Automatic entity labeling method vertical to industry policies and regulations

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an automatic entity labeling method perpendicular to industry policies and regulations. The method comprises the following steps: analyzing a document format, judging whether the document is a text document, dynamically splitting the document into text blocks based on semantic integrity, vectorizing the text, identifying entities and attributes, vectorizing the entities, clustering based on vector similarity and the like. According to the method, through full-process automation of semantic-driven text partitioning, LLM entity recognition and disambiguation, knowledge network construction and graph fusion, the tedious links of reading-understanding-labeling-rechecking in traditional manual labeling are replaced, the requirements for professional labeling personnel and quality inspection personnel are reduced through the automatic process, and in addition, the automatic labeling efficiency is improved. Through cooperation of a dynamic partitioning strategy (reducing LLM single processing capacity), vector retrieval (quickly positioning similar entities) and map fusion (incremental updating knowledge), the labeling period is greatly shortened.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to an automatic entity labeling method for policies and regulations in vertical industries. Background Technology

[0002] With the rapid development of artificial intelligence technology, the construction and application of large-scale models in vertical fields has become the core direction for promoting the intelligent transformation of industries. The training of such models is highly dependent on the professional knowledge of specific fields. This knowledge usually exists in unstructured forms such as massive business data, complex policies and regulations, and specific handling procedures. Transforming unstructured industry knowledge into training data that can be used by machine learning models requires a systematic data annotation process. Generally, the larger the scale of high-quality labeled datasets, the more effectively the training effect and performance of large-scale models in vertical fields can be improved. Currently, the mainstream technical solutions for vertical industry data annotation mainly follow this process: First, the system releases an annotation task containing several units to be annotated; second, annotators with domain expertise complete the manual annotation; and finally, quality inspectors manually review the annotation results to ensure accuracy. However, this process has significant drawbacks: 1. Quality control relies too heavily on manual labor: The accuracy of the annotation results depends entirely on the final manual review process, lacking an automated verification mechanism; 2. Long annotation cycle: There is a long time interval between manual annotation and manual quality inspection, resulting in low overall data processing efficiency; 3. High labor costs and high dependence: A large number of labeling and quality inspection personnel with professional knowledge are required, resulting in high labor costs.

[0003] 4. Delayed error feedback: The lack of real-time quality monitoring and error correction mechanisms during the annotation process makes it difficult to detect annotation errors in the early stages. These errors tend to accumulate until they are identified in the quality inspection stage, further extending the processing cycle and increasing rework costs. Therefore, how to improve the efficiency of data annotation in vertical fields, reduce the over-reliance on professional manpower, shorten the annotation cycle, and achieve real-time quality control and timely error correction in the annotation process has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0004] This invention aims to solve the technical problems existing in the prior art, such as low efficiency, long cycle, excessive reliance on manual quality inspection, and difficulty in timely detection and correction of annotation errors in vertical industry data annotation. Therefore, this invention provides an automatic entity annotation method for vertical industry policies and regulations. Its core lies in building a dynamic processing flow that integrates semantic understanding, entity disambiguation and knowledge graph construction to achieve real-time decomposition of policies and regulations and automatic annotation of business data.

