Semantic analysis-based case file named entity identification method and system

Abstract

The invention provides a case file named entity recognition method and system based on semantic analysis, and the method comprises the following steps: obtaining original text data of a judicial case electronic file, carrying out the data preprocessing and word level division of the original text data, and obtaining a lexical element sequence; processing the lexical element sequence by adopting a sliding window mechanism, extracting syntactic phrases and fragment feature information, and fusing to generate a candidate fragment feature set; entity types are defined, similarity calculation is carried out on the entity types and the candidate segment feature vectors, the entity type with the highest similarity score is obtained and associated with the candidate segments, and a flat entity set is generated; constructing a predefined structured template, and matching entities into slots of the predefined structured template by adopting a pre-trained classification model to obtain an initial structured instance set; according to the method, the key entities in the file can be efficiently identified, the structural relationship and event information between the entities can be automatically constructed, and the comprehensiveness, accuracy and processing efficiency of case semantic extraction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a named entity recognition method and system for case files based on semantic analysis. BACKGROUND

[0002] Intelligent analysis of electronic files of judicial cases is of great significance for improving judicial efficiency and assisting in case investigation. As a key pre-task of information extraction, named entity recognition aims to locate and classify entities with specific meanings from massive unstructured texts, such as suspects, places, times, agencies, charges, and tools. Further, understanding the roles played by these entities in cases and the structured relationships formed between them is crucial for in-depth case analysis.

[0003] A criminal hotspot feature mining method and system based on case element analysis, disclosed in CN109710712A, includes preprocessing case data using natural language processing, extracting case elements, and generating information about the corresponding commercial circle of the case. Data visualization is displayed in four modules: case heat map, case single point, commercial circle frame, and commercial circle labeling point. Through the case elements, the criminal characteristics of the commercial circle are analyzed, the similarity between each criminal commercial circle is calculated, and the similarity network structure of all criminal commercial circles is generated. The similarity network structure is analyzed, and the criminal characteristics of commercial circles with similar attributes are proposed.

[0004] Traditional sequence labeling-based methods or some generation NER methods that rely on autoregressive decoding have slow reasoning speed and limited throughput when processing long texts or requiring high concurrency. Most NER systems mainly focus on identifying independent and flattened entities, and have insufficient or inefficient extraction capabilities for complex relationships between entities, entity attributes, and structured events involved in cases. Moreover, case entity types are diverse and may evolve with new criminal methods. Traditional methods often require a large amount of re-labeled data and model adjustments to adapt to new entity types or fine-grained entity classification, resulting in poor adaptability. SUMMARY

[0005] Therefore, the present application provides a named entity recognition method and system for case files based on semantic analysis, which can efficiently identify key entities in files and automatically construct structured relationships between entities and event information, thereby improving the comprehensiveness, accuracy, and processing efficiency of case semantic extraction.

[0006] The technical solution of the present application is as follows: In a first aspect, the present application provides a named entity recognition method for case files based on semantic analysis, comprising the following steps: S1, obtain original text data of the electronic file of the judicial case, perform data preprocessing on the original text data and word-level division to obtain a word sequence; S2, adopt a sliding window mechanism to process the word sequence, extract syntax phrase and segment feature information, and fuse to generate a candidate segment feature set; S3, define an entity type, and perform similarity calculation on the candidate segment feature vector to obtain an entity type with the highest similarity score and the candidate segment, and associate the entity type with the candidate segment to generate a flat entity set; S4, construct a pre-defined structured template, and use a pre-trained classification model to match the entity to a slot of the pre-defined structured template to obtain an initial structured instance set; S5, take the flat entity as a node, identify semantic association between entities based on a pre-trained relation classification model, extract a subgraph structure matched with the structured template, map and fill corresponding entity and relation information, and obtain a final structured instance set; S6, use a pre-constructed criminal field knowledge graph and a legal regulation ontology library to verify and correct the extracted entity and structured instance and entity relation, integrate the corrected entity with the structured instance and entity relation, and output a final judicial case recognition result.

[0007] On the basis of the above technical scheme, preferably, in step S1, the original text data of the electronic file of the judicial case is obtained, the original text data is preprocessed and divided at the word level to obtain a word sequence, including the following sub-steps: S11, obtain the original text data of the electronic file of the judicial case, identify and separate different regions of the document; S12, perform paragraph segmentation on different regions of the document according to line breaks, and use a sentence segmenter to segment each paragraph into sentences to obtain a sentence sequence; S13, perform data preprocessing on the sentence sequence, and use a word segmenter to divide at the word level to obtain a word sequence.

[0008] On the basis of the above technical scheme, preferably, in step S2, the word sequence is processed using a sliding window mechanism to extract syntax phrase and segment feature information, and a candidate segment feature set is generated by fusion, including the following sub-steps: S21, preset a fixed length window and use a sliding window mechanism to enumerate continuous word segment to obtain a candidate word segment set; S22, perform block coding on the word sequence by using a pre-trained coding model, output the context representation vector of each word, and extract the context representation vector of the start and end word as boundary information; S23, extract the length of the word segment and generate a fixed-dimension width vector through an embedding layer; S24, judge whether the length of the token segment is greater than 1, if the length of the token segment is greater than 1, construct a feedforward neural network, calculate the corresponding attention score of each token in the token segment by using the feedforward neural network, perform softmax normalization processing on the attention score to obtain the corresponding attention weight, and calculate the weighted sum as the internal context vector according to the attention score of each token and the corresponding attention weight, if the length of the token segment is equal to 1, use the context representation vector of the token as the internal context vector; S25, splice and fuse the boundary information of the token segment, the fixed-dimension width vector and the internal context vector to generate a candidate segment feature set.

