CN-121996700-A - Complex system decision support method and device based on hybrid search enhancement generation

CN121996700ACN 121996700 ACN121996700 ACN 121996700ACN-121996700-A

Abstract

The invention belongs to the field of artificial intelligence auxiliary decision making, and particularly relates to a complex system decision support method and device based on hybrid search enhancement generation. According to the method, a knowledge graph is built aiming at documents in the field of complex systems, semantic embedded vectors of all text blocks are generated, double search channels of a vector search channel and a graph search channel are designed, parallel search is conducted on user inquiry, the vector search channel calculates semantic similarity of the user inquiry and all the text blocks according to the semantic embedded vectors of all the text blocks, the graph search channel conducts graph traversal on the knowledge graph, matching degree scores of all the text blocks are calculated on the basis of the traversal results, sorting results of the search results in the two search channels are obtained according to the similarity or the scores, fusion scores of all the text blocks in the search results are calculated by means of a reciprocal rank fusion algorithm, fusion results are obtained, decision suggestions are generated on the basis of the fusion results, and accuracy and reliability of emergency decision support of the complex system are improved.

Inventors

XU MINGLIANG
ZHANG BO
FAN MINGJIE
JIN ZHAO
CHEN DONG

Assignees

郑州大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. A complex system decision support method based on mixed retrieval enhancement generation is characterized by comprising the following steps: Dividing text blocks of a document in the field of complex systems, generating semantic embedded vectors of the text blocks, and constructing a knowledge graph based on entities and relations in the text blocks; The method comprises the steps of inputting a user query into a vector search channel and a graph search channel in parallel to obtain search results, wherein the vector search channel is used for calculating semantic similarity of the user query and all text blocks according to semantic embedded vectors of the text blocks, and selecting the first K v text blocks as the vector search results after sorting according to the semantic similarity descending order; According to the sequencing results of each text block in the two retrieval channels in the retrieval results, calculating the fusion score of each text block in the retrieval results by adopting a reciprocal rank fusion algorithm to obtain fusion results; and generating a decision suggestion based on the fusion result.
2. The complex system decision support method based on hybrid search enhancement generation of claim 1, wherein the traversal result comprises a set of seed entities identified in a knowledge-graph with entities determined from a user query and an extended set of entities from the seed entities that are co-community extended and multi-hop neighbor extended.
3. The complex system decision support method based on hybrid search enhancement generation of claim 2, wherein the matching score of each text block is determined according to the seed entity matching degree of the text block and the seed entity set, the extended entity matching degree of the text block and the extended entity set, and the semantic similarity of the text block and the user query.
4. The complex system decision support method based on hybrid search enhancement generation according to claim 1, wherein when a decision suggestion is generated based on a fusion result, firstly performing de-duplication processing on text blocks with semantic similarity greater than or equal to a first preset threshold in the fusion result, and generating the decision suggestion based on the structural evidence by taking the text blocks after the de-duplication processing as the structural evidence.
5. The complex system decision support method based on hybrid search enhancement generation according to claim 1, wherein when a decision suggestion is generated based on a fusion result, firstly performing de-duplication processing on text blocks with semantic similarity greater than or equal to a first preset threshold value in a fusion sorting list, selecting a first maximum number of text blocks with highest fusion score to perform filtering processing when the number of the text blocks after the de-duplication processing is greater than the maximum number, and using the text blocks after the de-duplication and filtering processing as structural evidence to generate the decision suggestion based on the structural evidence.
6. The complex system decision support method based on hybrid search enhancement generation according to claim 5, wherein when generating decision suggestions based on the fusion result, firstly performing deduplication processing on text blocks with semantic similarity greater than or equal to a first preset threshold in the fusion sorting list, selecting a first maximum number of text blocks with highest fusion score for filtering when the number of text blocks after deduplication processing is greater than the maximum number, adding meta information to each text block after deduplication and filtering processing as structural evidence, and generating the decision suggestions based on the structural evidence, wherein the meta information comprises a source chapter, a page number, a sorting result in a vector retrieval channel and a sorting result in a graph retrieval channel.
7. The complex system decision support method based on hybrid search enhancement generation according to claim 2, wherein communities in the same community extension from seed entities are knowledge modules that use community detection algorithms to modularize knowledge graphs and use large language models to generate community summary descriptions.
8. The complex system decision support method based on hybrid search enhancement generation according to claim 1 is characterized in that the entity in the process of constructing a knowledge graph is obtained by performing entity disambiguation by adopting a strategy combining rule matching and semantic similarity, wherein the strategy is that if the names of the entities are consistent, the entity is directly judged to be the same entity, if the names of the entities are not completely consistent, cosine similarity of description vectors corresponding to the two entities is calculated, and when the cosine similarity exceeds a second preset threshold, the entity is merged into the same entity.
9. The complex system decision support method based on hybrid search enhancement generation of any of claims 4-6, wherein generating decision suggestions based on structured evidence is by inputting structured evidence and user queries into a large language model, and generating decision suggestions using the large language model.
10. A complex system decision support device based on hybrid search enhancement generation, comprising a processor, wherein the processor is configured to implement the complex system decision support method based on hybrid search enhancement generation of any one of claims 1-9.