[0005] To achieve the above objectives, the present invention specifically adopts the following technical solution: This invention proposes an automatic entity annotation method for vertical industry policies and regulations, the method comprising the following steps: (a) Document format parsing: After obtaining the document to be processed, parse it; (ii) Is it a text document: Determine whether the parsed text is a text document (including Word, PDF, TXT, etc.); if not, convert it to text using OCR or audio conversion technology before proceeding to the next step; if yes, proceed directly to the next step. (III) Dynamically split into text blocks based on semantic integrity: Using chapter titles, section titles, article numbers, clause start marks and paragraph separators in the text document as cutting points, the text document is divided into text blocks that do not overlap and maintain independent semantic integrity. (iv) Text vectorization: Each segmented text block is vectorized, and a unique identifier chunk-id is generated for each text block; (V) Entity Recognition and Attributes: For the segmented and vectorized text blocks, the Large Language Model (LLM) is used to analyze and identify entities one by one. Each entity is assigned an entity type (used to distinguish entity categories, such as "service object" or "business scenario"), entity description, entity name, and associated chunk-id (i.e., the unique identifier of the text block from which the entity originates). (vi) Entity vectorization: The entity type, entity description, entity name and associated chunk-id assigned in step (v) are combined and vectorized to obtain the entity combination vector; (vii) Clustering based on vector similarity: Calculate the entity combination vector (including semantic similarity and structured attribute matching degree), and classify the entity combination vector with similarity exceeding a preset threshold into the preliminary similar entity set; (viii) Entity disambiguation: The initial similar entity set is checked one by one by the large language model LLM to verify semantic consistency and attribute matching degree. After disambiguation, the final similar entity set is obtained. Multiple entity combination vectors in the final similar entity set are disambiguated into a single entity combination. (ix) Constructing a knowledge network: retrieve associated text fragments by using the chunk-id in the entity combination vector, obtain text fragments and relation networks similar to the selected entity by using vector retrieval, and integrate the retrieved text and relation networks by using the Large Language Model (LLM) to construct the knowledge network of the current document; (x) Obtain entities from the knowledge network: Extract entity information from the knowledge network constructed in step (ix) using vector retrieval; (xi) Determine if the entity is unique in the knowledge graph: Determine if the entity information obtained in step (x) is unique in the initial knowledge graph; if not, proceed to step (xii); if yes, proceed directly to step (xiv). (xii) Graph entity disambiguation: The entity information obtained in step (x) is compared with the existing entity set in the initial knowledge graph. When the similarity reaches the preset threshold, the entities in the current document are disambiguated with the existing entities in the initial knowledge graph to eliminate cross-document entity ambiguity. (xiii) Graph entity merging: Merge the disambiguated entities into the initial knowledge graph and record the mapping relationship between the merged entities and the entities in the initial knowledge graph; (xiv) Knowledge graph relationship integration: Calculate the matching degree based on the entity attributes (name, type, description) of the current document and the attribute features of the business information in the initial knowledge graph. When the matching degree reaches the threshold, mark the entities and related text fragments of the current document as related to the business information and establish a relationship; if the threshold is not reached, directly insert the entity information of the current document into the initial knowledge graph and establish a relationship with the original graph. (xv) Construct a comprehensive knowledge graph: Through the merging of graph entities and the integration of graph relationships, a comprehensive knowledge graph covering policies, regulations and business scenarios is formed.

[0006] As a preferred technical solution of the present invention, the dynamic text segmentation based on semantic integrity described in step (iii) uses the sentence boundaries of the text as the basis for segmentation, ensuring that each segmented text block independently expresses complete semantics and avoiding semantic breaks caused by fixed-length segmentation.

[0007] As a preferred technical solution of the present invention, the Large Language Model (LLM) mentioned in step (V) includes, but is not limited to, the Qwen-Max model, the entity type is used to distinguish the specific category of the entity, and the entity description is used to provide a detailed explanation of the entity to distinguish entities with similar names.

[0008] As a preferred technical solution of the present invention, the preset threshold in step (vii) is dynamically adjusted according to the entity type, for example, the semantic similarity threshold is 0.89 and the structured attribute matching threshold is 91%.

[0009] As a preferred technical solution of the present invention, the entity disambiguation in step (viii) verifies the semantic consistency and attribute matching degree of the preliminary similar entity set through a large language model (LLM) to ensure that the entities after disambiguation are unique and accurate.

[0010] As a preferred technical solution of the present invention, the knowledge network construction in step (ix) includes text backtracking retrieval and graph structure retrieval: the original text fragments associated with entities are retrieved by chunk-id, the relationship network similar to the entities in the initial knowledge graph is obtained by vector retrieval, and finally the knowledge network of the current document is generated by integration.

[0011] As a preferred technical solution of the present invention, the graph entity disambiguation in step (twelfth) disambiguates the entities in the current document with similar entities in the initial knowledge graph through vector similarity calculation and attribute matching (including name, type and description), thereby eliminating cross-document entity ambiguity.

[0012] As a preferred technical solution of the present invention, the matching degree calculation in step (xiv) is based on the semantic similarity between entity attribute features (such as name, type, description) and business information attribute features (such as business name, business description), and the matching degree threshold is dynamically adjusted according to the vertical industry field.