[0009] On the basis of the above technical scheme, preferably, in step S3, the entity type is defined, and similarity calculation is performed on the candidate segment feature vector to obtain the entity type with the highest similarity score and the candidate segment, and a flat entity set is generated, including the following sub-steps: S31, define the entity type in the field of judicial cases, and use a pre-trained type coding model to encode to obtain the embedding vector corresponding to each entity type; S32, calculate the cosine similarity between the embedding vector corresponding to each entity type and each candidate segment feature vector by using cosine similarity to obtain a similarity score; S33, obtain the number of labels corresponding to each entity type according to business requirements, if the number of labels is 1, associate the candidate segment corresponding to the highest similarity score with the entity type to generate a flat entity set, if the number of labels is multiple, preset a similarity threshold, obtain the entity types with a similarity score greater than the similarity threshold, and calculate the token-level overlap rate between the entity types with overlap, if the token-level overlap rate is greater than 0.5, retain the entity type with a higher similarity score, obtain multiple entity types, associate the multiple entity types with the corresponding candidate segment to generate a flat entity set.

[0010] On the basis of the above technical scheme, preferably, in step S4, a pre-defined structured template is constructed, a pre-trained classification model is used to match the entity to the slot of the pre-defined structured template to obtain an initial structured instance set, including the following sub-steps: S41, construct a pre-defined structured template by combining the domain knowledge of legal experts and statistical analysis of a large number of structured judgment documents; S42, encode the text paragraph by using a pre-trained encoding model, and splice and input the text paragraph and the entity type distribution feature into a classifier to obtain the probability that the text paragraph contains each pre-defined structured template; S43, preset a probability threshold, and activate the structured template with an output probability greater than the probability threshold; S44, according to the activated structured template, filtering the entities of the same type in the flat entity set as candidate entities to fill the slot; S45, vector splicing each candidate entity with the filled slot, and inputting into the pre-trained feedforward neural network model to output the filling probability, selecting the candidate entity with the highest filling probability to fill the slot to obtain the initial structured instance set.

[0011] On the basis of the above technical solutions, preferably, in step S5, the flat entity is taken as a node, the semantic association between entities is identified based on the pre-trained relationship classification model, the subgraph structure matching the structured template is extracted, and the corresponding entity and relationship information is mapped and filled to obtain the final structured instance set, including the following substeps: S51, taking the flat entity as a node in the graph and associating attribute information, the attribute information including entity type, text content and filling slot information; S52, traversing the flat entity set, combining any two entities to generate an entity pair, and using the pre-trained relationship classification model to predict the relationship between the entity pair as an edge connecting the corresponding nodes to obtain an initial case graph; S53, using a subgraph matching algorithm to find a subgraph structure matching the structured instance set on the initial case graph, and mapping the nodes and edges in the subgraph structure to the roles and relationships in the structured template, and filling the corresponding entity and relationship information to obtain the final structured instance set.

[0012] On the basis of the above technical solutions, preferably, in step S6, the extracted entities and structured instances and entity relationships are verified and corrected using a pre-constructed criminal field knowledge graph and a legal regulation ontology library, and the corrected entities and structured instances and entity relationships are integrated to output the final judicial case recognition result, including the following substeps: S61, selecting Claude4 as a teacher model, and setting an input template, the input template including task description, judicial case labeling and output format, the task description being to identify entities, relationships and structured instances from criminal case texts; S62, the teacher model identifies the unlabeled original text according to the input template, predicts to generate flat entities, structured instances and relationship pseudo-labels, and eliminates labels with a teacher model output probability lower than a screening threshold; S63, constructing an external knowledge base, the external knowledge base including an entity dictionary and a relationship ontology and rule base; According to the entity dictionary, checking the entity type, alias and attribute in the flat entity and the pseudo-label, marking the entities with type mismatch, alias error or attribute missing; According to the relationship ontology, the relationship between entities and the relationship type, direction and attribute in the pseudo label are checked, and the relationship with an incorrect type, direction or attribute is marked; According to the rule base, whether the structured instance and the pseudo label violate judicial logic constraints or laws and regulations is checked, and the instance that violates the constraints is marked. S64, according to the entity verification result, the relationship verification result and the structured instance verification result, prompting manual correction, integrating the corrected flat entity with the structured instance and the entity relationship, and outputting the identification result of the judicial case.

[0013] In a second aspect, the present application also provides a named entity recognition system for case files based on semantic analysis, which comprises the named entity recognition method for case files based on semantic analysis, and the system comprises: The collection processing module is configured to obtain original text data of the electronic case files of the judicial case, perform data preprocessing on the original text data, and divide the original text data at the word level to obtain a word sequence. The candidate segment extraction module is configured to process the word sequence by using a sliding window mechanism, extract syntactic phrases and segment feature information, and fuse to generate a candidate segment feature set. The entity extraction module is configured to define entity types, perform similarity calculation on the candidate segment feature vectors, associate the entity type with the highest similarity score with the candidate segment, and generate a flat entity set. The structure information extraction module is configured to construct a pre-defined structured template, match the entities to the slots of the pre-defined structured template by using a pre-trained classification model, and obtain an initial structured instance set. The relationship information extraction module is configured to take the flat entities as nodes, identify semantic associations between the entities based on a pre-trained relationship classification model, extract subgraph structures matched with the structured template, map and fill corresponding entity and relationship information, and obtain a final structured instance set. The correction output module is configured to verify and correct the extracted entities, structured instances and entity relationships by using a pre-constructed criminal domain knowledge graph and a law and regulation ontology library, integrate the corrected entities with the structured instances and the entity relationships, and output a final judicial case identification result.

[0014] In a third aspect, the present application also provides a terminal device, which comprises a memory, a processor, and a named entity recognition method program for case files based on semantic analysis stored in the memory and executable on the processor, and the named entity recognition method program for case files based on semantic analysis is configured to implement the steps of the named entity recognition method for case files based on semantic analysis.