Description

Complex system decision support method and device based on hybrid search enhancement generation Technical Field The invention belongs to the field of artificial intelligence auxiliary decision making, and particularly relates to a complex system decision support method and device based on hybrid search enhancement generation. Background The emergency decision scene of the complex system has the characteristics of sparse events, difficult prediction and serious consequences, and has very limited ways for accumulating decision experience through real cases. Complex systems typically involve interactions, nonlinear causality and dynamic evolution processes of multiple subsystems, requiring decision-making systems capable of complex causal reasoning and multivariate trade-offs based on expertise in very short time. In recent years, a large language model (Large Language Model, LLM) has strong capability in text understanding, logical reasoning, knowledge generation and other aspects, and provides a new idea for solving the problem of intensive emergency decision. However, the general large language model often faces the problems of 1) knowledge illusion that the model may generate professional content that looks reasonable but actually wrong, 2) knowledge outgrowth that the timeliness limit of the pre-training data results in the inability to acquire the latest domain knowledge, and 3) lack of domain depth that the professional domain knowledge duty ratio in the general training data is limited, and difficulty in supporting deep reasoning when dealing with highly specialized domains. The search enhancement generation (RETRIEVAL-Augmented Generation, RAG) technique can effectively alleviate the above-described problems by searching for relevant external knowledge prior to decision generation. The existing RAG method mainly comprises two types, namely a vector retrieval-based method, wherein query and documents are encoded into points in a high-dimensional vector space, nearest neighbor search is carried out by utilizing metrics such as cosine similarity, the method is successful in tasks such as open-domain question answering and fact checking, but has the following limitations that the method only depends on semantic similarity, is difficult to capture causal relations among knowledge, cannot utilize structured relations of knowledge to carry out multi-hop reasoning, and knowledge which is far in semantic distance but is related with logic strength is easy to miss. The second category is a method based on knowledge graph, for example, chinese patent application document with publication number CN120929785A discloses a building elevator detection diagnosis decision method based on graph retrieval enhancement agent, which utilizes graph retrieval enhancement generation to carry out multidimensional retrieval and generates answers with reasoning paths based on the retrieved sub-graph structure. The method based on the knowledge graph explicitly represents the entity and the relation as the node and the edge, can characterize the structural characteristics of the knowledge, and supports multi-hop reasoning based on the relation path. However, a single graph retrieval strategy has the defects that the graph retrieval is possibly omitted for the knowledge of semantic similarity but lack of explicit relation links, the graph construction quality directly influences the retrieval effect, errors of entity disambiguation and relation extraction can be propagated to the downstream, and the flexibility and the diversity of natural language expression are difficult to process. In summary, single search enhancement generation strategy knowledge coverage is incomplete, affecting decision accuracy. When the vector retrieval enhancement and the graph retrieval enhancement are directly weighted and fused, the simple weighting is easily led by a certain channel due to inconsistent scoring dimension of different retrieval channels, and the fusion effect is poor. Therefore, the existing decision methods all result in lower decision accuracy. Disclosure of Invention The invention aims to provide a complex system decision support method and device based on mixed retrieval enhancement generation, which are used for solving the problem of low complex system decision accuracy caused by insufficient knowledge coverage of a single retrieval enhancement strategy or poor multi-source fusion effect in the prior art. The invention provides a complex system decision support method based on mixed search enhancement generation, which aims to solve the technical problems and comprises the steps of conducting text block division on a complex system field document, generating semantic embedded vectors of text blocks, constructing a knowledge graph based on entities and relations in the text blocks, enabling a user to inquire parallel input vector search channels and graph search channels to obtain search results, enabling the vector search channels