A Named Entity Recognition Method and System for Case Files Based on Semantic Analysis

Abstract

This invention proposes a named entity recognition method and system for case files based on semantic analysis. The method includes the following steps: acquiring the original text data of electronic case files of judicial cases; preprocessing the original text data and dividing it into words to obtain a word sequence; processing the word sequence using a sliding window mechanism to extract syntactic phrases and fragment feature information, and fusing them to generate a candidate fragment feature set; defining entity types and calculating similarity with the candidate fragment feature vectors, obtaining the entity type with the highest similarity score and associating it with the candidate fragment to generate a flat entity set; constructing a predefined structured template, and using a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances. This method can achieve efficient identification of key entities in case files and automatically construct structured relationships and event information between entities, improving the comprehensiveness, accuracy, and processing efficiency of case semantic extraction.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for named entity recognition of case files based on semantic analysis. Background Technology

[0002] Intelligent analysis of electronic case files in judicial cases is of great significance for improving judicial efficiency and assisting in case investigation. Named entity recognition, as a key preliminary task for information extraction, aims to locate and classify entities with specific meanings from massive amounts of unstructured text, such as persons involved in the case, locations, times, institutions, charges, and tools used in the crime. Furthermore, understanding the roles these entities play in the case and the structured relationships formed between them is crucial for in-depth case analysis.

[0003] A method and system for mining crime hotspot features based on case element analysis, disclosed in CN109710712A, includes: preprocessing case data using natural language processing to extract case elements and generate information on the business district to which the case belongs; visualizing the data using four modules: "case heat map," "case single point," "business district frame," and "business district marker point"; performing crime feature profiling analysis on the crime business districts using the case elements, calculating the feature similarity between each crime business district and generating a similarity network structure for all crime business districts, analyzing the similarity network structure, and proposing crime features of business districts with similar attributes.

[0004] Traditional sequence labeling-based methods or some generative NER methods that rely on autoregressive decoding suffer from slow inference speed and limited throughput when processing long texts or scenarios requiring high concurrency. Furthermore, most NER systems primarily focus on identifying independent, flat entities, and are insufficient or inefficient in extracting complex relationships between entities, entity attributes, and structured events involved in cases. Moreover, case entity types are diverse and may evolve with the emergence of new methods of crime. When adapting to new entity types or fine-grained entity classification, traditional methods often require a large amount of re-labeled data and model adjustments, resulting in poor adaptability. Summary of the Invention

[0005] In view of this, the present invention proposes a named entity recognition method and system for case files based on semantic analysis, which can not only achieve efficient recognition of key entities in the case files, but also automatically construct structured relationships and event information between entities, thereby improving the comprehensiveness, accuracy and processing efficiency of case semantic extraction.

[0006] The technical solution of this invention is implemented as follows: In a first aspect, this invention provides a named entity recognition method for case files based on semantic analysis, comprising the following steps:

[0007] S1. Obtain the original text data of the electronic case file of the judicial case, perform data preprocessing and word-level segmentation on the original text data to obtain the word sequence;

[0008] S2 uses a sliding window mechanism to process the word sequence, extract syntactic phrase and fragment feature information, and fuses them to generate a candidate fragment feature set;

[0009] S3, define entity types and calculate similarity with candidate fragment feature vectors, obtain the entity type with the highest similarity score and associate it with the candidate fragment to generate a flat entity set;

[0010] S4. Construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances.

[0011] S5 uses flat entities as nodes, identifies semantic associations between entities based on a pre-trained relation classification model, extracts subgraph structures that match the structured template, maps and fills in the corresponding entity and relation information, and obtains the final set of structured instances.

[0012] S6 utilizes a pre-built criminal domain knowledge graph and legal ontology to verify and correct the extracted entities, structured instances, and entity relationships. It then integrates the corrected entities, structured instances, and entity relationships to output the final judicial case identification results.

[0013] Based on the above technical solutions, preferably, step S1, which involves obtaining the original text data of the electronic case file of a judicial case, performing data preprocessing and word-level segmentation on the original text data to obtain a word sequence, includes the following sub-steps:

[0014] S11, Obtain the original text data of electronic case files for judicial cases, identify and separate different areas of the document;

[0015] S12, the document is divided into paragraphs according to line breaks, and each paragraph is divided into sentences using a sentence splitter to obtain a sentence sequence;

[0016] S13. Perform data preprocessing on the sentence sequence and use a word segmenter to perform word-level segmentation to obtain a word sequence.

[0017] Based on the above technical solutions, preferably, step S2, which uses a sliding window mechanism to process the word sequence, extract syntactic phrase and fragment feature information, and fuses them to generate a candidate fragment feature set, includes the following sub-steps:

[0018] S21, a fixed-length window is preset and a sliding window mechanism is used to enumerate continuous word segments to obtain a set of candidate word segments;

[0019] S22, the word sequence is block-encoded using a pre-trained encoding model, outputting the context representation vector of each word, and extracting the context representation vectors of the start and end words as boundary information;

[0020] S23, extract the length of the word segment and generate a fixed-dimensional width vector through the embedding layer;

[0021] S24. Determine if the word segment length is greater than 1. If the word segment length is greater than 1, construct a feedforward neural network and use the feedforward neural network to calculate the corresponding attention score for each word in the word segment. Perform softmax normalization on the attention score to obtain the corresponding attention weight. Calculate the weighted sum based on the attention score of each word and the corresponding attention weight as the internal context vector. If the word segment length is equal to 1, use the context representation vector of the word as the internal context vector.