[0015] In a fourth aspect, the present application further provides a storage medium, which is a computer storage medium, and the computer storage medium has stored thereon a program of the named entity recognition method for case files based on semantic analysis, and the program of the named entity recognition method for case files based on semantic analysis, when executed by a processor, implements the steps of the named entity recognition method for case files based on semantic analysis.

[0016] The named entity recognition method and system for case files based on semantic analysis of the present application have the following beneficial effects over the prior art: (1) By fusing the high-efficiency flat recognition mechanism and the structured collaborative modeling for named entity recognition, the key entities in the case files can be efficiently recognized, and the structured relationships and event information between the entities can be automatically constructed, thereby improving the comprehensiveness, accuracy and processing efficiency of the case semantic extraction; (2) By recognizing the semantic association between the entities through the relationship classification model, the understanding of the case information is deepened, which is conducive to discovering the complex relationships hidden in the text, and the subgraph matching algorithm is used to quickly find the subgraph structure matching the structured template, thereby improving the efficiency of information extraction; (3) By using the hierarchical or partially parallel decoding strategy, different levels of information can be processed in parallel, thereby improving the processing efficiency, and the hierarchical decoding strategy covers all levels of the case structure, thereby ensuring the comprehensive extraction of information and providing more abundant data support for case analysis; (4) By checking the entity dictionary, the relationship ontology and the rule base, the errors of entity type mismatch, unreasonable relationship and violation of judicial logic constraints are effectively identified and corrected, thereby improving the recognition accuracy of entities, relationships and structured instances. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 A flowchart of the named entity recognition method for case files based on semantic analysis of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0020] As shown in Figure 1 the first aspect, the present application provides a named entity recognition method for case files based on semantic analysis, comprising the following steps: S1, obtaining original text data of a judicial case electronic file, performing data preprocessing on the original text data and word-level division to obtain a word sequence.

[0021] The step S1 in the embodiment includes the following sub-steps: S11, obtaining original text data of a judicial case electronic file, identifying and separating different regions of the document; It should be noted that for the electronic file in PDF format, first, a PDF analysis library (pdfminer.six) is used to extract the original text stream; then, based on rules (such as fixed format of header and footer, font size and position of title) or a simple machine learning model (such as a classifier based on text block features), different regions (such as cover, table of contents, main text, evidence list, court record, etc.) of the document are identified and separated.

[0022] S12, performing paragraph division on different regions of the document according to line breaks, and using a sentence segmenter to perform sentence segmentation on each paragraph to obtain a sentence sequence; It should be noted that for the extracted main text part, first, preliminary paragraph division is performed according to line breaks, and a sentence segmenter is applied to each paragraph. A rule segmenter combining common Chinese punctuation marks (。!?;) and specific legal document terms (such as “hereby the judgment is as follows:”) can be used; or a pre-trained NLP toolkit (such as a Chinese model of spaCy or NLTK, and fine-tuning for legal text) is used for sentence segmentation.

[0023] S13, performing data preprocessing on the sentence sequence, and using a word segmenter to perform word-level division to obtain a word sequence.

[0024] It should be noted that the preprocessing includes OCR error correction, noise removal and format unification, wherein the OCR error correction can construct a replacement dictionary for common OCR errors, for more complex errors, a sequence-to-sequence correction model based on Transformer can be trained, or the text correction capability of a model such as ERNIE-Gram is used; noise removal uses regular expressions to remove page numbers, fixed content in page headers, watermark text, and non-text special control characters; format unification converts full-width characters to half-width characters, and unifies the representation format of numbers and dates; a WordPiece tokenizer that matches the pre-trained Transformer model is selected to cut the processed sentence sequence into token sequences.

[0025] In this embodiment, through OCR error correction, noise removal and format unification, the quality of the original text data is effectively improved, a more reliable data foundation is provided for subsequent named entity recognition and structured information extraction, and the processing efficiency and model adaptability are improved.

[0026] S2, using a sliding window mechanism to process the token sequence, extracting syntactic phrase and fragment feature information, and fusing to generate a candidate fragment feature set.

[0027] The step S2 in this embodiment includes the following sub-steps: S21, presetting a fixed length window and using a sliding window mechanism to enumerate continuous token fragments to obtain a candidate token fragment set; It should be noted that by setting the minimum entity length L min =1 and the maximum entity length L max =10, or according to the statistical analysis of the entity length in the judicial field, all continuous token fragments within this length range are enumerated; and a lightweight Chinese dependency syntax analyzer (such as a fast dependency analysis module based on LTP or HanLP) is used to extract all noun phrases, verb phrases and parallel noun phrases in the text; these syntactic components are usually good candidates for constructing named entities; and the length limit can be applied to the extracted syntactic phrases.

[0028] S22, block encoding the token sequence through a pre-trained encoding model, outputting the context representation vector of each token, and extracting the context representation vector of the start and end tokens as boundary information.

[0029] It should be noted that the Transformer encoder model pre-trained on large-scale Chinese legal texts is selected; for texts exceeding the maximum input length of the model, a sliding window mechanism is used to set the window size and step to block the text; for overlapping word units, the context representation vector can take the average of multiple encoding results; the encoder outputs the context representation vector of each word unit, and extracts the context representation vector of the start and end word units as boundary information.

[0030] S23, extract the length of the word unit segment and generate a fixed dimension width vector through the embedding layer to represent the length information of the segment.

[0031] S24, determine whether the length of the word unit segment is greater than 1, if the length of the word unit segment is greater than 1, construct a feedforward neural network, use the feedforward neural network to calculate the corresponding attention score of each word unit in the word unit segment, perform softmax normalization processing on the attention score to obtain the corresponding attention weight, and calculate the weighted sum as the internal context vector according to the attention score of each word unit and the corresponding attention weight, if the length of the word unit segment is equal to 1, use the context representation vector of the word unit as the internal context vector; The expression of the feedforward neural network is:

[0032] In the formula, h k is the context representation vector of the word unit, v , W att , b att is a learnable parameter; The expression of the softmax normalization processing is:

[0033] In the formula, a k is the attention score of the word unit; The expression of the internal context vector calculation is:

[0034] In the formula, h internal is the internal context vector.

