Search

CN-122019702-A - Knowledge graph retrieval method based on multi-channel recall and relationship enhancement

CN122019702ACN 122019702 ACN122019702 ACN 122019702ACN-122019702-A

Abstract

The invention provides a knowledge graph retrieval method based on multi-channel recall and relationship enhancement, belongs to the technical field of information retrieval, and provides a knowledge graph retrieval method and system integrating three-channel concurrent recall, unified scoring and split clipping, extraction relationship enhancement and double-threshold validity check, and path-level reasoning and unified sequencing. The method comprises the steps of realizing cross-channel score alignment and tracing through a unified scoring registry, stably denoising through a bit-dividing threshold value, reliably enhancing a relation through a prompt chain, label mapping and a structure template, converging node confidence, edge weight, structure regularization, time recency, service/rule priority and context relativity into a single path score through PathScore, and presenting a front end in a forest-path double view and outputting an auditable evidence list. The system adopts the pluggable design of the event bus and the SPI, and ensures the mobility and the SLA under different compliance environments.

Inventors

  • WANG HUI
  • MIAO XUDONG
  • WANG YUNNING

Assignees

  • 上海卓越睿新数码科技股份有限公司

Dates

Publication Date
20260512
Application Date
20251224

Claims (8)

  1. 1. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement is characterized by comprising the following steps of: s1, preprocessing and multi-vector embedding, namely finishing hierarchical word segmentation, synonymous mapping and syntactic maintenance to generate multi-vector; s2, three-channel concurrent recall, wherein the three-channel concurrent recall specifically comprises inverted row, node and relation concurrent recall candidates; s3, unified scoring and registration, specifically comprising the steps of calibrating and weighting the scores of all channels, synthesizing the scores into unified dimension scores, and writing the unified scoring registry; S4, dividing and cutting, namely respectively calculating dividing thresholds according to nodes and relations, and deleting low-quality long tail candidates; s5, relation enhancement, namely outputting the relation enhancement of the prompt chain after classifying the prompt chain; S6, checking the structure and the duplication removal, wherein the method specifically comprises the steps of omitting duplication according to the leaf priority policy in the UID-Pair and setting a threshold value for the root/trunk/She Fenceng according to the checking direction, depth and loop of the structure template; s7, composition and path reasoning, specifically comprising constructing a graph by using effective edges, constructing a forest by using a BFS unit, enumerating paths by using a DFS unit, and unifying path sequencing in a local domain by using a unified path score; s8, outputting and auditing, specifically comprising generating a route map and an evidence list, and recording evolution from the channel score to the route score; and S9, deployment and operation, namely adopting event bus driven stateless micro-services, and packaging hot switching of inverted row, vector, graph and LLM by SPI.
  2. 2. The method for retrieving a knowledge-graph based on multi-channel recall and relationship enhancement according to claim 1, wherein in step S3, the unified scoring registry comprises a unified dimension score of record candidates, source channel, weight, time stamp, evidence pointer, domain identifier and context gating factor, and provides a hierarchical statistics process.
  3. 3. The knowledge graph retrieval method based on multi-channel recall and relationship enhancement according to claim 2, wherein in the step S3, the method further comprises normalizing or equivalently calibrating the scores of different retrieval channels to make the scores fall into a unified number axis, and updating parameters by combining with a history statistics or verification set.
  4. 4. The knowledge-graph retrieval method based on multi-channel recall and relationship enhancement according to claim 3, wherein in the step S4, the quantile threshold value comprises a quantile threshold value calculated according to the node and the relationship respectively.
  5. 5. The method for retrieving a knowledge-graph based on multi-channel recall and relationship enhancement according to claim 4, wherein in the step S6, the structural templates include a rule set of relationship direction, maximum depth, loop constraint and hierarchy compatibility matrix.
  6. 6. The method for retrieving a knowledge-graph based on multi-channel recall and relationship enhancement according to claim 5, wherein the step S7 of unifying path scores further comprises synthesizing factors into a single scalar for unifying ordering of different length and component paths, the factors comprising node confidence, edge weight, structural regularities, temporal recency, regulatory priority, contextual relevance.
  7. 7. The method for retrieving a knowledge-graph based on multi-channel recall and relationship enhancement according to claim 6, wherein in step S5, the method further comprises, L1 master node alignment; L2 name endpoint extensions including aliases, abbreviations, synonyms; L3 vector neighbor contrast; l4 evidence constraint extraction, which comprises extracting triples and slots from evidence segments; L5 conflict resolution comprises label mapping, homonymy disambiguation and upper and lower merging.
  8. 8. The multi-channel recall and relationship-enhanced knowledge-graph retrieval method of claim 7 wherein, in step S5, the hint-chain relationship-enhanced output comprises triplets, confidence, evidence pointers, and time stamps.