[0013] As a preferred technical solution of the present invention, the method further includes a real-time quality monitoring mechanism: in the process of entity disambiguation in step (viii) and graph entity disambiguation in step (xii), inconsistencies or errors in annotation are identified in real time through vector similarity detection and realism evaluation of large language model LLM, thereby reducing the reliance on manual quality inspection.

[0014] As a preferred technical solution of the present invention, when the method is applied to a vertical industry (such as social insurance), it further includes: dynamically generating sub-graphs (such as pension insurance sub-graphs and work injury insurance sub-graphs) based on a comprehensive knowledge graph, and supporting real-time intelligent question answering, and quickly outputting the policy basis and related policy fragments corresponding to the business through semantic understanding and intent recognition of a large language model LLM.

[0015] The beneficial effects of this invention are as follows: This invention automates the entire process of semantic-driven text segmentation, LLM entity recognition and disambiguation, knowledge network construction, and graph fusion, replacing the tedious "read-understand-annotate-review" steps of traditional manual annotation. The automated process reduces the need for professional annotators and quality control personnel. Furthermore, the synergy of dynamic segmentation strategies (reducing the amount of data processed per LLM cycle), vector retrieval (rapidly locating similar entities), and graph fusion (incrementally updating knowledge) significantly shortens the annotation cycle. In addition, LLM's semantic understanding capabilities and vector similarity detection mechanism avoid individual differences in manual annotation (such as different interpretations of "deceased worker"), ensuring consistency in entity naming, types, and relationships. Simultaneously, real-time quality monitoring (such as LLM verification in entity disambiguation) enables early detection and correction of errors, further improving annotation accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0017] This invention proposes an automatic entity annotation method for vertical industry policies and regulations, the method comprising the following steps: (a) Document format parsing: After obtaining the document to be processed, parse it; (ii) Is it a text document: Determine whether the parsed text is a text document (including Word, PDF, TXT, etc.); if not, convert it to text using OCR or audio conversion technology before proceeding to the next step; if yes, proceed directly to the next step. (III) Dynamically split into text blocks based on semantic integrity: Using chapter titles, section titles, article numbers, clause start marks and paragraph separators in the text document as cutting points, the text document is divided into text blocks that do not overlap and maintain independent semantic integrity. (iv) Text vectorization: Each segmented text block is vectorized, and a unique identifier chunk-id is generated for each text block; (V) Entity Recognition and Attributes: For the segmented and vectorized text blocks, the Large Language Model (LLM) is used to analyze and identify entities one by one. Each entity is assigned an entity type (used to distinguish entity categories, such as "service object" or "business scenario"), entity description, entity name, and associated chunk-id (i.e., the unique identifier of the text block from which the entity originates). (vi) Entity vectorization: The entity type, entity description, entity name and associated chunk-id assigned in step (v) are combined and vectorized to obtain the entity combination vector; (vii) Clustering based on vector similarity: Calculate the entity combination vector (including semantic similarity and structured attribute matching degree), and classify the entity combination vector with similarity exceeding a preset threshold into the preliminary similar entity set; (viii) Entity disambiguation: The initial similar entity set is checked one by one by the large language model LLM to verify semantic consistency and attribute matching degree. After disambiguation, the final similar entity set is obtained. Multiple entity combination vectors in the final similar entity set are disambiguated into a single entity combination. (ix) Constructing a knowledge network: retrieve associated text fragments by using the chunk-id in the entity combination vector, obtain text fragments and relation networks similar to the selected entity by using vector retrieval, and integrate the retrieved text and relation networks by using the Large Language Model (LLM) to construct the knowledge network of the current document; (x) Obtain entities from the knowledge network: Extract entity information from the knowledge network constructed in step (ix) using vector retrieval; (xi) Determine if the entity is unique in the knowledge graph: Determine if the entity information obtained in step (x) is unique in the initial knowledge graph; if not, proceed to step (xii); if yes, proceed directly to step (xiv). (xii) Graph entity disambiguation: The entity information obtained in step (x) is compared with the existing entity set in the initial knowledge graph. When the similarity reaches the preset threshold, the entities in the current document are disambiguated with the existing entities in the initial knowledge graph to eliminate cross-document entity ambiguity. (xiii) Graph entity merging: Merge the disambiguated entities into the initial knowledge graph and record the mapping relationship between the merged entities and the entities in the initial knowledge graph; (xiv) Knowledge graph relationship integration: Calculate the matching degree based on the entity attributes (name, type, description) of the current document and the attribute features of the business information in the initial knowledge graph. When the matching degree reaches the threshold, mark the entities and related text fragments of the current document as related to the business information and establish a relationship; if the threshold is not reached, directly insert the entity information of the current document into the initial knowledge graph and establish a relationship with the original graph. (xv) Construct a comprehensive knowledge graph: Through the merging of graph entities and the integration of graph relationships, a comprehensive knowledge graph covering policies, regulations and business scenarios is formed; In step (iii), the dynamic text segmentation based on semantic integrity uses the sentence boundaries of the text as the basis for segmentation, ensuring that each segmented text block independently expresses complete semantics and avoiding semantic breaks caused by fixed-length segmentation. Among them, the large language model LLM mentioned in step (v) includes, but is not limited to, the Qwen-Max model, the entity type is used to distinguish the specific category of the entity, and the entity description is used to provide a detailed explanation of the entity to distinguish entities with similar names; In step (vii), the preset threshold is dynamically adjusted according to the entity type, for example, the semantic similarity threshold is 0.89 and the structured