Legal document entity extraction method based on Chinese semantic enhancement
By using multimodal feature fusion and entity-aware coding correction method in Chinese entity extraction in the judicial field, the problems of unsatisfactory characterization effect and insufficient entity distinction are solved, and efficient entity recognition and Improvement of fine-grained recognition.
Patent Information
- Application Number
- CN202410987528.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-12
AI Technical Summary
Chinese entity extraction faces the problems of unsatisfactory representation of rare characters and insufficient entity distinction in the judicial field, especially when fine-grained entity recognition is required, the existing technology is difficult to meet the high rigor and complexity requirements of the judicial field.
Using the legal document entity extraction method based on Chinese semantic enhancement, the multimodal feature fusion representation is used, and the word-level information of paramorphic and word segmentation information is fused with attention mechanism, and the entity perception coding correction device is introduced to enhance entity distinction and at a high level. Adaptive learning entity category coding in semantic space to improve the accuracy of entity recognition.
It significantly improves the extraction effect and entity distinction of Chinese entity extraction in the judicial field, improves the accuracy of fine-grained entity recognition, meets the diversified needs of the judicial field for entity extraction, and realizes efficient identification on judicial data sets.
Smart Images

Figure CN118940759A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of judicial intelligence, and in particular to a legal document entity extraction method based on Chinese semantic enhancement. Background Art
[0002] In recent years, artificial intelligence and deep learning have been increasingly used in various vertical fields. In the judicial field, the research on artificial intelligence mainly focuses on legal theory, ethical issues, algorithmic risks, etc. How to deeply integrate law and technology to create intelligent applications of logical reasoning and decision-making capabilities is a practical need for smart justice. Faced with massive and complex legal documents, intelligent entity extraction can largely solve problems such as inaccurate labeling, unclear recognition, and low efficiency. Furthermore, entity extraction plays an important role in tasks such as legal question and answer, relationship extraction, judicial decision prediction, similar case retrieval, and judicial knowledge graph construction.
[0003] Entity extraction requires inputting text of any length and identifying the boundaries and types of different entities based on the extracted word representations. This sequence labeling task requires a deep exploration of word embedding representations. After the emergence of pre-trained models such as BERT (Bidirectional Encoder Representations from Transformers), thanks to its semantic mining and grammatical mastery capabilities acquired after extensive text learning, word embedding acquisition for Chinese entity extraction has shifted to extraction through pre-trained models.
[0004] In the judicial field, Chinese entity extraction uses the BERT+BiLSTM+CRF framework, such as Figure 1As shown in the figure, BERT and other pre-trained models are used to extract word embeddings, LSTM (Bidirectional Long Short-Term Memory) learns semantic information that is beneficial to the task in context, and CRF (Conditional Random Field) can correct the annotations in combination with the transfer relationship between entity category labels to obtain logically smooth entity recognition sequence annotation results. This method performs well in tests on Chinese legal document datasets such as the Judgment Documents Network, public drug-related cases, and prosecution opinions. However, unlike the general field, Chinese entity extraction faces greater challenges in the judicial field. First, due to differences in case types and specific scenarios, entities in legal documents often involve uncommon characters. The extraction effect of these word embeddings may not be ideal if they only rely on pre-trained models, and additional information enhancement is required. Secondly, the legal entity extraction task is oriented to business applications and requires more fine-grained entity divisions, such as suspects and victims, while in the general field they are all names, so this requires the model to have higher entity discrimination. In addition, the high rigor in the judicial field leads to more standardized and detailed entity descriptions. In the general field, "xx city xx village" will be described as "xx street xx community xx unit xx room". The large-span entities reduce the performance of the general field model.
[0005] Therefore, technicians in this field are committed to developing a legal document entity extraction method based on Chinese semantic enhancement. The solution of the present invention, JMCNER (Judicial Multimodal Chinese Named Entity Recognition), combines multimodal information to enhance the quality of word embedding representation, and introduces an entity perception module to correct the encoding, so that the entity distinction is further widened to adapt to the particularity of entity extraction in the judicial field. Summary of the invention
[0006] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to improve the extraction effect of Chinese entity extraction in the judicial field and increase the entity differentiation.
