Search

CN-121981229-A - Emergency field knowledge graph construction method capable of realizing full-link audit

CN121981229ACN 121981229 ACN121981229 ACN 121981229ACN-121981229-A

Abstract

The application belongs to the technical field of emergency management, and particularly discloses a full-link auditable emergency field knowledge graph construction method, which comprises the steps of converting unstructured text in an emergency plan document into semantically continuous text blocks; the method comprises the steps of converting text blocks conforming to normative judgment into structured rule data, grouping the rule data, acquiring entities belonging to predefined entity types in each group of rule data and relations among the entities based on a large language model, mapping all different expressions pointing to the same entity to normative names, identifying entities which are lack of connection with other entities based on the semantics of emergency plan documents, supplementing bridging edges connected with isolated entities, forming a knowledge graph with a graph structure of a data source-entity type-entity, wherein the bridging edges are provided with evidence anchor points, the confidence is higher than a preset threshold, connecting the knowledge graph of a plurality of emergency plan documents with nodes of a preset disaster chain, and representing different disaster events by the nodes in the preset disaster chain.

Inventors

  • JI ZHEYUAN
  • XIE XIA

Assignees

  • 海南大学

Dates

Publication Date
20260505
Application Date
20260130

Claims (8)

  1. 1. The utility model provides a method for constructing an emergency domain knowledge graph capable of being audited by a full link, which is characterized by comprising the following steps: converting unstructured text in the emergency plan document into semantically continuous text blocks; Converting the text block conforming to the normalization judgment into a plurality of structured rule data, grouping the structured rule data, and grouping the rule data with similar semantics into the same group; Acquiring entities belonging to a predefined entity type in each group of rule data based on a large language model and preset prompt words, and a relationship consistent with the emergency plan document among the entities, and mapping all different expressions pointing to the same entity to the same standard name; Identifying an isolated entity which is lack of connection with other entities based on the semantics of the emergency plan document, and supplementing a bridging edge connected with the isolated entity based on the context of the emergency plan document and a preset predicate whitelist to form a knowledge graph with a graph structure of a data source-entity type-entity, wherein the bridging edge is provided with an evidence anchor point, and the confidence level is higher than a preset threshold value; and respectively connecting the knowledge maps of the emergency plan documents with nodes of a preset disaster chain, wherein the nodes in the preset disaster chain are used for representing different disaster events.
  2. 2. The full-link auditable emergency domain knowledge graph construction method of claim 1, wherein converting unstructured text in an emergency plan document into semantically continuous text blocks comprises: dividing a long text in an emergency plan document according to punctuation marks at the end of a sentence to obtain a plurality of clauses; and sequentially placing the clauses into a plurality of empty text blocks with the upper limit of the preset token, copying the text with the preset length at the tail of the current text block to the prefix of the next empty text block after the current text block is filled, and then sequentially placing the follow-up text.
  3. 3. The full-link auditable emergency domain knowledge graph construction method of claim 1, further comprising: storing the knowledge graph into a graph database, and establishing unique constraint on the graph database level to ensure that the same entity type is not repeatedly established in the same data source range; and deleting entity nodes which are only connected with the entity types in the knowledge graph, and cleaning the relation labels.
  4. 4. The full-link auditable emergency domain knowledge graph construction method of claim 1, further comprising: The method comprises the steps of constructing an entity index layer, a rule index layer and a context index layer, wherein the entity index layer is used for vectorizing and storing the entity and the attribute of the entity, and is also used for quickly aligning natural language problems input by a user to specific nodes in a knowledge graph, the rule index layer is used for vectorizing and storing the structured rule data, and is also used for supplementing original regulation details and context evidence when reasoning is carried out based on the knowledge graph, and the context index layer is used for storing vector representations of text blocks so as to provide original evidence support when retrieving.
  5. 5. The utility model provides a but all-link audit's emergent field knowledge graph construction device which characterized in that includes: The conversion module is used for converting unstructured text in the emergency plan document into text blocks with continuous semantics; The grouping module is used for converting the text block conforming to the standardability judgment into a plurality of structured rule data, grouping the structured rule data and grouping the rule data with similar semantics into the same group; The acquisition module is used for acquiring entities belonging to the predefined entity types in each group of rule data based on the large language model and the preset prompt word, and the relationship between the entities is consistent with the emergency plan document, and mapping all different expressions pointing to the same entity to the same standard name; The first connection module is used for identifying isolated entities which are lack of connection with other entities based on the semantics of the emergency plan document, supplementing bridging edges for connecting the isolated entities based on a preset predicate whitelist, and forming a knowledge graph with a graph structure of a data source-entity type-entity, wherein the bridging edges are provided with evidence anchor points, and the confidence level is higher than a preset threshold value; The second connection module is used for respectively connecting the knowledge maps of the emergency plan documents with the nodes of the preset disaster chain, and the nodes in the preset disaster chain are used for representing different disaster events.
  6. 6. An electronic device, comprising: at least one memory for storing a computer program; At least one processor for executing the memory-stored program, which when executed is adapted to carry out the full-link auditable emergency domain knowledge graph construction method according to any of claims 1-4.
  7. 7. A computer readable storage medium storing a computer program, characterized in that the computer program, when run on a processor, causes the processor to perform the full-link auditable emergency domain knowledge graph construction method according to any of claims 1-4.
  8. 8. A computer program product, which when run on a processor causes the processor to perform the full-link auditable emergency domain knowledge graph construction method of any of claims 1-4.