attribute matching threshold is 91%. In step (viii), the entity disambiguation is verified by using a large language model (LLM) to check the semantic consistency and attribute matching degree of the preliminary similar entity set, in order to ensure that the entities after disambiguation are unique and accurate. Among them, the knowledge network construction described in step (ix) includes text backtracking retrieval and graph structure retrieval: the original text fragments associated with entities are retrieved by chunk-id, the relationship network similar to the entity in the initial knowledge graph is obtained by vector retrieval, and finally the knowledge network of the current document is generated by integration; In step (xii), the graph entity disambiguation is performed by calculating vector similarity and matching attributes (including name, type, and description) to disambiguate the entities in the current document with similar entities in the initial knowledge graph, thereby eliminating cross-document entity ambiguity. In step (xiv), the matching degree calculation is based on the semantic similarity between entity attribute features (such as name, type, description) and business information attribute features (such as business name, business description), and the matching degree threshold is dynamically adjusted according to the vertical industry field. The method also includes a real-time quality monitoring mechanism: in the process of entity disambiguation in step (viii) and graph entity disambiguation in step (xii), inconsistencies or errors in annotation are identified in real time through vector similarity detection and realism evaluation of large language model LLM, reducing the reliance on manual quality inspection; When the method is applied to vertical industries (such as social insurance), it also includes: dynamically generating sub-graphs (such as pension insurance sub-graphs and work injury insurance sub-graphs) based on a comprehensive knowledge graph, and supporting real-time intelligent question answering. Through semantic understanding and intent recognition of the large language model LLM, it can quickly output the policy basis and related policy fragments corresponding to the business. The aforementioned method for automatically annotating entities related to policies and regulations in vertical industries works through three core mechanisms: semantic-driven text processing, entity management empowered by a large language model (LLM), and multi-dimensional knowledge fusion. Its specific working principle is as follows: For semantic-driven text segmentation and vectorization: For the complex structure of policy and regulatory texts (including chapters, sections, articles, clauses, etc.), a dynamic segmentation strategy based on semantic integrity is first used. The text is segmented into non-overlapping text blocks that maintain independent semantic integrity, using chapter titles, section titles, article numbers, clause start marks, and paragraph separators in the document as cutting points. This step avoids semantic breaks caused by fixed-length segmentation and ensures that each text block can independently express a complete meaning (such as a policy clause or a complete business scenario description). After segmentation, each text block is vectorized (e.g., by generating vector representations through a pre-trained language model), and a unique identifier, chunk-id, is generated for each text block. Vectorization converts the text into a machine-understandable numerical form, which facilitates subsequent entity recognition and relationship matching; the chunk-id provides a traceability basis for the association between entities and the original text. The core purpose of this step is to reduce the complexity of LLM processing (the amount of information in a single block of text is controllable) while ensuring semantic integrity (avoiding semantic confusion across blocks), thereby improving the accuracy of subsequent entity recognition. For LLM-enabled entity recognition and disambiguation: For the segmented text blocks, the Large Language Model (LLM) is used to analyze and identify entities (such as policy or business-related entities such as "deceased employee" and "work injury determination") block by block. LLM not only extracts entity descriptions, but also assigns entity type (such as "service object" and "business scenario") and entity description (such as "deceased employee is an employee who died due to a work-related accident, and his / her dependent relatives can enjoy pension benefits") to distinguish entities with similar names (such as "deceased employee" and "work injury employee"). Subsequently, the entity description, type, and associated chunk-id are combined and vectorized to form a multi-dimensional vector representation of the entity. By calculating the vector similarity of all entities (including semantic similarity and structured attribute matching), entities with similarity exceeding the threshold are included in the preliminary similar entity set. Then, LLM is used to verify them one by one, eliminate ambiguity and align them into a single entity (e.g., merging "dead worker" and "injured worker" into "injured worker") to ensure the uniqueness of the entity. This step leverages the semantic understanding capabilities of LLM to achieve accurate entity identification and ambiguity resolution, solving the problem of entity naming confusion caused by individual differences in traditional manual annotation. For knowledge network construction and graph fusion: After completing entity processing of the current document, the original text fragments associated with the entities are retrieved back through chunk-id, and vector retrieval is used to obtain similar relationship networks (such as the relationship between "deemed work-related injury" and "work-related injury determination") from the initial knowledge graph. LLM integrates the retrieved text fragments and relationship networks to construct the knowledge network of the current document, forming a preliminary connection between policy clauses and business scenarios. Furthermore, the knowledge network of the current document is integrated with the initial knowledge graph: Graph entity disambiguation: By comparing vector similarity, entities in the current document are disambiguated with existing entities in the initial knowledge graph (e.g., merging newly identified "work injury insurance" with existing "work injury insurance" entities in the graph), thus eliminating cross-document entity ambiguity. Entity merging and relationship integration: Merge the disambiguated entities into the initial knowledge graph and record the mapping relationships; at the same time, based on the matching degree between entity attributes (name, type, description) and business information attributes (business name, business description), establish the association between policy fragments and businesses (such as the association between the policy clause "death within 48 hours after ineffective rescue" and the business "work-related death determination"). Ultimately, through entity merging and relationship integration, a comprehensive knowledge graph covering policies, regulations, and business scenarios is formed, providing structured training data for large-scale models in vertical domains.