[0035] S25, splice and fuse the boundary information of the word unit segment, the fixed dimension width vector and the internal context vector to generate a candidate segment feature set; the expression is: .

[0036] In the formula, the boundary information of the token segment, the fixed-dimension width vector, and the internal context vector are spliced, and two layers of a feedforward network are used FFN span dimension reduction is performed, and the FFN span The output dimension of the output is consistent with the dimension of the type embedding, and a candidate segment feature set is generated.

[0037] In this embodiment, continuous token segments are enumerated through a sliding window mechanism, and a candidate token segment set can be efficiently generated, unnecessary calculations are reduced, the length limitation of model input is avoided by using a block coding mode for long texts, and the processing efficiency is improved. Furthermore, the context representation vector of the starting and ending token is extracted as boundary information, and a fixed-dimension width vector is generated, which can more comprehensively represent the features of the token segment. The boundary information, the width vector, and the internal context vector are spliced and fused through a feedforward neural network, and the internal context vector is calculated through an attention mechanism, which can capture the semantic relationship within the token segment, thereby improving the model performance and adaptability.

[0038] In step S3, the entity types are defined, and similarity calculation is performed on the candidate segment feature vectors to obtain the entity type with the highest similarity score and the candidate segment, and a flat entity set is generated.

[0039] In step S3 in this embodiment, the following sub-steps are included: In step S31, the entity types in the judicial case field are defined, and a pre-trained type encoding model is used for encoding to obtain the embedding vectors corresponding to each entity type.

[0040] It should be noted that an entity type list and its natural language description are maintained for the judicial field, such as person (PER), location (LOC), time (TIME), and money (MONEY). In addition, a lightweight pre-trained type encoding model (such as the Chinese version of DistilBERT or a variant of Sentence-BERT) is used as a type encoder; the natural language description of each type is input into the encoder, and the [CLS] vector or average pooling vector output by the encoder is taken as the embedding representation of the type.

[0041] In step S32, the cosine similarity between the embedding vectors of each entity type and each candidate segment feature vector is calculated using the cosine similarity, and the similarity score is obtained, which is expressed as:

[0042] In the formula, Score(span(i,j),type_label k ) The similarity score is Rep span(i,j) The candidate segment feature vector isRep type_labelk embedding vectors of entity types; S33, according to the business requirement, the number of labels corresponding to each entity type is obtained, if the number of labels is 1, the candidate segment corresponding to the highest similarity score is associated with the entity type to generate a flat entity set; if the number of labels is multiple, a similarity threshold is preset, the entity types corresponding to the similarity score greater than the similarity threshold are obtained, and the word-level overlap rate between the entity types with overlap is calculated, if the word-level overlap rate is greater than 0.5, the entity type with high similarity score is retained, multiple entity types are obtained, and the multiple entity types are associated with the corresponding candidate segment to generate a flat entity set.

[0043] In this embodiment, the candidate segment feature vector and the entity type embedding vector both contain rich semantic information, the candidate segment can be accurately associated with the entity type through similarity calculation, the low-quality matching result can be filtered out by presetting the similarity threshold, the accuracy of entity recognition is improved, according to the business requirement, the single label and the multi-label situation can be flexibly handled, the entity recognition requirement in different scenes is met, for the entity types with overlap, the word-level overlap rate is calculated and the entity type with high similarity score is retained, the repeated recognition and the wrong association can be avoided, and the flat entity set is efficiently generated, which provides accurate and reliable entity recognition result for subsequent analysis of the electronic case file of the judicial case.

[0044] S4, a pre-defined structured template is constructed, a pre-trained classification model is used to match the entity to the slot of the pre-defined structured template, and an initial structured instance set is obtained.

[0045] In this embodiment, step S4 includes the following sub-steps: S41, combining the domain knowledge of legal experts and the statistical analysis of a large number of structured judgment documents, a pre-defined structured template is constructed.

[0046] It should be noted that JSONSchema or self-defined YAML format is used to store and manage these patterns; for example, the structured template definition of the injury event is as follows: schema_name: IntentionalInjuryEvent description: describes the core elements of the injury event slots: -name: perpetrator type_constraint: PER_Suspect# points to the flat entity type is_list: true description:perpetrator -name:victim type_constraint:PER_Victim is_list:true description:victim -name:injury_location_body type_constraint:BODY_PART # e.g. "head", "chest", etc. fine-grained entity is_list:true description:injury location -name:injury_degree type_constraint:INJURY_LEVEL # e.g. "minor injury level 1", "serious injury level 2" is_list:false description:injury degree -name:instrument_used type_constraint:EVI_Weapon is_list:true description:instrument used -name:event_time type_constraint:TIME_Exact is_list:false description:event time -name:event_location type_constraint:LOC_CrimeScene is_list:false description:event location

[0047] S42, encode the text passage using a pre-trained encoding model (e.g. a Transformer encoder) and concatenate the input with the entity type distribution features to the classifier to obtain the probability of the text passage containing each pre-defined structured template.

[0048] S43, set a probability threshold, obtain and activate the structured template whose output probability is greater than the probability threshold; S44, according to the activated structured template, screen the entities of the same type in the flat entity set as candidate entities to fill the slot; S45, vector splice each candidate entity with the filled slot, and input into the pre-trained feedforward neural network model, output the filling probability, select the candidate entity with the highest filling probability to fill the slot, and obtain the initial structured instance set.