[0022] S25, the boundary information of word segments, the fixed-dimensional width vector and the internal context vector are concatenated and fused to generate a candidate segment feature set.

[0023] Based on the above technical solutions, preferably, step S3, which involves defining entity types and calculating similarity with candidate fragment feature vectors to obtain the entity type with the highest similarity score and associate it with the candidate fragment to generate a flat entity set, includes the following sub-steps:

[0024] S31, define entity types in the domain of judicial cases, and encode them using a pre-trained type encoding model to obtain the embedding vectors corresponding to each entity type;

[0025] S32, use cosine similarity to calculate the cosine similarity between the embedding vector corresponding to each entity type and the feature vector of each candidate segment, and obtain the similarity score;

[0026] S33. Based on business requirements, obtain the number of tags corresponding to each entity type. If there is only one tag, obtain the candidate segment corresponding to the highest similarity score and associate it with the entity type to generate a flat entity set. If there are multiple tags, preset a similarity threshold, obtain the entity type corresponding to the similarity score greater than the similarity threshold, and calculate the word-level overlap rate between overlapping entity types. If the word-level overlap rate is greater than 0.5, retain the entity type with the high similarity score to obtain multiple entity types. Associate the multiple entity types with the corresponding candidate segments to generate a flat entity set.

[0027] Based on the above technical solutions, preferably, step S4, which involves constructing a predefined structured template and using a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances, includes the following sub-steps:

[0028] S41, combining the domain knowledge of legal experts with statistical analysis of a large number of structured judgment documents, constructs a predefined structured template;

[0029] S42, use a pre-trained encoding model to encode the text paragraph, and combine it with entity type distribution features to concatenate the input into the classifier to obtain the probability that the text paragraph contains each predefined structured template;

[0030] S43, preset probability threshold, obtain and activate the structured template corresponding to the output probability being greater than the probability threshold;

[0031] S44, Based on the activated structured template, filter entities in the flat entity set whose types match and use them as candidate entities to fill the slots;

[0032] S45. Each candidate entity and the filling slot are concatenated into a vector and input into a pre-trained feedforward neural network model. The filling probability is output, and the candidate entity with the highest filling probability is selected to fill the slot, thus obtaining the initial set of structured instances.

[0033] Based on the above technical solutions, preferably, step S5, which uses flat entities as nodes, identifies semantic relationships between entities based on a pre-trained relation classification model, extracts subgraph structures that match the structured template, maps and fills in the corresponding entity and relation information to obtain the final set of structured instances, includes the following sub-steps:

[0034] S51, treat flat entities as nodes in the graph and associate them with attribute information, which includes entity type, text content and fill slot information;

[0035] S52, traverse the flat entity set, combine any two entities to generate entity pairs, and use a pre-trained relation classification model to predict the relationship between entity pairs, which are then used as edges to connect the corresponding nodes to obtain the initial case graph.

[0036] S53. The subgraph matching algorithm is used to find the subgraph structure that matches the structured instance set on the initial case graph. 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.

[0037] Based on the above technical solutions, preferably, step S6 involves using a pre-constructed criminal domain knowledge graph and legal ontology to verify and correct the extracted entities, structured instances, and entity relationships, and then integrating the corrected entities with the structured instances and entity relationships to output the final judicial case identification result. This includes the following sub-steps:

[0038] S61, Select Claude4 as the teacher model and set the input template, which includes a task description, judicial case annotation and output format. The task description is to identify entities, relationships and structured instances from criminal case text.

[0039] S62, the teacher model identifies the unlabeled original text based on the input template, predicts and generates pseudo-labels for flat entities, structured instances and relationships, and removes labels whose output probability is lower than the screening threshold.

[0040] S63, Construct an external knowledge base, which includes an entity dictionary, a relation ontology, and a rule base;

[0041] Based on the entity dictionary, check the entity type, alias, and attributes in flat entities and pseudo-tags, and mark entities with mismatched types, incorrect aliases, or missing attributes;

[0042] Based on the relation ontology, check the relationship type, direction and attributes in the relationships between entities and pseudo tags, and mark relationships that are not allowed in type, have the wrong direction or unreasonable attributes.

[0043] Based on the rule base, check whether structured instances and pseudo-labels violate judicial logic constraints or laws and regulations, and mark instances that violate the constraints.

[0044] 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.

[0045] Secondly, the present invention also provides a named entity recognition system for case files based on semantic analysis, including a named entity recognition method for case files based on semantic analysis, the system comprising:

[0046] The collection and processing module is used to acquire the original text data of electronic case files of judicial cases, perform data preprocessing and word-level segmentation on the original text data, and obtain word sequence;

[0047] The candidate segment extraction module is used to process the word sequence using a sliding window mechanism, extract syntactic phrase and segment feature information, and fuse them to generate a candidate segment feature set;

[0048] The entity extraction module is used to define entity types and calculate similarity with the feature vectors of candidate segments. It obtains the entity type with the highest similarity score and associates it with the candidate segment to generate a flat entity set.

[0049] The structural information extraction module is used to construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances.

[0050] The relation information extraction module is used to identify semantic associations between entities based on a pre-trained relation classification model, with flat entities as nodes, and to extract subgraph structures that match the structured template, map and fill in the corresponding entity and relation information to obtain the final set of structured instances.

[0051] The correction output module is used to verify and correct the extracted entities, structured instances, and entity relationships using a pre-built criminal domain knowledge graph and legal ontology database. It then integrates the corrected entities with the structured instances and entity relationships to output the final judicial case identification results.