Description

Emergency field knowledge graph construction method capable of realizing full-link audit Technical Field The application belongs to the technical field of emergency management, and particularly relates to a knowledge graph construction method for an all-link auditable emergency field. Background The existing knowledge graph construction method generally uses a natural language processing model based on deep learning to automatically extract when extracting entities and relations, firstly, a large amount of training corpus is required to be manually marked for training the model, however, a large-scale high-quality marking corpus is lacking in the emergency field, so that the supervised learning model is difficult to train or poor in effect, and the traditional model is mostly extracted at sentence level, however, long-distance dependence (such as that a main body of action is in a first section and specific execution conditions are in a third section) of a cross-paragraph and a cross-form often exists in the emergency plan, the traditional method is extremely easy to lose the cross-sentence relations, and the model only can extract a predefined Schema type, and is easy to generate a large amount of unclassified entities or generate various non-standardized redundant relation types in the face of complex and variable job descriptions in the plan, so that the graph is excessively large and inappropriate. The prior knowledge graph construction method also uses a large language model (Large Language Model, LLM) to construct, directly inputs the complete or abstracted emergency plan as background information (Context) of a Prompt word (Prompt) to the large model, and utilizes the understanding and generating capacity of the model to answer the questions about the plan by the user, however, the LLM is a probability model, and is easy to generate legal rules or incorrect emergency measures which seem reasonable but do not exist in practice, the uncontrollable illusion is a deadly potential safety hazard in the emergency field of the people's fate, and the answer directly generated by the LLM is a ' black box ' result, so that the user cannot know which version in the plan is specifically based on, and once the decision is wrong, the responsibility cannot be traced, and the interpretability is lacking. In conclusion, the emergency field knowledge graph constructed by the existing knowledge graph construction method is low in reliability. Disclosure of Invention Aiming at the defects of the prior art, the application aims to provide a full-link auditable emergency domain knowledge graph construction method, which aims to solve the problems that the existing emergency domain knowledge graph construction method is difficult to train or has poor effect, cross sentence relation in an emergency plan is easy to lose, a large number of uncategorized entities are easy to generate, and the reliability of the constructed emergency domain knowledge graph is low due to lack of interpretability. In order to achieve the above object, in a first aspect, the present application provides a method for constructing an emergency domain knowledge graph capable of being audited by a full link, including: converting unstructured text in the emergency plan document into semantically continuous text blocks; Converting the text block conforming to the normalization judgment into a plurality of structured rule data, grouping the structured rule data, and grouping the rule data with similar semantics into the same group; Acquiring entities belonging to a predefined entity type in each group of rule data based on a large language model and preset prompt words, and a relationship consistent with the emergency plan document among the entities, and mapping all different expressions pointing to the same entity to the same standard name; Identifying an isolated entity which is lack of connection with other entities based on the semantics of the emergency plan document, and supplementing a bridging edge connected with the isolated entity based on the context of the emergency plan document and a preset predicate whitelist to form a knowledge graph with a graph structure of a data source-entity type-entity, wherein the bridging edge is provided with an evidence anchor point, and the confidence level is higher than a preset threshold value; and respectively connecting the knowledge maps of the emergency plan documents with nodes of a preset disaster chain, wherein the nodes in the preset disaster chain are used for representing different disaster events. The application performs semantically maintained segmentation and grouping on the unstructured texts such as paragraphs and tables in the pre-case document, converts normalization clauses into structured rule data, can obviously reduce fragmentation and semantic breakage of the long document, ensures the completeness of clauses and improves the extraction stability of cross-sentence relations, extracts