[0018] This invention automates the entire process of semantic-driven text segmentation, LLM entity recognition and disambiguation, knowledge network construction, and graph fusion, replacing the tedious "read-understand-annotate-review" steps of traditional manual annotation. The automated process reduces the need for professional annotators and quality control personnel. Furthermore, the synergy of dynamic segmentation strategies (reducing the amount of data processed per LLM cycle), vector retrieval (rapidly locating similar entities), and graph fusion (incrementally updating knowledge) significantly shortens the annotation cycle. In addition, LLM's semantic understanding capabilities and vector similarity detection mechanism avoid individual differences in manual annotation (such as different interpretations of "deceased worker"), ensuring consistency in entity naming, types, and relationships. Simultaneously, real-time quality monitoring (such as LLM verification in entity disambiguation) enables early detection and correction of errors, further improving annotation accuracy. In summary, this invention effectively solves the problems of low efficiency, lengthy cycle, high dependence on manual labor, and delayed error feedback in vertical domain data annotation, providing a reliable data foundation for building high-quality large-scale models in vertical domains. Example: Application of automatic entity annotation in social insurance industry policies and regulations: This embodiment takes the social insurance field as its background and addresses the needs of building an industry knowledge base. It integrates 2,800 policies and regulations, 529 policy interpretation documents, 1,000+ business handling records, and 10,000+ business Q&A pairs. It achieves efficient conversion of policy knowledge into business knowledge through automated entity annotation methods. During system operation, the system first uses a dynamic chunking strategy based on semantic integrity to divide the policy and regulatory text into 81,200 semantically complete text blocks, using chapter titles, section titles, article numbers, clause start marks, and paragraph separators as cutting points. A unique identifier, chunk-id, is generated for each text block. Subsequently, each text block is vectorized. Based on this, a large language model (such as Qwen-Max) is driven to analyze the blocks one by one, identify entities such as "deceased workers," and assign them types such as "service recipients" and descriptive information such as "deceased workers are employees who died due to work-related accidents, and their dependents are entitled to pension benefits." This facilitates subsequent alignment and disambiguation. For example, for the text fragment "died from a sudden illness during working hours and at the workplace, or died within 48 hours after emergency treatment," the model accurately extracts the entity "deceased workers" and distinguishes it from "injured workers" through type and description. In the entity disambiguation process, the system automatically calculates the vector similarity of all entities (e.g., the semantic similarity between "dead worker" and "injured worker" is 0.89 and the structured attribute matching degree is 91%), classifies high-similarity entities into the preliminary similarity set, and then verifies the disambiguation into a single entity through a large language model (e.g., merged into "injured worker") to ensure entity uniqueness. Subsequently, the system retrieves the original text fragments associated with entities by backtracking through chunk-id, and obtains the relationship network in the initial knowledge graph (such as the business entity "work injury determination") by combining vector retrieval. This network is then integrated by a large language model to generate the knowledge network of the current document, establishing the dependency relationship between policy fragments and business scenarios (such as the association between "deemed work injury" and "work injury determination"). Finally, based on the integrated knowledge network, the system dynamically generates sub-domain graphs such as "pension insurance sub-graph" and "work injury insurance sub-graph," and supports real-time intelligent question answering on the business side. When business personnel query, the system outputs relevant policy basis and specific fragments within 150ms through semantic understanding and intent recognition of the large language model, with a hit rate of 98.8%. In terms of technical performance, this method significantly outperforms traditional solutions: the accuracy of policy and regulation semantic integrity decomposition reaches 91.3% (compared to only 84% by traditional rule engines), the accuracy of policy and business annotation is improved to 93.2% (compared to 90% by traditional manual annotation), and the comprehensiveness is increased from 70% to 90%; the response time for decomposing complex text documents is reduced to less than 1 second (accelerated by 10 times through the VLLM inference framework); the processing time for 100 policies is significantly reduced from 240 hours to 10 hours, improving work efficiency by 24 times, effectively supporting the needs of large-scale model training and intelligent application on the business side in the social insurance field.