[0052] Thirdly, the present invention also provides a terminal device, the terminal device comprising: 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, the named entity recognition method program for case files based on semantic analysis being configured to implement the steps of a named entity recognition method for case files based on semantic analysis.

[0053] Fourthly, the present invention also provides a storage medium, which is a computer storage medium, storing a named entity recognition method program for case files based on semantic analysis. When the named entity recognition method program for case files based on semantic analysis is executed by a processor, it implements the steps of the named entity recognition method for case files based on semantic analysis.

[0054] The named entity recognition method and system for case files based on semantic analysis of the present invention have the following advantages over the prior art:

[0055] (1) By integrating efficient flat recognition mechanism with structured collaborative modeling for named entity recognition, it is possible to achieve efficient recognition of key entities in case files and automatically construct structured relationships and event information between entities, thereby improving the comprehensiveness, accuracy and processing efficiency of case semantic extraction.

[0056] (2) By identifying semantic relationships between entities through the relation classification model, we can deepen our understanding of case information and facilitate the discovery of complex relationships hidden in the text. The subgraph matching algorithm is used to quickly find the subgraph structure that matches the structured template, which improves the efficiency of information extraction.

[0057] (3) By using a layered or partially parallel decoding strategy, information at different levels can be processed in parallel, improving processing efficiency. At the same time, the layered decoding strategy covers all levels of the case structure, ensuring comprehensive extraction of information and providing richer data support for case analysis.

[0058] (4) Through the verification of entity dictionary, relation ontology and rule base, errors such as entity type mismatch, unreasonable relations and violation of judicial logic constraints are effectively identified and corrected, thereby improving the accuracy of entity, relation and structured instance identification. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a flowchart of a named entity recognition method for case files based on semantic analysis according to the present invention. Detailed Implementation

[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0062] like Figure 1 As shown, in a first aspect, the present invention provides a named entity recognition method for case files based on semantic analysis, comprising the following steps:

[0063] S1: Obtain the original text data of the electronic case file of the judicial case, perform data preprocessing and word-level segmentation on the original text data to obtain the word sequence.

[0064] Step S1 in this embodiment includes the following sub-steps:

[0065] S11, Obtain the original text data of electronic case files for judicial cases, identify and separate different areas of the document;

[0066] It should be noted that for PDF format electronic case files, the original text stream is first extracted using a PDF parsing library (pdfminer.six); then, based on rules (such as fixed formatting of headers and footers, font size and position of titles) or simple machine learning models (such as classifiers based on text block features), different areas of the document (such as cover, table of contents, body text, evidence list, court transcripts, etc.) are identified and separated.

[0067] S12, the document is divided into paragraphs according to line breaks, and each paragraph is divided into sentences using a sentence splitter to obtain a sentence sequence;

[0068] It should be noted that for the extracted main text, preliminary paragraph segmentation is first performed based on line breaks. A sentence segmenter is then applied to each paragraph. This can be done using a rule-based segmenter that combines common Chinese punctuation marks (。!?;) with specific legal terminology (such as "The judgment is hereby made as follows:"); or by using a pre-trained NLP toolkit (such as a Chinese model from spaCy or NLTK, fine-tuned for legal text) for sentence segmentation.

[0069] S13. Perform data preprocessing on the sentence sequence and use a word segmenter to perform word-level segmentation to obtain a word sequence.

[0070] It should be noted that preprocessing includes OCR error correction, noise removal, and format unification. OCR error correction can involve building a replacement dictionary for common OCR errors. For more complex errors, a Transformer-based sequence-to-sequence correction model can be trained, or the text correction capabilities of models such as ERNIE-Gram can be used. Noise removal uses regular expressions to remove fixed content in page numbers and headers, watermark text, and special control characters that are not text. Format unification converts full-width characters to half-width characters and standardizes the representation format of numbers and dates. Finally, the WordPiece word segmenter, which is compatible with the pre-trained Transformer model, is used to segment the processed sentence sequence into a word sequence.

[0071] In this embodiment, OCR error correction, noise removal, and format unification effectively improve the quality of the original text data, providing a more reliable data foundation for subsequent named entity recognition and structured information extraction, and enhancing processing efficiency and model adaptability.

[0072] S2 uses a sliding window mechanism to process the word sequence, extract syntactic phrase and fragment feature information, and fuse them to generate a candidate fragment feature set.

[0073] Step S2 in this embodiment includes the following sub-steps:

[0074] S21, a fixed-length window is preset and a sliding window mechanism is used to enumerate continuous word segments to obtain a set of candidate word segments;

[0075] It should be noted that by setting a minimum entity length L min =1 and maximum entity length L max =10, or based on the statistical analysis of the length of entities in the judicial domain, enumerate all continuous word segments within this length range; and use a lightweight Chinese dependency parser (such as a fast dependency parsing module based on LTP or HanLP) to extract all noun phrases, verb phrases and parallel noun phrases in the text; these syntactic components are usually good candidates to form named entities; and length restrictions can be imposed on the extracted syntactic phrases.

[0076] S22 uses a pre-trained encoding model to encode the word sequence in blocks, outputs the context representation vector of each word, and extracts the context representation vectors of the start and end words as boundary information.

[0077] 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 stride to encode the text in blocks; for overlapping words, their context representation vectors can be the average of multiple encoding results; the encoder outputs the context representation vector of each word and extracts the context representation vectors of the start and end words as boundary information.

[0078] S23 extracts the length of the word segment and generates a fixed-dimensional width vector through the embedding layer to represent the length information of the segment.