[0007] To achieve the above objectives, the present invention provides a legal document entity extraction method based on Chinese semantic enhancement, including multimodal feature fusion representation and word-level information fusion of radical and word segmentation information based on an attention mechanism.
[0008] Furthermore, the word positions reserved by BERT-like models are used as word segmentation tokens to extract word embeddings combined with word segmentation information.
[0009] Furthermore, the radical features are extracted to enrich the representation of rare characters.
[0010] Furthermore, the representation is mainly based on character embedding, supplemented by radical features, to obtain a fused representation with information enhancement.
[0011] Further, the following steps are included:
[0012] Step 1: Multimodal feature fusion representation;
[0013] Step 2: high-dimensional semantic encoding;
[0014] Step 3: Entity category decoding.
[0015] Furthermore, the step 1 comprises the following steps:
[0016] Step 1.1: Preprocessing of legal texts;
[0017] Step 1.2, multimodal feature extraction;
[0018] Step 1.3: Multimodal feature fusion.
[0019] Furthermore, the legal text preprocessing includes text normalization, radical extraction, and introduction of word segmentation information.
[0020] Further, an entity-aware encoding corrector is included.
[0021] Furthermore, the entity-aware encoding modifier initializes the encoding of the entity category in the high-dimensional semantic space.
[0022] Furthermore, the entity-aware encoding corrector adaptively learns the encoding of entity categories in a high-dimensional semantic space to increase the encoding differences between different entity categories.
[0023] The specific text distribution of legal documents in the judicial field, relying solely on pre-trained models to extract the representation of uncommon characters is not satisfactory, and additional information enhancement is required. The present invention integrates radicals and word-level information containing segmentation information based on the attention mechanism, greatly enhancing the representation ability of uncommon characters.
[0024] Multimodal feature extraction: word segmentation information enhancement has been widely used in the Chinese field. Word information helps to enhance word meaning and eliminate ambiguity. The present invention uses the word unit position reserved by the BERT-type model as a word segmentation symbol, and relies on the semantic mining ability of the pre-trained model to extract word embedding combined with word segmentation information; on the other hand, Chinese characters are pictographic characters and their shapes can also express meanings. The radicals are related to the meaning of the characters. Extracting radical features can enrich the representation of rare characters, while also avoiding the heavy cost of learning the overall character shape.
[0025] Multimodal feature fusion: The final representation is based on character embeddings, while radical features are used as auxiliary enhancements. Combining the mechanism principles behind the attention mechanism, the representation is fused with radical features as queries and character embeddings as key-value pairs. This enhances the embedding similarity of characters with the same radical, and uses the residual structure to ensure the core position of character embeddings, thereby ultimately obtaining an information-enhanced fusion representation.
[0026] The present invention conducts ablation experiments on multimodal information features, compares them with the BERT+CRF baseline, and further compares them with only word segmentation information or radical information, and verifies the effectiveness of multimodal features on a judicial dataset. For multimodal feature fusion, compared with feature concatenation and multi-layer perceptron linear relationship aggregation commonly used in other methods, the fusion representation obtained by the attention mechanism used in the present invention achieves better results.
[0027] The entity extraction task in legal documents requires finer category granularity and more precise expression scope. For example, the fine-grained distinction between suspects and victims or address information specific to house numbers, compared with the general field, the model has challenges in identifying and dividing the entity boundaries of these subtle differences. The present invention introduces an entity-aware encoding corrector, adaptively learns the encoding of each entity category in a high-dimensional semantic space, and adjusts the encoding based on this, thereby increasing the encoding differences between different entity categories, thereby enhancing the entity distinction. The fine-grained and precise legal entity requires a higher entity distinction. To this end, the present invention initializes an encoding in a high-dimensional semantic space for each entity category. The previous high-quality fusion representation is encoded after BiLSTM learns syntactic and semantic information, and calculates similarity with it, narrowing the distance between its possible entity categories in the semantic space, thereby increasing its distance from other entity encodings, and ultimately enhancing the entity distinction. The entity boundary will also become clearer and more obvious. The present invention compares the performance with and without the encoding corrector to illustrate the effectiveness of the module.
