003eHow to use science and technology to deal with and prevent public risks facing mankind has become a major topic of the modernization of governance systems and capabilities in various countries, and also a topic of great concern for industry and academia. This study applies the extraction model of legal document elements to the comprehensive governance of public risk, randomly samples four types of cases from 200 indictments for analysis, and compares the results of the proposed Bleem model with ABCNN, match-lstm, Bert and other models.\u003c/p\u003e\u003cp\u003eIt can be seen from the four cases shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e\u0026thinsp;~\u0026thinsp;18 that the summary generation of the Bleem model proposed in this study is the best, and the generated summary is relatively standard and readable, indicating that the model can quickly learn the summary mode of various criminal cases. In the case of the crime of causing trouble and intentional homicide, Bert did not recognize it well, and even in the case of intentional homicide, there was a lot of repetition. In the case of multiple crimes, other comparative models failed to identify key information such as concealing and concealing crimes, which indicates that these models have poor effect on the generation of multiple crimes summaries.\u003c/p\u003e \u003cp\u003eThe above analysis shows that the Bleem model proposed by us can improve the rapid and accurate identification of criminal elements in criminal facts and is of great significance to the comprehensive management of public risks for crimes that are easy to lead to public risks, such as theft, intentional homicide and provocation.\u003c/p\u003e"},{"header":"7 Conclusions","content":"\u003cp\u003eThe legal intelligence derived from the combination of AI technology and legal theory has had a profound impact on the field of law. AI technology has emerged in a large number of applications in case analysis, intelligent court platforms, evidence collection, etc. However, the homogeneity of element information in legal documents and the diversity of problems in legal reading hinder the application of AI technology. In order to solve these problems, this research constructed a document element extraction model based on text implication - Bleem. This model used the Bert layer as the coding layer to obtain the context sensitive word vector representation, and then used Attention and pooling layers to model the global features, and used the global features to judge whether the document sentence contains element semantics. In the experimental results shown, Bleem model achieved 96% accuracy on the dataset, which was superior to other models. At the same time, our research used the attention weight visualization to prove that our Bleem model can well focus on the sentence fragments associated with document sentences and element example sentences.\u003c/p\u003e \u003cp\u003eThen, this research designed an intelligent analysis system for legal documents, including three parts: document crawling, reading analysis, and analysis tools, to help legal researchers obtain the summary information and reading understanding information of legal documents. Finally, we applied the model of extracting elements of legal documents to the comprehensive governance of public risks. The experimental results show that our model can quickly and accurately identify the criminal elements in the criminal facts for crimes that are easy to lead to public risks, such as theft, intentional killing, provocation, etc.\u003c/p\u003e \u003cp\u003eOur research can effectively promote the application of legal services in the comprehensive governance of public risks. From the experimental results, the Bleem model has achieved high accuracy, but further discussion is needed on the interpretability. Next, we will carry out the implementation deployment of the model and verify the effect of the intelligent legal text analysis and retrieval system in practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have not disclosed any funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEnquiries about data availability should be directed to the authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict Interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe paper does not deal with any ethical problems.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed consent\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe declare that all the authors have informed consent.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eYu J, Xia J. E-justice evaluation factors: The case of Smart Court of China[J]. \u003cem\u003eInformation Development,\u003c/em\u003e 2021, 37(4): 658-670.