[0079] S24. Determine if the word segment length is greater than 1. If the word segment length is greater than 1, construct a feedforward neural network and use the feedforward neural network to calculate the corresponding attention score for each word in the word segment. Perform softmax normalization on the attention score to obtain the corresponding attention weight. Calculate the weighted sum based on the attention score of each word and the corresponding attention weight as the internal context vector. If the word segment length is equal to 1, use the context representation vector of the word as the internal context vector.

[0080] The expression for a feedforward neural network is:

[0081]

[0082] In the formula, h k This is the context representation vector for a word. v , Watt , b att These are learnable parameters;

[0083] The expression for softmax normalization is:

[0084]

[0085] In the formula, a k The attention score for each word;

[0086] The expression for calculating the internal context vector is:

[0087]

[0088] In the formula, h internal This is the internal context vector.

[0089] S25, the boundary information of the word segment, the fixed-dimensional width vector, and the internal context vector are concatenated and fused to generate a candidate segment feature set; the expression is:

[0090] .

[0091] In the formula, the boundary information of word segments, the fixed-dimensional width vector, and the internal context vector are concatenated and processed through a two-layer feedforward network. FFN span To achieve fusion and dimensionality reduction, FFN span The output dimension is consistent with the type embedding dimension, generating a candidate fragment feature set.

[0092] This embodiment uses a sliding window mechanism to enumerate continuous word segments, which can efficiently generate a set of candidate word segments and reduce unnecessary computation. For long texts, block encoding can avoid the limitation of model input length and improve processing efficiency. Furthermore, extracting the context representation vectors of the start and end words as boundary information and generating a fixed-dimensional width vector can more comprehensively represent the features of word segments. By concatenating and fusing the boundary information, width vector and internal context vector through a feedforward neural network, and calculating the internal context vector through an attention mechanism, the semantic relationships within word segments can be captured, improving model performance and adaptability.

[0093] S3 defines entity types and calculates their similarity with the feature vectors of candidate segments. It then associates the entity type with the highest similarity score with the candidate segment to generate a flat entity set.

[0094] Step S3 in this embodiment includes the following sub-steps:

[0095] S31 defines entity types in the field of judicial cases and encodes them using a pre-trained type encoding model to obtain the embedding vectors corresponding to each entity type.

[0096] It should be noted that a list of entity types and their natural language descriptions are maintained for the judicial field, such as Person (PER), Location (LOC), Time (TIME), and Money (MONEY).

[0097] 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 the type encoder; the natural language description of each type is input into the encoder, and the [CLS] vector or average pooling vector output by it is taken as the embedding representation of that type.

[0098] S32, cosine similarity is used to calculate the cosine similarity between the embedding vector corresponding to each entity type and the feature vector of each candidate segment, and the similarity score is obtained. The expression is:

[0099]

[0100] In the formula, Score(span(i,j),type_label k ) To score the similarity, Rep span(i,j) For the feature vector of the candidate segment, Rep type_labelk An embedding vector for the entity type;

[0101] S33. Based on business requirements, obtain the number of tags corresponding to each entity type. If there is only one tag, obtain the candidate segment corresponding to the highest similarity score and associate it with the entity type to generate a flat entity set. If there are multiple tags, preset a similarity threshold, obtain the entity type corresponding to the similarity score greater than the similarity threshold, and calculate the word-level overlap rate between overlapping entity types. If the word-level overlap rate is greater than 0.5, retain the entity type with the high similarity score to obtain multiple entity types. Associate the multiple entity types with the corresponding candidate segments to generate a flat entity set.

[0102] In this embodiment, both the candidate fragment feature vector and the entity type embedding vector contain rich semantic information. Through similarity calculation, candidate fragments can be accurately associated with entity types. A preset similarity threshold can filter out low-quality matching results, improving the accuracy of entity recognition. According to business needs, single-label and multi-label cases can be flexibly handled to meet the entity recognition needs in different scenarios. For entity types with overlap, by calculating the word-level overlap rate and retaining entity types with high similarity scores, duplicate recognition and incorrect association can be avoided, and a flat entity set is efficiently generated, providing accurate and reliable entity recognition results for subsequent analysis of electronic case files in judicial cases.

[0103] S4. Construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances.

[0104] In this embodiment, step S4 includes the following sub-steps:

[0105] S41, combining the domain knowledge of legal experts with statistical analysis of a large number of structured judgment documents, constructs a predefined structured template.

[0106] It should be noted that these schemas can be stored and managed using JSONSchema or a custom YAML format; for example, the structured template for an injury event is defined as follows:

[0107] schema_name:IntentionalInjuryEvent

[0108] Description: The core elements describing a harm event.

[0109] slots:

[0110] -name:perpetrator

[0111] type_constraint:PER_Suspect # Points to the flat solid type

[0112] is_list:true

[0113] Description: Perpetrator

[0114] -name:victim

[0115] type_constraint:PER_Victim

[0116] is_list:true

[0117] Description: Victim

[0118] -name:injury_location_body

[0119] type_constraint:BODY_PART # For example, fine-grained entities such as "head" and "chest".

[0120] is_list:true

[0121] Description: Injured area

[0122] -name:injury_degree

[0123] type_constraint:INJURY_LEVEL # For example, "minor injury level one", "serious injury level two".

[0124] is_list:false

[0125] Description: Severity level

[0126] -name:instrument_used

[0127] type_constraint:EVI_Weapon

[0128] is_list:true

[0129] Description: Tools used

[0130] -name:event_time

[0131] type_constraint:TIME_Exact

[0132] is_list:false

[0133] Description: Time of the event

[0134] -name:event_location

[0135] type_constraint:LOC_CrimeScene

[0136] is_list:false

[0137] Description: Location of the incident.

