CN-121981257-A - Artificial intelligence model training and reasoning management system based on knowledge graph

CN121981257ACN 121981257 ACN121981257 ACN 121981257ACN-121981257-A

Abstract

The invention discloses an artificial intelligent model training and reasoning management system based on a knowledge graph, which comprises a data analysis module, a graph construction module, a context mapping module, a problem analysis module, a reasoning chain generation module, a graph reasoning module and a feedback updating module, wherein the data analysis module is used for collecting multisource perception data and forming an entity set and an event set, the graph construction module is used for constructing a context knowledge graph based on the entity set and the event set, the context mapping module is used for establishing a reversible mapping relation between the context knowledge graph and original multisource perception data, the problem analysis module is used for carrying out semantic analysis on a problem input by a user and cutting the problem to obtain a context Wen Zitu, the reasoning chain generation module is used for generating a reasoning chain meeting time sequence consistency and causal consistency, the graph reasoning module is used for generating graph reasoning embedding and combined question-answer generation models to output question-answer results based on an improved T-GNN model, and the feedback updating module is used for updating context knowledge graphs and model parameters. The invention realizes the structured modeling and credibility closed-loop management of the question-answer reasoning process.

Inventors

CAO HANJUN
HUA YE
HAN XIAOXIAO
SHEN TONG
YUAN LEI
CHEN YIMING
ZHOU QIANG
Xin Youshun

Assignees

中国烟草总公司安徽省公司

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. An artificial intelligence model training and reasoning management system based on a knowledge graph is characterized by comprising the following modules: the data analysis module is used for collecting multi-source perception data related to the question-answering task to form an entity set and an event set; The map construction module is used for constructing a context knowledge map based on the entity set and the event set; The context mapping module is used for establishing a reversible mapping relation between the context knowledge graph and the original multi-source perception data; The problem analysis module is used for carrying out semantic analysis on the problem input by the user, resolving the problem into a plurality of semantic intention components, mapping the semantic intention components to corresponding nodes, relations or sub-graph structures in the contextual knowledge graph, and executing up-down Wen Zitu clipping to obtain up-down Wen Zitu; The inference chain generation module is used for executing path search under time sequence constraint in the context subgraph and generating an inference chain meeting the consistency of time sequence and causal direction; The graph reasoning module is used for inputting the reasoning chain and the corresponding context subgraph into the improved T-GNN model to generate graph reasoning embedding, and inputting the graph reasoning embedding and the user problem together into the question and answer generating model to generate a question and answer result; And the feedback updating module is used for collecting feedback information of the user on the question and answer result, forming an inference reliability index, and updating the context knowledge graph and the improved T-GNN model parameters according to the inference reliability index.
2. The knowledge-based artificial intelligence model training and reasoning management system according to claim 1, wherein the modules are realized by the following method: Firstly, acquiring multisource perception data related to questions and answers, executing semantic analysis to obtain entity, event and attribute information, and performing time labeling to form an entity set and an event set; step two, constructing a context knowledge graph based on the entity set and the event set; Step three, configuring context state attributes for nodes and relationship edges in the context knowledge graph, setting context anchor point identifiers, and establishing a reversible mapping relationship between the context knowledge graph and original multi-source perception data; performing semantic analysis on the problem input by the user, constructing the problem into a plurality of semantic intention components, mapping the semantic intention components to corresponding nodes, relations or sub-graph structures in the context knowledge graph, and performing up-down Wen Zitu clipping to obtain up-down Wen Zitu; Step five, path searching under time sequence constraint is executed in the context subgraph, and an inference chain meeting the time sequence consistency and the causal direction consistency is generated; Inputting an inference chain and a corresponding context subgraph into an improved T-GNN model, generating graph inference embedding based on a message transfer clipping mechanism of the inference chain constraint, a time weight modulation mechanism of the context perception and a staged time sequence inference mechanism driven by a problem target, and inputting the graph inference embedding and a user problem together into a question and answer generation model to generate a question and answer result; And step seven, collecting feedback information of the user on the question and answer result to form an inference reliability index, and updating a context knowledge graph and improved T-GNN model parameters according to the inference reliability index.