Description

Knowledge graph retrieval method based on multi-channel recall and relationship enhancement Technical Field The specification relates to the technical field of information retrieval, in particular to a knowledge graph retrieval method based on multi-channel recall and relationship enhancement. Background At present, when a large language model of a search enhancement generation system supported by a knowledge graph is used for carrying out a question-answer reasoning task of natural language, noise or irrelevant paths exist in the searched graph information, so that the accuracy of answer generation of the large language model in the question-answer reasoning task is reduced. The prior Chinese patent, publication No. CN120950652A, discloses a video retrieval method for matching text information, which comprises the steps of responding to the received natural language question information, obtaining a corresponding candidate path set by utilizing a knowledge graph retrieval system, carrying out coarse filtration and fine filtration on the candidate path set by utilizing a large language model to obtain a coarse filtration path set and a target path set, constructing a structure prompting instruction, respectively obtaining a first answer result output by the knowledge graph retrieval system and a second answer result output by the large language model, calculating confidence coefficient values corresponding to the first answer result and the second answer result, screening and correspondingly fusing the first answer result and the second answer result according to the confidence coefficient values, and generating a target answer result corresponding to the natural language question information. The traditional inverted row and vector fine row has the problems of non-unification of cross-channel dimension, fragmentation of results, lack of path level explanation, difficulty in rechecking under compliance requirements and the like on heterogeneous and weak connection data, and is difficult to consider between coverage rate and precision, interpretation and time delay in high concurrency and multi-compliance scenes. Disclosure of Invention The invention provides a knowledge graph retrieval method based on multi-channel recall and relation enhancement, which provides three-channel (inverted row+node vector+relation vector) concurrent recall, realizes cross-channel score calibration (channel calibration function), domain self-adaptive weighting (source weight) and posterior robustness processing through a unified scoring registry (NSR), unifies dimensions, solves the problem of unstable rearrangement, and simultaneously reserves score sources and evidence pointers to lay a foundation for subsequent interpretability. In some embodiments, the method comprises the steps of: s1, preprocessing and multi-vector embedding, namely finishing hierarchical word segmentation, synonymous mapping and syntactic maintenance to generate multi-vector; s2, three-channel concurrent recall, wherein the three-channel concurrent recall specifically comprises inverted row, node and relation concurrent recall candidates; s3, unified scoring and registration, specifically comprising the steps of calibrating and weighting the scores of all channels, synthesizing the scores into unified dimension scores, and writing the unified scoring registry; S4, dividing and cutting, namely respectively calculating dividing thresholds according to nodes and relations, and deleting low-quality long tail candidates; s5, relation enhancement, namely outputting the relation enhancement of the prompt chain after classifying the prompt chain; S6, checking the structure and the duplication removal, wherein the method specifically comprises the steps of omitting duplication according to the leaf priority policy in the UID-Pair and setting a threshold value for the root/trunk/She Fenceng according to the checking direction, depth and loop of the structure template; s7, composition and path reasoning, specifically comprising constructing a graph by using effective edges, constructing a forest by using a BFS unit, enumerating paths by using a DFS unit, and unifying path sequencing in a local domain by using a unified path score; s8, outputting and auditing, specifically comprising generating a route map and an evidence list, and recording evolution from the channel score to the route score; and S9, deployment and operation, namely adopting event bus driven stateless micro-services, and packaging hot switching of inverted row, vector, graph and LLM by SPI. By the technical means, the traditional search only returns a keyword matching result list, and the technology returns a path-based result (such as a risk path of 'enterprise A, guarantee, enterprise B, debt default') with logic relation, evidence chain and auditable, and the application solves the problem of accurately and interpretably finding the associated information and logic path from massive heterogeneous