[0049] It should be noted that the pre-defined structured template reduces the workload of manual definition and information extraction, improves the efficiency of information extraction, uses the pre-trained classification model to encode and classify the text paragraph, improves the accuracy of the activation of the structured template, and through vector splicing and feedforward neural network model to predict the filling probability, select the candidate entity with the highest score to fill the slot, improve the accuracy of slot filling, the structured template can be extended and modified according to different case types and requirements, has good flexibility, and can process multiple pre-defined structured templates at the same time, suitable for complex case analysis scene, efficiently generates the initial structured instance set, provides standardized and accurate information support for subsequent judicial case electronic file analysis.

[0050] S5, taking the flat entity as a node, identifying the semantic association between entities based on the pre-trained relationship classification model, and extracting the subgraph structure matched with the structured template, mapping and filling the corresponding entity and relationship information, and obtaining the final structured instance set.

[0051] The step S5 in the embodiment includes the following sub-steps: S51, taking the flat entity as a node in the graph, and associating attribute information, the attribute information including entity type, text content and filled slot information; It should be noted that all the identified flat entities are taken as initial nodes in the graph, and each node includes attribute information such as entity type, text content and filled slot information.

[0052] S52, traverse the flat entity set, combine any two entities to generate an entity pair, and use the pre-trained relationship classification model to predict the relationship between the entity pair, connect the corresponding nodes with edges, and obtain the initial case graph; It should be noted that the pre-trained relationship classification model is used to predict the relationship between the entity pair. The model can receive the entity pair and their context information as input, predict whether there is a pre-defined relationship type between them by learning the semantic association between entities, and connect the entity pairs with edges according to the prediction result of the relationship classification model, forming the initial case graph.

[0053] S53, adopt subgraph matching algorithm to find the subgraph structure matching the structured instance set on the initial case graph, and map the nodes and edges in the subgraph structure to the roles and relationships in the structured template, and fill in the corresponding entity and relationship information to obtain the final structured instance set.

[0054] It should be noted that the subgraph matching algorithm is used to find the subgraph structure matching the predefined structured template on the initial case graph. The subgraph matching algorithm compares the subgraph in the graph with the slots and relationships of the structured template to find the required subgraph. The nodes and edges in the subgraph structure are mapped to the roles and relationships in the structured template, and the corresponding entity and relationship information is filled in. The final structured instance set is obtained, which contains the key information required for case analysis.

[0055] This embodiment deepens the understanding of case information by identifying the semantic association between entities through the relationship classification model, which is beneficial to discovering complex relationships hidden in the text. The subgraph matching algorithm is used to quickly find the subgraph structure matching the structured template, which improves the efficiency of information extraction. By representing the case information as a structured instance set, it provides in-depth and accurate information support for subsequent judicial case electronic file analysis.

[0056] For each identified flat entity, a document-unique ID is assigned, which is unique within the document and used to identify and reference the entity. When outputting structured information, for slots that need to reference these entities, their IDs are used directly instead of repeating entity text, which can avoid repeating the same entity text in the output and reduce redundant information.

[0057] For information that needs to be standardized, such as numerical values and dates, standardized processing is performed to improve data consistency and comparability, facilitating subsequent machine processing and storage. For slots that need to extract or generate text descriptions from the original text, a lightweight controlled text generation model can be used, such as a Pointer-Generator Network or a small-scale pre-trained Seq2Seq model. This model generates short and accurate text based on slot descriptions and related context, with a strict limit on output length of up to 50 tokens to ensure efficient tokenization and reduce computational resource consumption.

[0058] The final output uses standardized JSON or Protobuf and other compact serialization formats, which facilitates subsequent machine processing and storage, improves data processing efficiency and accuracy.

[0059] For example, an event instance can be represented as an object containing schema type, instance ID, and slot filling values, such as: { "schema_id":"FraudEvent_v1.2", "instance_id":"uuid_xyz", "slots":{ "perpetrator":[ { "entity_id":"flat_entity_id_001", "text":"Zhang San" } ], "victim":[ { "entity_id":"flat_entity_id_002", "text":"Li Si" } ], "means and process_text":"xxxxxx", "involving money":{ "entity_id":"flat_entity_id_003", "value":50000, "currency":"RMB" }... } }。

[0060] This representation avoids repeating the output of entity text by referencing the ID of the identified entity, and only outputs the text segment for slots that require the original text.

[0061] In addition, for complex case structures with inherent hierarchy or dependency, a hierarchical or partially parallel decoding strategy is adopted, specifically: According to the complexity of judicial cases, the structured mode is divided into different levels; Level-0 is the core event layer, Level-1 is the associated information layer, such as the "personal background information" template related to the participants of the core event or the evidence template related to the core event; Level-2 is the derived information layer, which defines the organizational structure template and the social impact template.

[0062] Firstly, the model centrally processes the identification and filling of the core pattern instances of Level-0; once the instances of Level-0 are identified, the analysis of Level-1 patterns is triggered; the filling of the patterns related to the core event participants and the filling of the “evidence” templates related to the core event can be started in parallel; the Level-1 analysis of different branches can be independently carried out; if the output of Level-1 triggers the patterns of Level-2, the analysis continues downward; the results of the upper structure analysis can be used as the input or constraint condition of the lower structure analysis to ensure the consistency and relevance of the information.

[0063] In this embodiment, through the hierarchical or partially parallel decoding strategy, different levels of information can be processed in parallel, the processing efficiency is improved, and at the same time, the hierarchical decoding strategy covers all levels of the case structure, ensuring the comprehensive extraction of information and providing richer data support for case analysis.

[0064] S6, the pre-constructed criminal field knowledge graph and the legal regulation ontology library are used to verify and correct the extracted entities and structured instances and entity relationships, and the corrected entities and structured instances and entity relationships are integrated, and the final judicial case recognition result is output.