\u003c/li\u003e\n\u003cli\u003eXinlei C A I, Rongfu Q I U, Yanli C U I, et al. Power Grid Auxiliary Control System Based on Big Data Application and Artificial Intelligence Decision[C]//\u003cem\u003e2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE).\u003c/em\u003e IEEE, 2020: 154-157.\u003c/li\u003e\n\u003cli\u003eZhukova M N, Zolotarev V V, Zhukov V G, et al. Service robot security from unauthorized access by connection control[C]//\u003cem\u003e2019 12th International Conference on Developments in eSystems Engineering (DeSE).\u003c/em\u003e IEEE, 2019: 526-529.\u003c/li\u003e\n\u003cli\u003eTaran M O, Revunkov G I, Gapanyuk Y E. The text fragment extraction module of the hybrid intelligent information system for analysis of judicial practice of arbitration courts[C]//\u003cem\u003eInternational Conference on Neuroinformatics. \u003c/em\u003eSpringer, Cham, 2020: 242-248.\u003c/li\u003e\n\u003cli\u003eUthirapathy Samson Ebenezar,Sandanam Domnic. Topic Modelling and Opinion Analysis On Climate Change Twitter Data Using LDA And BERT Model.[J]. \u003cem\u003eProcedia Computer Science,\u003c/em\u003e2023,218.\u003c/li\u003e\n\u003cli\u003eHu W, Hu R, Xie N, et al. Image classification using multiscale information fusion based on saliency driven nonlinear diffusion filtering[J]. \u003cem\u003eIEEE transactions on image processing, \u003c/em\u003e2014, 23(4): 1513-1526.\u003c/li\u003e\n\u003cli\u003ePaul C, Rettinger A, Mogadala A, et al. Efficient graph-based document similarity[C]//\u003cem\u003eEuropean Semantic Web Conference.\u003c/em\u003e Springer, Cham, 2016: 334-349.\u003c/li\u003e\n\u003cli\u003eVoeste Anja. Spelling variation and text alignment[J]. Zeitschrift f\u0026uuml;r Sprachwissenschaft,2021,40(3).\u003c/li\u003e\n\u003cli\u003eWaszek D, Schlimm D. Calculus as method or calculus as rules? Boole and Frege on the aims of a logical calculus[J]. Synthese, 2021, 199(5): 11913-11943.\u003c/li\u003e\n\u003cli\u003eGhadage Y H, Shelke S D. Speech to text conversion for multilingual languages[C]//\u003cem\u003e2016 International Conference on Communication and Signal Processing (ICCSP).\u003c/em\u003e IEEE, 2016: 0236-0240.\u003c/li\u003e\n\u003cli\u003ePark B, Colaresi M, Greene K. Beyond a bag of words: Using PULSAR to extract judgments on specific human rights at scale[J]. \u003cem\u003ePeace Economics, Peace Science and Public Policy, \u003c/em\u003e2018, 24(4).\u003c/li\u003e\n\u003cli\u003eLi L, Qin B, Liu T. Contradiction detection with contradiction-specific word embedding[J]. \u003cem\u003eAlgorithms,\u003c/em\u003e 2017, 10(2): 59.\u003c/li\u003e\n\u003cli\u003eUpmeier zu Belzen A, Engelschalt P, Kr\u0026uuml;ger D. Modeling as scientific reasoning\u0026mdash;The role of abductive reasoning for Modeling competence[J]. \u003cem\u003eEducation Sciences,\u003c/em\u003e 2021, 11(9): 495.\u003c/li\u003e\n\u003cli\u003eRhanoui M, Mikram M, Yousfi S, et al. A CNN-BiLSTM model for document-level sentiment analysis[J]. \u003cem\u003eMachine Learning and Knowledge Extraction,\u003c/em\u003e 2019, 1(3): 832-847.\u003c/li\u003e\n\u003cli\u003eDave H, Jaswal S. Multiple text document summarization system using hybrid summarization technique[C]//\u003cem\u003e2015 1st International Conference on Next Generation Computing Technologies (NGCT).\u003c/em\u003e IEEE, 2015: 804-808.\u003c/li\u003e\n\u003cli\u003eLi Z, Cai J, He S, et al. Seq2seq dependency parsing[C]//\u003cem\u003eProceedings of the 27th International Conference on Computational Linguistics.\u003c/em\u003e 2018: 3203-3214.\u003c/li\u003e\n\u003cli\u003eXie, Shufang,Xia, Yingce,Wu, Lijun,Huang, Yiqing,Fan, Yang,Qin, Tao. End-to-end entity-aware neural machine translation[J]. \u003cem\u003eMachine Learning,\u003c/em\u003e2022(prepublish).\u003c/li\u003e\n\u003cli\u003eYu Tianxiang,Xin Yang,Zhu Hongliang,Tang Qifeng,Chen Yuling. Network Penetration Intrusion Prediction Based on Attention Seq2seq Model[J]. \u003cem\u003eSecurity and Communication Networks,\u003c/em\u003e2022,2022.\u003c/li\u003e\n\u003cli\u003eAlaa Joukhadar,Nada Ghneim,Ghaida Rebdawi. Impact of Using Bidirectional Encoder Representations from Transformers (BERT) Models for Arabic Dialogue Acts Identification[J]. ISI,2021,26(5).\u003c/li\u003e\n\u003cli\u003eZhang N N, Xing Y. Questions and answers on legal texts based on BERT-BiGRU[C]//Journal of Physics: Conference Series. \u003cem\u003eIOP Publishing,\u003c/em\u003e 2021, 1828(1): 012035.\u003c/li\u003e\n\u003cli\u003eYuan J, Wei Z, Gao Y, et al. Overview of SMP-CAIL2020-Argmine: The Interactive Argument-Pair Extraction in Judgement Document Challenge[J]. \u003cem\u003eData Intelligence,\u003c/em\u003e 2021, 3(2): 287-307.\u003c/li\u003e\n\u003cli\u003eIslamaj R, Wilbur W J, Xie N, et al. PubMed Text Similarity Model and its application to curation efforts in the Conserved Domain Database[J]. \u003cem\u003eDatabase,\u003c/em\u003e 2019, 2019.