[0138] S42 uses a pre-trained encoding model (such as a Transformer encoder) to encode the text paragraphs, and combines the entity type distribution features to concatenate the input into the classifier to obtain the probability that the text paragraphs contain each predefined structured template.

[0139] S43, preset probability threshold, obtain and activate the structured template corresponding to the output probability being greater than the probability threshold;

[0140] S44, Based on the activated structured template, filter entities in the flat entity set whose types match and use them as candidate entities to fill the slots;

[0141] S45. Each candidate entity and the filling slot are concatenated into a vector and input into a pre-trained feedforward neural network model. The filling probability is output, and the candidate entity with the highest filling probability is selected to fill the slot, thus obtaining the initial set of structured instances.

[0142] It should be noted that the use of predefined structured templates reduces the workload of manual definition and information extraction, improving the efficiency of information extraction. Using a pre-trained classification model to encode and classify text paragraphs improves the accuracy of structured template activation. Furthermore, by predicting filling probabilities through vector concatenation and a feedforward neural network model, the highest-scoring candidate entity is selected to fill the slots, improving the accuracy of slot filling. The structured templates can be extended and modified according to different case types and needs, exhibiting good flexibility. Multiple predefined structured templates can be processed simultaneously, making it suitable for complex case analysis scenarios. It efficiently generates an initial set of structured instances, providing standardized and accurate information support for subsequent electronic case file analysis in judicial cases.

[0143] S5 uses flat entities as nodes, identifies semantic relationships between entities based on a pre-trained relation classification model, extracts subgraph structures that match the structured template, maps and fills in the corresponding entity and relation information, and obtains the final set of structured instances.

[0144] Step S5 in this embodiment includes the following sub-steps:

[0145] S51, treat flat entities as nodes in the graph and associate them with attribute information, which includes entity type, text content and fill slot information;

[0146] It should be noted that all the identified flat entities are used as the initial nodes in the graph. Each node contains attribute information, such as entity type, text content, and fill slot information.

[0147] S52, traverse the flat entity set, combine any two entities to generate entity pairs, and use a pre-trained relation classification model to predict the relationship between entity pairs, which are then used as edges to connect the corresponding nodes to obtain the initial case graph.

[0148] It should be noted that a pre-trained relation classification model is used to predict the relationships between entity pairs. This model can receive entity pairs and their contextual information as input, learn the semantic associations between entities, predict whether there is a predefined relation type between them, and connect the entity pairs with relationships with edges based on the prediction results of the relation classification model to form an initial case graph.

[0149] S53. The subgraph matching algorithm is used to find the subgraph structure that matches the structured instance set on the initial case graph. 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.

[0150] It should be noted that the subgraph matching algorithm is used to find subgraph structures that match the predefined structured template on the initial case graph. The subgraph matching algorithm finds the subgraphs that meet the requirements by comparing the slots and relationships between the subgraphs in the graph and the structured template. The nodes and edges in the subgraph structure are then mapped to the roles and relationships in the structured template, and the corresponding entity and relationship information is filled in. The final set of structured instances is obtained, which contains the key information required for case analysis.

[0151] This embodiment identifies semantic relationships between entities through a relation classification model, deepening the understanding of case information and facilitating the discovery of complex relationships hidden in the text. It employs a subgraph matching algorithm to quickly find subgraph structures that match the structured template, improving the efficiency of information extraction. By representing case information as a structured set of instances, it provides in-depth and accurate information support for subsequent analysis of electronic case files in judicial cases.

[0152] Each identified flat entity is assigned a unique ID within the document. This ID is unique within the document scope and is 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 the entity text. This avoids repeating the same entity text in the output and reduces redundant information.

[0153] For information that requires standardization, such as numerical values ​​and dates, standardized processing is performed to improve data consistency and comparability, and facilitates subsequent machine processing and storage. For slots that need to be extracted or summarized from the original text to generate text descriptions, lightweight controlled text generation models can be used, such as Pointer-Generator Network or small-scale pre-trained Seq2Seq models. These models generate short and accurate text based on slot descriptions and relevant context. The output length is strictly limited to a maximum of 50 tokens to ensure efficient tokenization and reduce computational resource consumption.

[0154] The final output uses a standardized, compact serialization format such as JSON or Protobuf, which facilitates subsequent machine processing and storage, improving the efficiency and accuracy of data processing.

[0155] For example, an event instance can be represented as an object containing a schema type, an instance ID, and the population values ​​for each slot, such as:

[0156] {

[0157] "schema_id":"FraudEvent_v1.2",

[0158] "instance_id":"uuid_xyz",

[0159] "slots":{

[0160] "Perpetrator":[

[0161] {

[0162] "entity_id":"flat_entity_id_001",

[0163] text":"Zhang San"

[0164] }

[0165] ],

[0166] "Victim":[

[0167] {

[0168] "entity_id":"flat_entity_id_002",

[0169] text":"Li Si"

[0170] }

[0171] ],

[0172] "Methods and Process of the Crime_text":"xxxxxx",

[0173] "Amount involved":{

[0174] "entity_id":"flat_entity_id_003",

[0175] "value":50000,

[0176] "currency":"RMB"

[0177] }...

[0178] }

[0179] }

[0180] This means that by referencing the ID of an already identified entity, duplicate entity text is avoided, and text fragments are only output for slots that require the original text.