[0065] The step S6 in this embodiment includes the following sub-steps: S61, selecting Claude4 as a teacher model, and setting an input template, the input template including a task description, a judicial case annotation, and an output format, the task description being to identify entities, relationships, and structured instances from criminal case texts; It should be noted that a large open source model such as Claude4Opus or a large open source model fine-tuned in the legal field is selected as the teacher model, the model has strong ability in processing complex texts and generating structured outputs, can provide high-quality pseudo-labels for the student model, and the input template Prompt is designed to contain the task description, a small amount of high-quality criminal case annotation examples, and the output format requirement. The teacher model is required to output flat entities and structured relationships in a specific JSON format to guide the student model to learn the correct annotation method.

[0066] S62, the teacher model identifies the original text without labels according to the input template, predicts and generates pseudo-labels of flat entities, structured instances, and relationships, and eliminates labels with an output probability of the teacher model lower than a filtering threshold; It should be noted that the annotation output of the teacher model on the unlabeled criminal text is collected as the pseudo-label, a confidence evaluation mechanism is designed, such as the probability based on the output of the teacher model or the consistency with the results of multiple different teacher models, to filter out low-quality pseudo-labels. For structured information, simple logical consistency checks can be performed to ensure the quality of the pseudo-labels. A distillation loss for flat entity recognition is also set, including logits matching of segment classification and probability distribution matching of type selection, minimizing the KL divergence between logits output by the student model and the teacher model, and the logits of the teacher model are usually smoothed by a temperature coefficient T>1 to improve the distillation effect; A distillation loss for structured information extraction is also included. If the teacher model can generate structured pseudo-labels, these structures can be directly imitated, such as distilling the logits of slot filling decisions. If the teacher model mainly provides text form of case summaries, the student model can be trained to recover the structure from the summary, or the structure generated by the student model is semantically close to the teacher's summary, which is achieved through an auxiliary semantic similarity loss; The distillation losses for flat entity recognition and structured information extraction are combined to form a comprehensive loss function, expressed as:

[0067] In the formula, α and β are adjustable weights, which can be determined by grid search on the performance of the validation set, and the common value range is 0.1 to 0.9; L total is the total loss function, used to measure the overall loss of the model during the training process, considering the loss of the real label and the distillation loss from different stages; L ground_truth is the loss function of the real label, used to measure the difference between the model's prediction result and the actual label, usually using a cross-entropy loss function; L distillation_S2 is the distillation loss for flat entity recognition, used in the knowledge distillation process to measure the difference between the outputs of the student model and the teacher model, L distillation_S3 is the distillation loss for structured information extraction.

[0068] S63, an external knowledge base is constructed, including an entity dictionary, a relationship ontology, and a rule base; According to the entity dictionary, check the entity type, alias, and attribute in the flat entity and pseudo-label, and mark the entities with type mismatch, alias error, or attribute missing; According to the relationship ontology, check the relationship between entities and the relationship type, direction, and attribute in the pseudo-label, and mark the relationships with type disallowed, direction error, or unreasonable attribute; According to the rule base, check whether the structured instance and the pseudo-label violate judicial logic constraints or laws and regulations, and mark the instances that violate the constraints; It should be noted that the entity dictionary includes common roles of involved persons, types of institutions, types of locations, legal and regulatory entries, charges, and tools used in the crime, which are used to verify whether the extracted entities conform to the domain specifications; the relation ontology defines typical relationships between entities and marks the attributes of the relationships, which are used to verify whether the relationships between entities are reasonable; the rule base consists of constraint rules defined based on legal logic and practical experience, which are used to verify whether structured information violates judicial logic constraints or laws and regulations.

[0069] S64 prompts manual correction based on entity verification results, relationship verification results, and structured instance verification results. It then integrates the corrected flat entities with structured instances and entity relationships to output the identification results of judicial cases.

[0070] It should be noted that a lightweight pronoun dereference module is built based on rules or statistics. Rules are used to handle common pronoun references and references frequently used in legal documents. For more complex references, a binary classification model based on entity pair features can be trained to determine whether a referential relationship exists. Different representations of the same real-world entity identified in the document are linked to the same unique entity ID. This is achieved by calculating string similarity, contextual embedding similarity, and utilizing an external alias knowledge base. The final result is output in JSONLines format, where each line is an independent JSON object representing an extracted flat entity or a structured event / relationship instance. Each object contains detailed metadata and core extracted content.

[0071] For example: Output of a flat entity: { "doc_id":"XYZ001", "span_text":"Zhang San", "start_char":100, "end_char":102, "entity_type":"PER_Suspect", "confidence":0.95, "source_module":"S2_FlatNER" }; A structured event output: { "doc_id":"XYZ001", "event_id":"EVT001", "event_type":"RobberyEvent", "confidence":0.88, "source_module":"S3_StructExtract", "slots":{ "perpetrator":[ "entity_id_zhangsan" ], "victim":[ "entity_id_lisi" ],... } }。

[0072] In this embodiment, the extracted information is verified by integrating entity dictionary, relationship ontology and rule base, which not only effectively identifies and corrects the errors of entity type mismatch, unreasonable relationship and violation of judicial logic constraints, significantly improves the identification accuracy of entity, relationship and structured instance, but also deeply integrates the knowledge graph of judicial field and the ontology library of laws and regulations into the identification process, so that the model can more accurately understand and process the domain-specific entity, relationship and structured information, significantly enhances the domain adaptability of the model. At the same time, the constraint rules defined based on legal logic and practical experience ensure that the extracted information strictly meets the requirements of judicial practice, thereby greatly improving the professionalism and reliability of the identification results.