[0028] The present invention uses a highly modular model to develop applications to ensure the high availability and scalability of the system. It mainly includes a data preprocessing module, a multimodal fusion representation module, a high-dimensional semantic encoding module, an entity category decoding module, and an application interface module. By collecting a large number of relevant annotated data sets from the judicial document network, a supervised learning method is used for model training to enhance the model's adaptability to the judicial vertical field. The system can be deployed in the cloud or locally to meet the needs of different scenarios, and can provide an API (Application Programming Interface) interface for easy integration into existing judicial information systems.
[0029] Compared with the prior art, the present invention has the following obvious substantial features and significant advantages:
[0030] 1. The present invention enhances Chinese semantics based on words and radicals, and uses entity-aware coding modifiers to increase entity differentiation. It is innovative and efficient, can meet the diverse needs of legal document entity extraction, and has broad industrial application prospects. It has excellent performance in technology, performance, production implementation, etc., and provides strong support for the informatization and intelligent development of the legal service industry.
[0031] 2. In order to improve the encoding quality of words in judicial documents such as judgments, the present invention uses words and radicals as auxiliary information to enhance the word embedding obtained by the BERT pre-training model, fully refine local semantics and enrich the representation of word meanings; and combines the attention mechanism to scientifically fuse features, explore the interactive relationship between features and avoid feature redundancy.
[0032] 3. In order to meet the business needs of the judicial field for the identification of fine-grained entities (suspects, victims, etc.), the present invention adopts an entity-aware encoding corrector to adjust the encoding so that the word encoding is closer to the fine-grained entity category to which it belongs in the high-dimensional semantic space, thereby increasing the entity distinction and alleviating the misclassification problem that may occur in the fine-grained entity recognition scenario due to the increase in categories.
[0033] 4. The F1 of the present invention reaches 92.72 on the Chinese judicial dataset CAILIE1.0 (Challenge of AI in Law-Information Extraction V1.0), which is 6.34% higher than the baseline BERT-BiLSTM-CRF, surpassing many Chinese information enhancement solutions. In addition, it is 4.02% higher than the judicial-specific SBENER (Semantic and Boundary Enhance Named Entity Recognition), and the accuracy of each fine-grained entity category is significantly improved.
[0034] 5. The present invention can automatically segment legal documents into sentences and extract entity information in parallel, meeting the needs of efficient processing of large-scale documents.
[0035] 6. The present invention has good adaptability to different types of legal documents (drug-related, theft, etc.) and can be used in different judicial scenarios.
[0036] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is the classic BERT+BiLSTM+CRF entity extraction framework;
[0038] Figure 2 It is a flow chart of a preferred embodiment of the present invention. DETAILED DESCRIPTION
[0039] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.
[0040] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.
[0041] like Figure 2 As shown, in order to solve the problem of entity extraction in legal documents in the Chinese field, the present invention enhances Chinese representation information and increases entity differentiation. The idea of multimodal feature fusion representation in the present invention can be directly extended to other task scenarios that require enhanced Chinese information representation capabilities. For example, entity extraction in other fields with word distribution differences such as biology, medicine, and finance, and other tasks that require Chinese information enhancement, such as intelligent question answering and relationship extraction.
[0042] 1. First embodiment:
[0043] The first embodiment is a multimodal fusion representation + BiLSTM + CRF without an entity-aware encoding modifier.
[0044] based on Figure 1 The framework is enriched, the original BERT-based single network is modified to extract embedding, and the rest of the network structure is retained for the time being.
[0045] 1.1. Multimodal fusion representation:
[0046] i. Legal text preprocessing
[0047] Text normalization: To facilitate subsequent processing, punctuation marks are unified into English form to prevent the impact caused by the mixing of Chinese and English punctuation marks. In addition, some unconventional characters need to be filtered out, such as special emoticons, non-printable characters, etc. The normalized text sequence x is obtained.