[0181] In addition, for complex case structures with inherent hierarchical or dependent relationships, a layered or partially parallel decoding strategy is adopted, specifically as follows:

[0182] Based on the complexity of judicial cases, the structured model is divided into different levels: Level-0 is the core event layer, Level-1 is the related information layer, such as the "personal background information" template related to the core event participants 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.

[0183] First, the model centrally processes the identification and filling of core pattern instances at Level-0; once a Level-0 instance is identified, the parsing of Level-1 patterns is triggered; pattern filling related to core event participants and "evidence" template filling related to core events can be initiated in parallel; Level-1 parsing of different branches can be performed independently; if the output of Level-1 triggers a Level-2 pattern, the parsing continues downward; the results of the upper-level structure parsing can be used as input or constraints for the lower-level structure parsing to ensure the consistency and relevance of information.

[0184] In this embodiment, a hierarchical or partially parallel decoding strategy can be used to process information at different levels in parallel, improving processing efficiency. At the same time, the hierarchical decoding strategy covers all levels of the case structure, ensuring comprehensive information extraction and providing richer data support for case analysis.

[0185] S6 utilizes a pre-built criminal domain knowledge graph and legal ontology to verify and correct the extracted entities, structured instances, and entity relationships. It then integrates the corrected entities, structured instances, and entity relationships to output the final judicial case identification results.

[0186] Step S6 in this embodiment includes the following sub-steps:

[0187] S61, Select Claude4 as the teacher model and set the input template, which includes a task description, judicial case annotation and output format. The task description is to identify entities, relationships and structured instances from criminal case text.

[0188] It should be noted that open-source large models such as Claude4Opus or those fine-tuned with legal domain instructions are selected as teacher models. These models have powerful capabilities in processing complex text and generating structured outputs, and can provide high-quality pseudo-labels for student models. The input template Prompt is designed, which includes a task description, a small number of high-quality criminal case annotation examples, and output format requirements. The teacher model is required to output flat entities and structured relationships in a specific JSON format to guide student models in learning the correct annotation methods.

[0189] S62, the teacher model identifies the unlabeled original text based on the input template, predicts and generates pseudo-labels for flat entities, structured instances and relationships, and removes labels whose output probability is lower than the screening threshold.

[0190] It should be noted that the labeled outputs of the teacher model on unlabeled criminal texts are collected as pseudo-labels. A confidence assessment mechanism is designed, such as filtering out low-quality pseudo-labels based on the probability of the teacher model output or the consistency with the results of multiple different teacher models. For structured information, a simple logical consistency check can be performed to ensure the quality of pseudo-labels.

[0191] The distillation loss for flat entity recognition is also set, including logits matching for fragment classification and probability distribution matching for type selection. The KL divergence between the logits output by the student model and the teacher model is minimized. The logits of the teacher model are usually smoothed by a temperature coefficient T>1 to improve the distillation effect.

[0192] It also includes distillation loss for structured information extraction. 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 case summaries in text form, the student model can be trained to recover the structure from the summary, or the structure generated by the student model can be semantically close to the teacher's summary, which can be achieved through auxiliary semantic similarity loss.

[0193] Combining the distillation loss from flattened entity recognition and structured information extraction, a comprehensive loss function is formed, expressed as:

[0194]

[0195] In the formula, α and β are adjustable weights that can be determined by grid search on the performance on the validation set, and their common values ​​range from 0.1 to 0.9. L total The total loss function measures the overall loss of the model during training, taking into account both the loss from the true label and the distillation loss from different stages. L ground_truth The loss function is the loss function for the true label, used to measure the difference between the model's prediction and the actual label. Commonly used loss functions include cross-entropy loss. L distillation_S2 The distillation loss for flattened entity recognition is used in the knowledge distillation process to measure the difference between the outputs of the student model and the teacher model. L distillation_S3 This refers to the distillation loss in the extraction of structured information.

[0196] S63, Construct an external knowledge base, which includes an entity dictionary, a relation ontology, and a rule base;

[0197] Based on the entity dictionary, check the entity type, alias, and attributes in flat entities and pseudo-tags, and mark entities with mismatched types, incorrect aliases, or missing attributes;

[0198] Based on the relation ontology, check the relationship type, direction and attributes in the relationships between entities and pseudo tags, and mark relationships that are not allowed in type, have the wrong direction or unreasonable attributes.

[0199] Based on the rule base, check whether structured instances and pseudo-labels violate judicial logic constraints or laws and regulations, and mark instances that violate the constraints.

[0200] 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.

[0201] 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.

[0202] 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.

[0203] For example: Output of a flat entity:

[0204] {

[0205] "doc_id":"XYZ001",

[0206] "span_text":"Zhang San",

[0207] "start_char":100,

[0208] "end_char":102,

[0209] "entity_type":"PER_Suspect",

[0210] "confidence":0.95,

[0211] "source_module":"S2_FlatNER"

[0212] };

[0213] A structured event output:

[0214] {

[0215] "doc_id":"XYZ001",

[0216] "event_id":"EVT001",

[0217] "event_type":"RobberyEvent",

[0218] "confidence":0.88,

[0219] "source_module":"S3_StructExtract",

[0220] "slots":{

[0221] "perpetrator":[

[0222] "entity_id_zhangsan"

[0223] ],

[0224] "victim":[

[0225] "entity_id_lisi

[0226] ],...