[0019] As an alternative, the present invention also provides an automatic entity labeling system for vertical industry policies and regulations, comprising: Document parsing module: Used to acquire and parse the document to be processed, and determine whether it is a text document; if not, it is converted into text using OCR or audio conversion technology; Text Blocking Module: Used to divide text into non-overlapping text blocks that maintain independent semantic integrity, using chapter titles, section titles, article numbers, paragraph start marks, and paragraph separators in a text document as cutting points; Text vectorization module: used to vectorize each text block and generate a unique identifier chunk-id; Entity recognition module: Used to perform entity recognition on segmented text blocks using a large language model (LLM), extracting entity type, entity name, entity description and associated chunk-id; Entity vectorization module: used to combine entity type, entity name, entity description and chunk-id and vectorize them to generate entity combination vectors; Entity clustering module: used to calculate the similarity of entity combination vectors (including semantic similarity and structured attribute matching degree), and classify entities exceeding the threshold into the preliminary similar entity set; Entity disambiguation module: used to validate the initial set of similar entities using a large language model (LLM) and eliminate ambiguity to a single entity combination; Knowledge Network Construction Module: Used to retrieve related text fragments by chunk-id, combine them with the initial knowledge graph relationship network obtained from vector retrieval, and integrate them to generate the knowledge network of the current document; The graph fusion module is used to determine the uniqueness of entities in the knowledge network within the initial knowledge graph, perform disambiguation, merging, and relation integration, and ultimately construct a comprehensive knowledge graph.

[0020] As an alternative, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the automatic entity annotation method for vertical industry policies and regulations as described above.