[0048] Radical extraction: Obtain the radical sequence x corresponding to the legal document x to be extracted through external dictionary search R。The external dictionary maintains a total of 218 semantic radicals. Different from radicals, semantic radicals unify radicals with the same meaning. For example, both "氵" and "水" belong to "水" because some radicals have evolved for the sake of rich character formation, and their original meanings are represented by semantic radicals. In addition, to classify non-Chinese characters into four categories: "digit [D]", "letter [A]", "punctuation [E]", and "other ASCII code [S]", representative semantic radicals are used as fillers.
[0049] Introduction of word segmentation information: Process the legal document sequence x, and use the widely used jieba word segmentation (an excellent third-party Chinese word segmentation tool based on probabilistic statistical algorithms) to obtain the text word segmentation sequence result, using a special word segmentation symbol " <spa>" represents the word segmentation break point, and adds the starting special word of the BERT-like model to the beginning of the sentence" <cls>"Get the word segmentation text sequence x CW , and record the index sequence x corresponding to the real text I This facilitates subsequent extraction of word embeddings.
[0050] From the radical sequence x R , word segmentation text sequence x CW , real text index sequence x I The subsequent model input is formed. Taking "suspect Li Mou." as an example, the composition of each sequence is as follows, where the label sequence y is an entity label sequence labeled in the "BIO" format ("O" represents a non-entity, "B" and "I" represent the beginning and the rest of the entity respectively, and the entity type is after the hyphen).
[0051] Table 1 Examples of text preprocessing results
[0052] Sequence name Content <![CDATA[Radical sequence x R > ['female', 'pi', 'person', 'wood', 'wood', '[E]'] <![CDATA[Tokenized text sequence x CW > ['[CLS]','suspect', 'Li', '[SPA]', '.'] <![CDATA[True text index sequence x I > [1,2,3,5,6,8] Labeled sequence y ['O', 'O', 'O', 'B-NHCS', 'I-NHCS', 'O']
[0053] ii. Multimodal feature extraction
[0054] Word segmentation text sequence feature extraction:
[0055] Benefiting from the learning ability of BERT-like models in a wide range of texts, the present invention uses the Chinese adaptive optimization version RoBERTa_wwm_ext (RoBERTa_wholewordmasking_ext) of its variant RoBERTa (RobustoptimizeBERTapproach, Robust Optimized BERT)
[0056] E CW =RoBERTa_wwm_ext(X R )ded data, and optimized RoBERTa) on the extended Chinese data using the full-word masking technology to extract features. We further filter special word-unit features based on the real text index sequence to select word embeddings with enhanced word segmentation information. CW , with a dimension of 768.
[0057] The model parameters used in this step are fixed and come from the open source model.
[0058] Radical sequence feature extraction:
[0059] The total number of radicals is 222, consisting of 218 actual radicals + 4 filler radicals. Since the number of radicals is small, a more convenient and efficient multi-layer perceptron can be used to learn their embedding, denoted by E R , the dimension is aligned with the character embedding to 768. Another reason why this scheme is feasible is that the radical itself integrates synonymous radicals. Originally, radicals with the same meaning may have different shapes or occupy different positions due to the needs of character creation, which requires feature extraction from their images to capture the differences, which is unnecessary for the radical.
[0060] E R =tanhW R x R +b R
[0061] This step involves W R , b R is the learnable parameter of the multilayer perceptron, and tanh is the nonlinear activation function used in it.
[0062] iii. Multimodal feature fusion
[0063] In the present invention, the word embedding E enhanced by word segmentation information CW and the radical is embedded in E R Perform attention mechanism cross operation to obtain the final fusion representation E fus Different from the widely used direct concatenation or multi-layer perceptron linear relationship learning, the attention mechanism fusion adopted in the present invention is designed based on the relationship between characters and radicals, and the feature fusion is more semantically oriented.
[0064] Specifically, the present invention uses radical features as query as auxiliary features, word embedding enhanced by word segmentation information as key-value pairs as core features, and obtains the query, key, and value of the attention mechanism through the following formula processing:
[0065] Q fus =tanhW freqQ x R +b freqQ
[0066] K fus =tanhW freqK x CW +b freqK
[0067] V fus =tanhW freqV x CW +b freqV
[0068] Among them, W freqQ , W freqK , W freqV , b freqQ , b freqK , b freqV is a learnable parameter, and tanh is a nonlinear activation function. fus , K fus 、V fus The query, key, and value used by the subsequent attention mechanism.