[0227] }

[0228] }

[0229] This embodiment integrates entity dictionaries, relational ontologies, and rule bases to verify the extracted information. This not only effectively identifies and corrects errors such as entity type mismatches, unreasonable relationships, and violations of judicial logic constraints, significantly improving the accuracy of entity, relation, and structured instance identification, but also deeply integrates the knowledge graph and legal ontology of the judicial field into the identification process. This enables the model to more accurately understand and process domain-specific entities, relations, and structured information, significantly enhancing the model's domain adaptability. At the same time, the constraint rules defined based on legal logic and practical experience ensure that the extracted information strictly conforms to the requirements of judicial practice, thereby greatly improving the professionalism and reliability of the identification results.

[0230] Secondly, the present invention also provides a named entity recognition system for case files based on semantic analysis, including a named entity recognition method for case files based on semantic analysis, the system comprising:

[0231] The collection and processing module is used to acquire the original text data of electronic case files of judicial cases, perform data preprocessing and word-level segmentation on the original text data, and obtain word sequence;

[0232] The candidate segment extraction module is used to process the word sequence using a sliding window mechanism, extract syntactic phrase and segment feature information, and fuse them to generate a candidate segment feature set;

[0233] The entity extraction module is used to define entity types and calculate similarity with the feature vectors of candidate segments. It obtains the entity type with the highest similarity score and associates it with the candidate segment to generate a flat entity set.

[0234] The structural information extraction module is used to construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances.

[0235] The relation information extraction module is used to identify semantic associations between entities based on a pre-trained relation classification model, with flat entities as nodes, and to extract subgraph structures that match the structured template, map and fill in the corresponding entity and relation information to obtain the final set of structured instances.

[0236] The correction output module is used to verify and correct the extracted entities, structured instances, and entity relationships using a pre-built criminal domain knowledge graph and legal ontology database. It then integrates the corrected entities with the structured instances and entity relationships to output the final judicial case identification results.

[0237] It should be noted that this system is a system corresponding to the above-mentioned hydrodynamic compensation-based mechanical leg motion control method. All implementation methods in the above method embodiments are applicable to the embodiments of this system and can achieve the same technical effect.

[0238] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0239] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0240] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0241] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0242] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0243] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0244] Furthermore, it should be noted that in the system and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Moreover, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order; some steps can be executed in parallel or independently of each other. Those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in hardware, firmware, software, or a combination thereof. This is something that those skilled in the art can achieve by using their basic programming skills after reading the description of the present invention.

[0245] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing system. The computing system can be a known general-purpose system. Therefore, the object of the present invention can also be achieved simply by providing a program product containing program code implementing the method or apparatus. That is, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any known storage medium or any storage medium developed in the future. It should also be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent to the present invention. Furthermore, the steps performing the above series of processes can naturally be performed in the order described, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel or independently of each other.

[0246] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A named entity recognition method for case files based on semantic analysis, characterized in that, Includes the following steps: S1. Obtain the original text data of the electronic case file of the judicial case, perform data preprocessing and word-level segmentation on the original text data to obtain the word sequence; S2 uses a sliding window mechanism to process the word sequence, extract syntactic phrase and fragment feature information, and fuses them to generate a candidate fragment feature set; S3, define entity types and calculate similarity with candidate fragment feature vectors, obtain the entity type with the highest similarity score and associate it with the candidate fragment to generate a flat entity set; S4. Construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances. S5 uses flat entities as nodes, identifies semantic associations between entities based on a pre-trained relation classification model, extracts subgraph structures that match the structured template, maps and fills in the corresponding entity and relation information, and obtains the final set of structured instances. S6 utilizes a pre-built criminal domain knowledge graph and legal ontology to verify and correct the extracted entities, structured instances, and entity relationships. It then integrates the corrected entities, structured instances, and entity relationships to output the final judicial case identification results. Step S6 includes the following sub-steps: S61, Select Claude4 as the teacher model and set the input template, which includes a task description, judicial case annotation and output format. The task description is to identify entities, relationships and structured instances from criminal case text. S62, the teacher model identifies the unlabeled original text based on the input template, predicts and generates pseudo-labels for flat entities, structured instances and relationships, and removes labels whose output probability is lower than the screening threshold. S63, Construct an external knowledge base, which includes an entity dictionary, a relation ontology, and a rule base; Based on the entity dictionary, check the entity type, alias, and attributes in flat entities and pseudo-tags, and mark entities with mismatched types, incorrect aliases, or missing attributes; Based on the relation ontology, check the relationship type, direction and attributes in the relationships between entities and pseudo tags, and mark relationships that are not allowed in type, have the wrong direction or unreasonable attributes. Based on the rule base, check whether structured instances and pseudo-labels violate judicial logic constraints or laws and regulations, and mark instances that violate the constraints. 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.

2. The named entity recognition method for case files based on semantic analysis as described in claim 1, characterized in that: Step S1, which involves obtaining the original text data of electronic case files for judicial cases, performing data preprocessing and word-level segmentation on the original text data to obtain a word sequence, includes the following sub-steps: S11, Obtain the original text data of electronic case files for judicial cases, identify and separate different areas of the document; S12, the document is divided into paragraphs according to line breaks, and each paragraph is divided into sentences using a sentence splitter to obtain a sentence sequence; S13. Perform data preprocessing on the sentence sequence and use a word segmenter to perform word-level segmentation to obtain a word sequence.

