CN-122019751-A - Knowledge index retrieval enhancement generation method based on map anchor points

CN122019751ACN 122019751 ACN122019751 ACN 122019751ACN-122019751-A

Abstract

The invention belongs to the technical field of search enhancement generation, and discloses a knowledge index search enhancement generation method based on a map anchor point. The method redefines the graph structure in the static knowledge representation as a dynamic evolution knowledge index, incrementally updates the graph structure through an iterative retrieval process, anchors key entities and relationships, generates a structured index to guide a large language model to evaluate knowledge sufficiency and constructs subsequent sub-queries. The final answer is generated by all the search documents together with the final evolution graph. Experiments in a plurality of multi-hop question-answer benchmark tests verify the effectiveness of the method and reveal that the method can more efficiently correlate key information distributed in a search document by adjusting the attention mechanism of a large language model.

Inventors

LIU ZHENGHAO
SUN BO
WU MINGYAN
GU Yu
YU MINGHE

Assignees

东北大学

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (7)

1. The method for generating the retrieval enhancement of the knowledge index based on the map anchor point is characterized by comprising the following steps: step 1, starting an iterative search process for a given initial query; In the iterative retrieval process, reasoning is carried out through a large language model, and based on the output of the previous reasoning of the large language model, the map updating, the reasoning track generation and the next query generation are simultaneously carried out, wherein the map is an evolution map and is used as a knowledge index of context perception, and key entities and relations are gradually accumulated in the iterative retrieval process; and 3, after the iterative search process meets the termination condition, aggregating all the searched documents, and generating a final answer for the initial query by combining a final evolution map.
2. The method for generating a search enhancement for knowledge index based on map anchor point according to claim 1, wherein, In the first place In the step of iterative search, the query generated according to the previous iteration Retrieving a set of related documents from a prior knowledge base : Wherein the method comprises the steps of , Representing the maximum iterative search times; Representing the use of a dense retriever to retrieve a set of documents relevant to a query from within a knowledge base.
3. The method for generating the retrieval enhancement of the knowledge index based on the map anchor point according to claim 1, wherein the iterative retrieval is specifically: According to a large language model In combination with initial queries Currently retrieved documents Output of previous iteration Updating the evolution map to obtain : Wherein, the Comprising the previous evolution pattern Intermediate reasoning results And queries ; Based on updated evolution patterns Currently retrieved documents And the previous output Generating an intermediate reasoning result of the current step And next query : Wherein the reasoning results Comprises judging the reasoning track and knowledge adequacy, if the judgment is insufficient, based on Generating And proceeds to the next iteration.
4. The method for generating the index of knowledge based on the map anchor point according to claim 3, wherein the specific process of updating the evolution map is as follows: adopting a gradual graph evolution mechanism to conduct evolution graph from Updated to In the slave direction Transition to When by at step Incorporating retrieved documents The evolution spectrum structure is gradually updated, so that a gradual evolution process is formed: Wherein the said The indexing operation is based on the previous reasoning track And queries From related documents Significant information related to the query is extracted and integrated into the evolution graph.
5. The method for generating the index of knowledge base on map anchor point according to claim 1, wherein the evolution map is The expression form of (a) is specifically as follows: defining an evolution profile Wherein As a set of entities, Converting the evolution map into a text sequence through linearization: Entities ; Relations: ; Wherein, the Is a linguistic function containing an entity and its attributes, Is an RDF triple linguistic function that includes a head entity, a relationship, and a tail entity.
6. The method for generating a search enhancement of knowledge index based on map anchor point according to claim 1, wherein, after completion of the search enhancement After iterative retrieval, if the knowledge output by the large language model in the step 2.2 is fully judged to be sufficient or the current iterative frequency reaches the maximum iterative retrieval frequency, the documents retrieved in all the steps are subjected to union aggregation to form a complete document set : 。
7. The method for generating a search enhancement for knowledge indexes based on map anchor points according to claim 6, wherein the method is based on a complete document set Generating instructions Initial query Aggregation document sets And final evolution profile According to a large language model Generating a final answer : ; Wherein the evolution profile As a structured knowledge anchor for assisting large model organization and interpretation Is provided.

Description

Knowledge index retrieval enhancement generation method based on map anchor points Technical Field The invention relates to the technical field of search enhancement generation, in particular to a knowledge index search enhancement generation method based on a map anchor point. Background The search enhancement generation model typically assists the large language model in generating accurate and dependable answers by searching for relevant paragraphs and entering them as context. However, conventional search enhancement generation methods often face the problem of insufficient knowledge acquisition, especially in a multi-hop question-and-answer scenario. In such a scenario, the inference process requires integration of evidence distributed over multiple different paragraphs, which is difficult to effectively cope with by conventional methods. To alleviate the limitations described above, prior studies have proposed iterative search strategies aimed at progressively finding and accumulating relevant knowledge to support answer generation. In addition, some methods implement more flexible and adaptive information searching by performing deep retrieval through interaction with external retrieval tools during reasoning. However, as the number of retrieval steps increases, these retrieval enhancement generation systems inevitably introduce more noise or irrelevant information. Studies have shown that too much extraneous information can significantly reduce the quality of the generated output. To improve the utilization of knowledge by large language models, existing work explores methods to reorder retrieved paragraphs or summarize salient information to filter noise. However, these methods typically process retrieved paragraphs in isolation, lacking modeling of relationships between different documents. As a result, evidence across documents may be overlooked or undercaptured, especially when complementary information from multiple sources needs to be aggregated. To better aggregate the diverse knowledge from multiple search paragraphs, more and more research has utilized structured representations (including patterns, tables, and graphs) to explicitly extract key information scattered in these documents. The partial work builds a graph to identify relevant information across multiple paragraphs or blocks, and connects extracted key entities and facts through graph structures, enabling large languages to better capture long text or long-range dependencies and multi-hop relationships in the cross-document, thereby facilitating the localization of relevant evidence. There are also studies to mitigate knowledge conflicts between internal and external information by exploring multiple inference paths on a constructed knowledge graph, identifying conflicting paths as input contexts. However, existing graph-based approaches mostly focus on static dependency capture or conflict recognition with graph structures, or tend to replace retrieved paragraphs entirely with structured representations, but lack a mechanism that can dynamically evolve with iterative retrieval processes and actively anchor key concepts to guide knowledge interpretation and sub-query generation. Therefore, how to design a dynamic indexing mechanism that can effectively anchor knowledge, reduce noise interference and support deep reasoning is still a problem to be solved in the current technology. Disclosure of Invention The invention aims to provide a knowledge index retrieval enhancement generation method based on a map anchor point, which solves the problems of difficult knowledge integration and insufficient reasoning depth in the prior art by introducing an evolution knowledge map as a knowledge anchor mechanism in an iterative retrieval enhancement generation framework. The invention converts the atlas from simple knowledge representation to an active indexing tool, and gradually updates the atlas in the iterative retrieval process to enhance the interpretation of the retrieval knowledge. By updating the atlas in each step of searching and generating the final answer when the knowledge is sufficient, the invention can more effectively organize and interpret the searched knowledge, thereby improving the reasoning accuracy of large language and the reliability of answer generation. The technical scheme of the invention is that the method for generating the retrieval enhancement of the knowledge index based on the map anchor point comprises the following steps: step 1, starting an iterative search process for a given initial query; In the iterative retrieval process, reasoning is carried out through a large language model, and based on the output of the previous reasoning of the large language model, the map updating, the reasoning track generation and the next query generation are simultaneously carried out, wherein the map is an evolution map and is used as a knowledge index of context perception, and key entities and relations are gradually