[0069] Next, the present invention uses the residual structure to enhance the attention mechanism to obtain the fusion representation E fus To further highlight the importance of word embedding, the formula is as follows:
[0070]
[0071] Among them, softmax plays the role of normalizing the similarity score and assigning weights to the values, that is, based on the radical similarity, the word embedding is enhanced to strengthen the representation of individual rare characters in the pre-trained model due to the distribution differences in the judicial vertical field. In addition, the residual structure is combined with the original word embedding to highlight its dominant position, and the final dimension is 768.
[0072] 1.2. High-dimensional semantic coding:
[0073] High-dimensional semantic coding generally uses BiLSTM to learn the contextual dependencies of texts, and can also obtain high-dimensional semantic space representations that are beneficial to vertical field entity extraction as training progresses. In the first embodiment of the present invention, this component is used without change, and its output is recorded as H, with a dimension of 768, and the formula is as follows:
[0074]
[0075] in Indicates the splicing operation, from left to right From right to left The dimensions are all 768, and all parameters can be learned.
[0076] 1.3. Entity category decoding:
[0077] Each word is classified according to the high-dimensional semantic coding obtained previously. For the entity extraction task of the "BIO" annotation type, the number of target categories is twice the number of entity categories plus 1. In general, each sample in a multi-classification task is independent of each other, and there is no need to consider the influence of other samples. Only a simple linear classifier is needed, which is converted into probability through softmax processing and the maximum probability is taken as the predicted category. However, for sequence annotation tasks, it is necessary to consider the association between word units, that is, the samples are no longer independent of each other. Therefore, the present invention uses the widely used CRF for final decoding, and the specific process is as follows:
[0078] A=W A H+b A
[0079] First, the label score matrix A is calculated, which consists of a linear layer where W A and b A To learn the parameters, for a sample sentence pair (x, y), the label sequence score S(x, y) is calculated as follows:
[0080]
[0081] in For labels from y i to y i+1 The transfer score, Divide the i-th word into y i The CRF model describes the probability p(y|x) of predicting y on all possible label sequence sets Y, which is expressed as follows:
[0082]
[0083] in, represents any possible sequence of annotation results, is the probability of predicting the result sequence.
[0084] In the training optimization phase, negative log-likelihood loss is used and gradient back propagation is used to minimize the loss function, where the loss function It is expressed as follows:
[0085]
[0086] Where p(y j |x j ) represents the CRF probability of the jth sentence pair. In the decoding stage, the Viterbi algorithm (a widely used dynamic programming algorithm for solving the shortest path problem) is used to obtain the highest score of the label sequence as the corresponding entity category for prediction.
[0087] 1.4. Performance comparison
[0088] Table 2 Comparison of multimodal information fusion representation and BERT representation on the CAILIE1.0 dataset
[0089]
[0090] As shown in Table 2, with the support of multimodal information fusion representation, the F1 score is improved by 2.19% compared with BERT. Performance improvement is achieved for most categories of fine-grained entities involved, and there is a significant improvement for some non-general entities, such as tools used in the crime.
[0091] 2. Second embodiment:
[0092] The second embodiment adds an entity-aware encoding modifier between the BiLSTM and CRF layers.
[0093] Based on the first embodiment, an entity-aware coding corrector is introduced, and the rest of the network structure is the same as that of the first embodiment.
[0094] 2.1. Multimodal fusion representation:
[0095] E is obtained through the process of the first embodiment fus As a multimodal fusion representation.