3. The named entity recognition method for case files based on semantic analysis as described in claim 2, characterized in that: Step S2, which involves processing the word sequence using a sliding window mechanism to extract syntactic phrase and fragment feature information and fusing them to generate a candidate fragment feature set, includes the following sub-steps: S21, a fixed-length window is preset and a sliding window mechanism is used to enumerate continuous word segments to obtain a set of candidate word segments; S22, the word sequence is block-encoded using a pre-trained encoding model, outputting the context representation vector of each word, and extracting the context representation vectors of the start and end words as boundary information; S23, extract the length of the word segment and generate a fixed-dimensional width vector through the embedding layer; S24. Determine if the word segment length is greater than 1. If the word segment length is greater than 1, construct a feedforward neural network and use the feedforward neural network to calculate the corresponding attention score for each word in the word segment. Perform softmax normalization on the attention score to obtain the corresponding attention weight. Calculate the weighted sum based on the attention score of each word and the corresponding attention weight as the internal context vector. If the word segment length is equal to 1, use the context representation vector of the word as the internal context vector. S25, the boundary information of word segments, the fixed-dimensional width vector and the internal context vector are concatenated and fused to generate a candidate segment feature set.

4. The named entity recognition method for case files based on semantic analysis as described in claim 3, characterized in that: Step S3, which involves defining entity types and calculating similarity with candidate fragment feature vectors to obtain the entity type with the highest similarity score and associate it with the candidate fragment to generate a flat entity set, includes the following sub-steps: S31, define entity types in the domain of judicial cases, and encode them using a pre-trained type encoding model to obtain the embedding vectors corresponding to each entity type; S32, use cosine similarity to calculate the cosine similarity between the embedding vector corresponding to each entity type and the feature vector of each candidate segment, and obtain the similarity score; S33. Based on business requirements, obtain the number of tags corresponding to each entity type. If there is only one tag, obtain the candidate segment corresponding to the highest similarity score and associate it with the entity type to generate a flat entity set. If there are multiple tags, preset a similarity threshold, obtain the entity type corresponding to the similarity score greater than the similarity threshold, and calculate the word-level overlap rate between overlapping entity types. If the word-level overlap rate is greater than 0.5, retain the entity type with the high similarity score to obtain multiple entity types. Associate the multiple entity types with the corresponding candidate segments to generate a flat entity set.

5. The named entity recognition method for case files based on semantic analysis as described in claim 4, characterized in that: Step S4, which involves constructing a predefined structured template and using a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances, includes the following sub-steps: S41, combining the domain knowledge of legal experts with statistical analysis of a large number of structured judgment documents, constructs a predefined structured template; S42, use a pre-trained encoding model to encode the text paragraph, and combine it with entity type distribution features to concatenate the input into the classifier to obtain the probability that the text paragraph contains each predefined structured template; S43, preset probability threshold, obtain and activate the structured template corresponding to the output probability being greater than the probability threshold; S44, Based on the activated structured template, filter entities in the flat entity set whose types match and use them as candidate entities to fill the slots; S45. Each candidate entity and the filling slot are concatenated into a vector and input into a pre-trained feedforward neural network model. The filling probability is output, and the candidate entity with the highest filling probability is selected to fill the slot, thus obtaining the initial set of structured instances.

6. The named entity recognition method for case files based on semantic analysis as described in claim 5, characterized in that: Step S5, which uses flat entities as nodes, identifies semantic relationships between entities based on a pre-trained relation classification model, extracts subgraph structures that match the structured template, maps and fills in the corresponding entity and relation information to obtain the final set of structured instances, includes the following sub-steps: S51, treat flat entities as nodes in the graph and associate them with attribute information, which includes entity type, text content and fill slot information; S52, traverse the flat entity set, combine any two entities to generate entity pairs, and use a pre-trained relation classification model to predict the relationship between entity pairs, which are then used as edges to connect the corresponding nodes to obtain the initial case graph. S53. The subgraph matching algorithm is used to find the subgraph structure that matches the structured instance set on the initial case graph. 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. A named entity recognition system for case files based on semantic analysis, comprising the named entity recognition method for case files based on semantic analysis as described in any one of claims 1-6, characterized in that: The system includes: The collection and processing module is used to acquire the original text data of electronic case files of judicial cases, perform data preprocessing and word-level segmentation on the original text data, and obtain word sequence; The candidate segment extraction module is used to process the word sequence using a sliding window mechanism, extract syntactic phrase and segment feature information, and fuse them to generate a candidate segment feature set; The entity extraction module is used to define entity types and calculate similarity with the feature vectors of candidate segments. It obtains the entity type with the highest similarity score and associates it with the candidate segment to generate a flat entity set. The structural information extraction module is used to construct a predefined structured template and use a pre-trained classification model to match entities to slots in the predefined structured template to obtain an initial set of structured instances. The relation information extraction module is used to identify semantic associations between entities based on a pre-trained relation classification model, with flat entities as nodes, and to extract subgraph structures that match the structured template, map and fill in the corresponding entity and relation information to obtain the final set of structured instances. The correction output module is used to verify and correct the extracted entities, structured instances, and entity relationships using a pre-built criminal domain knowledge graph and legal ontology database. It then integrates the corrected entities with the structured instances and entity relationships to output the final judicial case identification results.

8. A terminal device, characterized in that, The terminal device includes: a memory, a processor, and a semantic analysis-based named entity recognition method program for case files stored in the memory and executable on the processor, the semantic analysis-based named entity recognition method program for case files configured to implement the steps of the semantic analysis-based named entity recognition method for case files as claimed in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer storage medium, and the computer storage medium stores a named entity recognition method program for case files based on semantic analysis. When the named entity recognition method program for case files based on semantic analysis is executed by the processor, it implements the steps of the named entity recognition method for case files based on semantic analysis as described in any one of claims 1 to 6.

Images