[0021] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0022] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An automatic entity annotation method for vertical industry policies and regulations, characterized in that, The method includes the following steps: (a) Document format parsing: After obtaining the document to be processed, parse it; (ii) Is it a text document: Determine whether the parsed text is a text document; if not, convert it to text using OCR or audio conversion technology before proceeding to the next step; if yes, proceed directly to the next step. (iii) Dynamically split into text blocks based on semantic integrity: Using chapter titles, section titles, article numbers, clause start marks and paragraph separators in the text document as cutting points, the text document is divided into text blocks that do not overlap and maintain independent semantic integrity. (iv) Text vectorization: Each segmented text block is vectorized, and a unique identifier chunk-id is generated for each text block; (v) Entity recognition and attributes: For the segmented and vectorized text blocks, the Large Language Model (LLM) is used to analyze and identify entities one by one, and each entity is assigned an entity type, entity description, entity name and associated chunk-id; (vi) Entity vectorization: The entity type, entity description, entity name and associated chunk-id assigned in step (v) are combined and vectorized to obtain the entity combination vector; (vii) Clustering based on vector similarity: Calculate the entity combination vectors and classify the entity combination vectors with similarity exceeding a preset threshold into the preliminary similar entity set; (viii) Entity disambiguation: The initial similar entity set is checked one by one by the large language model LLM to verify semantic consistency and attribute matching degree. After disambiguation, the final similar entity set is obtained. Multiple entity combination vectors in the final similar entity set are disambiguated into a single entity combination. (ix) Constructing a knowledge network: retrieve associated text fragments by using the chunk-id in the entity combination vector, obtain text fragments and relation networks similar to the selected entity by using vector retrieval, and integrate the retrieved text and relation networks by using the Large Language Model (LLM) to construct the knowledge network of the current document; (x) Obtain entities from the knowledge network: Extract entity information from the knowledge network constructed in step (ix) using vector retrieval; (xi) Determine if the entity is unique in the knowledge graph: Determine if the entity information obtained in step (x) is unique in the initial knowledge graph; if not, proceed to step (xii); if yes, proceed directly to step (xiv). (xii) Graph entity disambiguation: The entity information obtained in step (x) is compared with the existing entity set in the initial knowledge graph. When the similarity reaches the preset threshold, the entities in the current document are disambiguated with the existing entities in the initial knowledge graph to eliminate cross-document entity ambiguity. (xiii) Merging of knowledge graph entities: Merge the disambiguated entities into the initial knowledge graph and record the mapping relationship between the merged entities and the entities in the initial knowledge graph; (xiv) Graph Relationship Integration: Calculate the matching degree based on the entity attributes of the current document and the attribute features of the business information in the initial knowledge graph. When the matching degree reaches the threshold, mark the entities and associated text fragments of the current document as related to the business information and establish a relationship. If the threshold is not reached, directly insert the entity information of the current document into the initial knowledge graph and establish a relationship with the original graph. (xv) Construct a comprehensive knowledge graph: Through the merging of graph entities and the integration of graph relationships, a comprehensive knowledge graph covering policies, regulations and business scenarios is formed.

2. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The dynamic text segmentation based on semantic integrity described in step (iii) uses the sentence boundaries of the text as the basis for segmentation.

3. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The large language model (LLM) mentioned in step (V) includes, but is not limited to, the Qwen-Max model. The entity type is used to distinguish the specific category of the entity, and the entity description is used to provide a detailed explanation of the entity to distinguish entities with similar names.

4. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The preset threshold mentioned in step (vii) is dynamically adjusted according to the entity type.

5. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The entity disambiguation in step (viii) verifies the semantic consistency and attribute matching degree of the preliminary similar entity set through a large language model (LLM) to ensure that the entities after disambiguation are unique and accurate.

6. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The knowledge network construction described in step (ix) includes text backtracking retrieval and graph structure retrieval: the original text fragments associated with entities are retrieved by chunk-id, the relationship network similar to the entities in the initial knowledge graph is obtained by vector retrieval, and finally the knowledge network of the current document is generated by integration.

7. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The graph entity disambiguation described in step (twelfth) uses vector similarity calculation and attribute matching to disambiguate the entities in the current document with similar entities in the initial knowledge graph, thereby eliminating cross-document entity ambiguity.

8. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The matching degree calculation in step (xiv) is based on the semantic similarity between entity attribute features and business information attribute features, and the matching degree threshold is dynamically adjusted according to the vertical industry field.

9. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, The method also includes a real-time quality monitoring mechanism: in the process of entity disambiguation in step (viii) and graph entity disambiguation in step (xii), inconsistencies or errors in annotation are identified in real time through vector similarity detection and realism assessment of large language model LLM.

10. The automatic entity annotation method for vertical industry policies and regulations according to claim 1, characterized in that, When the method is applied to vertical industries, it also includes: dynamically generating sub-graphs based on a comprehensive knowledge graph and supporting real-time intelligent question answering; and quickly outputting the policy basis and related policy fragments corresponding to the business through semantic understanding and intent recognition of a large language model (LLM).

Images