[0073] In a second aspect, the present application also provides a named entity recognition system for case files based on semantic analysis, which comprises the named entity recognition method for case files based on semantic analysis, and the system comprises: a collection processing module for obtaining original text data of a judicial case electronic file, performing data preprocessing on the original text data and word-level division to obtain a word sequence; a candidate segment extraction module for processing the word sequence by using a sliding window mechanism, extracting syntactic phrases and segment feature information, and fusing to generate a candidate segment feature set; an entity extraction module for defining entity types, calculating the similarity of candidate segment feature vectors, associating the entity type with the highest similarity score with the candidate segment, and generating a flat entity set; a structure information extraction module for constructing a pre-defined structured template, matching entities to the slot positions of the pre-defined structured template by using a pre-trained classification model, and obtaining an initial structured instance set; a relationship information extraction module for taking flat entities as nodes, identifying semantic associations between entities based on a pre-trained relationship classification model, extracting subgraph structures matching the structured template, mapping and filling corresponding entity and relationship information, and obtaining a final structured instance set; The correction output module is used for correcting and modifying the extracted entities and the structured instances and entity relationships by using the pre-constructed criminal field knowledge graph and the legal regulation ontology library, and integrating the modified entities and the structured instances and entity relationships, and outputting a final judicial case recognition result.

[0074] It should be noted that the system corresponds to the above-mentioned mechanical leg motion control method based on water power compensation. All implementation manners in the above-mentioned method embodiments are applicable to the embodiments of the system and can achieve the same technical effects.

[0075] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0076] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system and modules can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0077] In the embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented by other ways. For example, the above-described device embodiments are only schematic, and the division of the units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0078] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0079] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0080] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, and various program code storage media.

[0081] In addition, it should be noted that in the system and method of the present application, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present application. Moreover, the steps of performing the above series of processes can naturally be executed in time sequence according to the order of description, but do not necessarily have to be executed in time sequence, and some steps can be executed in parallel or independently of each other. It can be understood by those skilled in the art that all or any steps or components of the method and device of the present application can be implemented in hardware, firmware, software or a combination thereof in any computing device (including processors, storage media, etc.) or network of computing devices, which can be implemented by those skilled in the art using their basic programming skills after reading the description of the present application.

[0082] Therefore, the object of the present application can also be achieved by running a program or a set of programs on any computing system. The computing system can be a commonly known general system. Therefore, the object of the present application can also be achieved only by providing a program product containing program code for implementing the method or device. That is, such a program product also constitutes the present application, and a storage medium storing such a program product also constitutes the present application. Obviously, the storage medium can be any commonly known storage medium or any storage medium developed in the future. It should be noted that in the device and method of the present application, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present application. Moreover, the steps of performing the above series of processes can naturally be executed in time sequence according to the order of description, but do not necessarily have to be executed in time sequence. Some steps can be executed in parallel or independently of each other.

[0083] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for named entity recognition of a case file based on semantic analysis, characterized in that, The method comprises the following steps: S1, obtaining original text data of a judicial case electronic file, performing data preprocessing on the original text data and word-level division to obtain a word sequence; S2, processing the word sequence by using a sliding window mechanism, extracting syntax phrase and segment feature information, and fusing to generate a candidate segment feature set; S3, defining an entity type, and performing similarity calculation on the candidate segment feature vector to obtain an entity type with the highest similarity score and the candidate segment, and associating the entity type with the candidate segment to generate a flat entity set; S4, constructing a pre-defined structured template, and using a pre-trained classification model to match the entity to a slot of the pre-defined structured template to obtain an initial structured instance set; S5, taking the flat entity as a node, identifying semantic association between entities based on a pre-trained relationship classification model, extracting a subgraph structure matched with the structured template, mapping and filling corresponding entity and relationship information, and obtaining a final structured instance set; S6, using a pre-constructed criminal field knowledge graph and a legal regulation ontology library to verify and correct the extracted entity and structured instance and entity relationship, integrating the corrected entity with the structured instance and entity relationship, and outputting a final judicial case recognition result.

2. The method for named entity recognition of a case file based on semantic analysis as claimed in claim 1, wherein: In step S1, the original text data of the judicial case electronic file is obtained, the original text data is preprocessed and divided at the word level to obtain a word sequence, which comprises the following sub-steps: S11, obtaining original text data of a judicial case electronic file, identifying and separating different regions of the document; S12, performing paragraph segmentation on different regions of the document according to line breaks, and using a sentence segmenter to segment each paragraph into sentences to obtain a sentence sequence; S13, performing data preprocessing on the sentence sequence, and using a word segmenter to divide at the word level to obtain a word sequence.

3. The method of claim 2, wherein: In step S2, the word sequence is processed by using a sliding window mechanism, syntax phrase and segment feature information are extracted, and a candidate segment feature set is fused and generated, which comprises the following sub-steps: S21, presetting a fixed length window and enumerating continuous word segment sets by using a sliding window mechanism to obtain a candidate word segment set; S22, performing block coding on the word sequence by using a pre-trained coding model, outputting a context representation vector of each word, and extracting the context representation vector of the starting and ending word as boundary information; S23, extracting the length of the word segment and generating a fixed-dimension width vector through an embedding layer; S24, determining whether the length of the word segment is greater than 1, if the length of the word segment is greater than 1, a feedforward neural network is constructed, the attention score of each word in the word segment is calculated by using the feedforward neural network, the attention score is subjected to softmax normalization processing to obtain the corresponding attention weight, the weighted sum is calculated according to the attention score of each word and the corresponding attention weight as the internal context vector, if the length of the word segment is equal to 1, the context representation vector of the word is used as the internal context vector; S25, the boundary information of the word segment, the fixed-dimension width vector and the internal context vector are spliced and fused to generate a candidate segment feature set.