[0096] 2.2. High-dimensional semantic encoding:
[0097] In the face of fine-grained, high-precision entity segmentation scenarios in the judicial field, the encoding needs to have higher entity discrimination. To this end, the present invention introduces an entity-aware encoding modifier at the high-level semantic encoding stage based on the rich Chinese representation to meet the scenario requirements. After obtaining the high-dimensional semantic encoding H learned by BiLSTM, the processing flow of the entity-aware modifier is as follows:
[0098] This module is designed by the present invention to have the same number of encodings as the target classification entities, which are used to represent the encoding centers E of different entity categories in the high-dimensional semantic space. entity , calculate the similarity between each entity category encoding and the context encoding, and combine the attention mechanism and the residual structure to correct the encoding, so that the adjusted encoding Z can be closer to the encoding center of the entity category to which it belongs, thereby increasing the entity discrimination.
[0099] Q = tanhW Q H+b Q
[0100] K=tanhW K E entity +b K
[0101] V=tanhW V E entity +b V
[0102] Z=softmaxQK^TV+H
[0103] The formula involved is as above, where W Q , W K , W V , b Q , b K , b V is a learnable parameter, tanh is a nonlinear activation function, Q, K, V are the query, key, and value used by the subsequent attention mechanism, and the residual connection is used to further highlight the importance of the original code H.
[0104] 2.3. Entity category decoding:
[0105] The decoding process is consistent with the first embodiment, and the input for calculating the score matrix is changed from the output H of the BiLSTM to the Z of the entity-aware corrector.
[0106] 2.4. Performance comparison
[0107] Table 3 Comparison of the entity-aware encoding modifier before and after the introduction of the CAILIE1.0 dataset
[0108]
[0109] As shown in Table 3, after the introduction of the entity-aware encoding corrector, the F1 score increased by 1.98% compared to before the introduction, with significant performance improvements in almost all categories.
[0110] Table 4. Comparison of overall performance on the CAILIE1.0 dataset
[0111] Model Pre% Rec% F1%(macro) BERT-CRF 82.85 86.65 83.93 BERT-BiLSTM-CRF 86.46 86.11 86.38 PromptNER 85.62 87.22 86.41 MECT 85.21 86.34 85.77 Lattice-LSTM 83.60 87.07 85.30 FLAT-Lattice 85.53 88.23 86.86 LEBERT 87.28 87.56 87.42 SBENER 88.31 89.10 88.70 JMCNER 93.22 92.45 92.72
[0112] As shown in Table 4, the JMCNER of the present invention performs best compared with other named entity recognition schemes for Chinese (where BERT-CRF and BERT-BiLSTM-CRF are baseline schemes, PromptNER is a novel scheme for extraction in prompt form, MECT is a scheme for enriching word meaning representation, and Lattice-LSTM, FLAT-Lattice, and LEBERT are word enhancement schemes), and also surpasses another legal-oriented named entity recognition scheme SBENER, with an F1 score improvement of 4.02%.
[0113] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.< / cls> < / spa>
Claims
1. A legal document entity extraction method based on Chinese semantic enhancement, characterized in that: It includes multimodal feature fusion representation and word-level information that integrates radical and word segmentation information based on the attention mechanism.
2. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 1 is characterized in that: The word positions reserved by BERT-like models are used as word segmentation tokens to extract word embeddings combined with word segmentation information.
3. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 1 is characterized in that: Extract radical features to enrich the representation of rare characters.
4. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 1 is characterized in that: The representation is mainly based on character embedding, supplemented by radical features, to obtain a fused representation with information enhancement.
5. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 1 is characterized in that: The following steps are involved: Step 1: Multimodal feature fusion representation; Step 2: high-dimensional semantic encoding; Step 3: Entity category decoding.
6. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 5 is characterized in that: The step 1 comprises the following steps: Step 1.1: Preprocessing of legal texts; Step 1.2, multimodal feature extraction; Step 1.3: Multimodal feature fusion.
7. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 6 is characterized in that: The legal text preprocessing includes text normalization, radical extraction, and introduction of word segmentation information.
8. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 1 is characterized in that: Includes entity-aware encoding modifiers.
9. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 8 is characterized in that: The entity-aware encoding modifier initializes the encoding of entity categories in a high-dimensional semantic space.
10. The legal document entity extraction method based on Chinese semantic enhancement as claimed in claim 8, characterized in that: The entity-aware encoding corrector adaptively learns the encoding of entity categories in a high-dimensional semantic space and increases the encoding differences between different entity categories.