\u003c/li\u003e\n\u003cli\u003eYin W, Sch\u0026uuml;tze H. Convolutional neural network for paraphrase identification[C]//\u003cem\u003eProceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. \u003c/em\u003e2015: 901-911.\u003c/li\u003e\n\u003cli\u003eYin W, Sch\u0026uuml;tze H, Xiang B, et al. Abcnn: Attention-based convolutional neural network for modeling sentence pairs[J]. \u003cem\u003eTransactions of the Association for Computational Linguistics, \u003c/em\u003e2016, 4: 259-272.\u003c/li\u003e\n\u003cli\u003eWang Xiaomin,Wang Haoriqin,Zhao Guocheng,Liu Zhichao,Wu Huarui. ALBERT over Match-LSTM Network for Intelligent Questions Classification in Chinese[J]. \u003cem\u003eAgronomy, \u003c/em\u003e2021,11(8).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"soft-computing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"soco","sideBox":"Learn more about [Soft Computing](https://www.springer.com/journal/500)","snPcode":"500","submissionUrl":"https://submission.nature.com/new-submission/500/3","title":"Soft Computing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Social public risk, Legal intelligence, Legal text analysis and retrieval, Bleem model, Bert model","lastPublishedDoi":"10.21203/rs.3.rs-2994403/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-2994403/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe amalgamation of information retrieval systems and soft computing techniques establishes a robust framework to confront the challenges and seize the opportunities presented by the vast expanse of big data. As the volume, diversity, and velocity of data continue to proliferate, further advancements in this domain are poised to significantly contribute to various realms, encompassing healthcare, finance, e-commerce, and scientific research, ultimately propelling innovation and facilitating decision-making in the epoch of big data. The application of artificial intelligence (AI) technology to the analysis and comprehension of legal documents holds the potential to expedite the acquisition of case-specific information by legal researchers, thereby expediting their research endeavors. This paper proposes a legal text analysis and retrieval model, rooted in the Bleem model (Bert-based Legal Paper's Element Extracting Model). Initially, our model employs Bert as the coding layer to extract the semantic information embedded within document sentences and element exemplars. Subsequently, we leverage the Attention mechanism to align the semantic essence of element example sentences with document sentences, while simultaneously computing their respective attention weights. Comparative experiments and attention visualization are then employed to validate the efficacy of the Bleem model. The experimental results corroborate the superiority of the Bleem model in terms of accuracy and F1 scores. The visualization of the attention mechanism effectively reveals the inner workings of the Bleem model and unveils its capacity to explore pertinent fragments within document sentences and element examples. Building upon the aforementioned model, we have devised an intelligent legal text analysis and retrieval system, empowering legal researchers to swiftly acquire pivotal data through case briefs. This application has effectively fostered the integration of legal services within the holistic management of public risks.\u003c/p\u003e","manuscriptTitle":"Design of intelligent legal text analysis and information retrieval system based on BERT model","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2023-06-07 20:20:34","doi":"10.21203/rs.3.rs-2994403/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2023-06-06T08:08:32+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2023-06-06T05:36:05+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2023-06-02T19:05:40+00:00","index":"","fulltext":""},{"type":"submitted","content":"Soft Computing","date":"2023-05-29T04:11:31+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"soft-computing","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"soco","sideBox":"Learn more about [Soft Computing](https://www.springer.com/journal/500)","snPcode":"500","submissionUrl":"https://submission.nature.com/new-submission/500/3","title":"Soft Computing","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"67fe68f1-6844-47c6-872b-f7e363b744ad","owner":[],"postedDate":"June 7th, 2023","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2023-06-07T20:20:34+00:00","versionOfRecord":[],"versionCreatedAt":"2023-06-07 20:20:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-2994403","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-2994403","identity":"rs-2994403","version":["v1"]},"buildId":"cBFmMYwuxLRRLfASyISRj","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}