4. The method for named entity recognition of a case file based on semantic analysis as claimed in claim 3, wherein: In step S3, the entity type is defined, and similarity calculation is performed with the candidate segment feature vector to obtain the entity type with the highest similarity score and associate the candidate segment to generate a flat entity set, including the following sub-steps: S31, define the entity type in the judicial case field, and use the pre-trained type coding model for coding to obtain the embedding vector corresponding to each entity type; S32, calculate the cosine similarity between the embedding vector corresponding to each entity type and the feature vector of each candidate segment using cosine similarity to obtain the similarity score; S33, obtain the number of labels corresponding to each entity type according to the business requirements, if the number of labels is 1, associate the candidate segment corresponding to the highest similarity score with the entity type to generate a flat entity set; if the number of labels is multiple, a preset similarity threshold is obtained, the entity types with a similarity score greater than the similarity threshold are obtained, and the word-level overlap rate between the entity types with overlapping is calculated, if the word-level overlap rate is greater than 0.5, the entity type with a higher similarity score is retained, multiple entity types are obtained, and the multiple entity types are associated with the corresponding candidate segment to generate a flat entity set.

5. The method for named entity recognition of a case file based on semantic analysis as claimed in claim 4, wherein: In step S4, the pre-defined structured template is constructed, the entity is matched to the slot of the pre-defined structured template using a pre-trained classification model, and an initial structured instance set is obtained, including the following sub-steps: S41, construct a pre-defined structured template in combination with the domain knowledge of legal experts and statistical analysis of a large number of structured judgment documents; S42, use a pre-trained coding model to encode the text paragraph, and combine the entity type distribution features to splice the input into a classifier to obtain the probability that the text paragraph contains each pre-defined structured template; S43, a probability threshold is preset, and the structured template corresponding to the output probability greater than the probability threshold is obtained and activated; S44, according to the activated structured template, filter the entities of the same type in the flat entity set as candidate entities to fill the slot; S45, splice each candidate entity and the filled slot to obtain the initial structured instance set.

6. The method for named entity recognition of a case file based on semantic analysis as claimed in claim 5, wherein: In step S5, the flat entity is taken as a node, the semantic association between entities is identified based on a pre-trained relationship classification model, and the subgraph structure matching the structured template is extracted, and the corresponding entity and relationship information is mapped and filled to obtain a final structured instance set, including the following sub-steps: S51, take the flat entity as a node in the graph, and associate attribute information, the attribute information includes entity type, text content and filled slot information; S52, traverse the flat entity set, combine any two entities to generate an entity pair, and use a pre-trained relationship classification model to predict the relationship between the entity pair as an edge to connect the corresponding nodes to obtain an initial case graph; S53, a subgraph matching algorithm is used to find a subgraph structure matching the structured instance set on the initial case graph, and the nodes and edges in the subgraph structure are mapped to the roles and relationships in the structured template, and the corresponding entity and relationship information is filled in to obtain the final structured instance set.

7. The named entity recognition method for case files based on semantic analysis as described in claim 6, characterized in that, In step S6, the extracted entities and structured instances and entity relationships are verified and corrected using a pre-constructed criminal field knowledge graph and a legal regulation ontology library, and the corrected entities and structured instances and entity relationships are integrated to output the final judicial case recognition result, including the following substeps: S61, Claude4 is selected as the teacher model, and an input template is set, including task description, judicial case labeling, and output format, wherein the task description is to identify entities, relationships, and structured instances from criminal case texts; S62, the teacher model identifies the unlabeled original text according to the input template, predicts the generation of flat entities, structured instances, and relationship pseudo-labels, and eliminates labels with a teacher model output probability lower than a screening threshold; S63, an external knowledge base is constructed, including an entity dictionary and a relationship ontology and rule base; According to the entity dictionary, the entity types, aliases, and attributes in the flat entities and pseudo-labels are checked, and entities with type mismatch, alias errors, or attribute omissions are marked; According to the relationship ontology, the relationship types, directions, and attributes between entities and in the pseudo-labels are checked, and relationships with type disallowed, direction errors, or unreasonable attributes are marked; According to the rule base, it is checked whether the structured instances and pseudo-labels violate judicial logic constraints or laws and regulations, and the instances that violate the constraints are marked; S64, according to the entity verification result, the relationship verification result, and the structured instance verification result, the artificial correction is prompted, the corrected flat entities and structured instances and entity relationships are integrated, and the recognition result of the judicial case is output.

8. A semantic analysis based named entity recognition system for case files, comprising the semantic analysis based named entity recognition method according to any one of claims 1-7, characterized in that: The system comprises: A collection and processing module is configured to obtain original text data of a judicial case electronic file, perform data preprocessing on the original text data, and divide the original text data at a word level to obtain a word sequence; A candidate segment extraction module is configured to process the word sequence using a sliding window mechanism, extract syntax phrases and segment feature information, and fuse to generate a candidate segment feature set; An entity extraction module is configured to define entity types, calculate similarity with a candidate segment feature vector, associate the entity type with the highest similarity score with a candidate segment, and generate a flat entity set; A structure information extraction module is configured to construct a pre-defined structured template, match entities to slots of the pre-defined structured template using a pre-trained classification model, and obtain an initial structured instance set; A relationship information extraction module is configured to take flat entities as nodes, identify semantic associations between entities based on a pre-trained relationship classification model, extract a subgraph structure matching the structured template, map and fill in corresponding entity and relationship information, and obtain a final structured instance set; The correction output module is configured to correct and modify the extracted entities, structured instances and entity relationships by using the pre-constructed criminal field knowledge graph and the legal regulation ontology library, integrate the modified entities, structured instances and entity relationships, and output a final judicial case identification result.

9. A terminal device, comprising: The terminal device comprises a memory, a processor, and a program of the named entity recognition method of the case file based on semantic analysis stored on the memory and capable of running on the processor, and the program of the named entity recognition method of the case file based on semantic analysis is configured to implement the steps of the named entity recognition method of the case file based on semantic analysis according to any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium is a computer storage medium, and the program of the named entity recognition method of the case file based on semantic analysis is stored on the computer storage medium and is capable of being executed by the processor to implement the steps of the named entity recognition method of the case file based on semantic analysis according to any one of claims 1 to 7.

Images