3. The knowledge-based artificial intelligence model training and reasoning management system of claim 2, wherein the first step specifically comprises: collecting multi-source perception data related to questions and answers, wherein the multi-source perception data comprises question text, voice transcription text, image or video frame data and corresponding structured business data which are input by a user; performing word segmentation processing on the problem text and the voice transcription text by adopting a positive maximum matching word segmentation method to obtain a word sequence, performing named entity recognition on the word sequence by adopting a sequence labeling method of a hidden Markov model to obtain entity candidate items, and simultaneously triggering a word list to recognize event candidate items through a matching event; Performing target detection on the image or video frame data by adopting YOLOv target detection models, outputting target class labels and corresponding boundary frame information, and completing cross-modal alignment according to the type corresponding relation between the target class labels and entity candidates; Extracting attribute information from the structured business data according to the mapping relation between the field names and the entity types, and associating the attribute information to corresponding entity candidate items; And taking the time stamp in the structured business data as reference time, and marking the time of the entity candidate item and the event candidate item to form an entity set and an event set containing the time mark.
4. The system for training and reasoning about an artificial intelligence model based on a knowledge graph as set forth in claim 2, wherein the second step comprises: Mapping each entity in the entity set into a graph node, and mapping each event in the event set into an event node; establishing a relationship edge between corresponding nodes according to the co-occurrence relationship, the time sequence relationship and the semantic association relationship of the entity and the event in the multi-source perception data; type marking and direction marking are carried out on the relation edges; And when the new multi-source perception data arrives, the context knowledge graph is updated in an increment mode according to the newly generated entity set and the event set.
5. The knowledge-based artificial intelligence model training and reasoning management system of claim 2, wherein the third step specifically comprises: configuring a context state attribute for each node in a context knowledge graph, wherein the context state attribute comprises a source identifier, a time annotation identifier and a data type identifier; Configuring corresponding context state attributes for each relationship side in a context knowledge graph, wherein the context state attributes of the relationship side comprise a relationship type identifier and a direction identifier; Setting a unique context anchor point identifier for each node and each relation edge in the context knowledge graph; establishing a one-to-one correspondence between nodes and relationship edges in the context knowledge graph and corresponding problem text, voice transcription text, image or video frame data and corresponding structured business data in the original multi-source perception data through the context anchor point identification to form a reversible mapping relationship between the context knowledge graph and the original multi-source perception data; and when the context knowledge graph is updated, synchronously updating the corresponding relation between the context anchor point identification and the original multi-source perception data.
6. The knowledge-based artificial intelligence model training and reasoning management system as set forth in claim 2, wherein the fourth step comprises: Carrying out semantic analysis on a problem input by a user, carrying out word segmentation processing by adopting a forward maximum matching word segmentation method, carrying out named entity recognition by adopting a sequence labeling method of a hidden Markov model to obtain entity terms and event terms in the problem, carrying out semantic intention recognition on the problem based on a preset intention keyword list, and forming a plurality of semantic intention components; Respectively mapping semantic intention components to nodes and relation edges in a context knowledge graph, taking the nodes and relation edges which finish mapping as starting points, and acquiring adjacent nodes and relation edges which are directly connected with the starting points from the context knowledge graph to form candidate contexts Wen Zitu; Performing clipping processing on the candidate context subgraphs, reserving nodes and relationship edges matched with semantic intention components on entity types, event types or relationship types during clipping, and removing unmatched nodes and relationship edges to obtain a context Wen Zitu; the method comprises the steps of carrying out context semantic coding on a problem input by a user based on a transducer encoder, outputting context expression vectors corresponding to words in a problem text, carrying out vector aggregation processing on the context expression vectors, and obtaining semantic embedded vectors of the user problem by carrying out average pooling operation.
7. The knowledge-based artificial intelligence model training and reasoning management system of claim 2, wherein the fifth step specifically comprises: Reading time marking information of each node in the context subgraph, and relation type, direction identification and time marking information of each relation edge, and constructing a time sequence constraint set and a causal direction constraint set; selecting a node obtained by mapping with a user problem from the context subgraph as an initial node set, and generating a candidate path by adopting a layered expansion path searching mode; After each expansion generates a new node, executing time sequence consistency check, and comparing the time label of the new node with the time label of the end node of the candidate path when the time sequence consistency check is executed, and only reserving the candidate path branches which meet the condition that the time label of the new node is not earlier than the time label of the end node; After each expansion generates a new relationship edge, executing causal direction consistency check, and when the relationship type of the new relationship edge is determined to be causal when the causal direction consistency check is performed, keeping the corresponding direction identification consistent with the existing direction of the candidate path, and only keeping the candidate path branches passing the causal direction consistency check; Performing repeated node elimination processing on the candidate path branches, eliminating the candidate path branches containing repeated nodes, setting a maximum path length threshold value for the candidate path branches, and stopping expansion of the candidate path branches exceeding the maximum path length threshold value; And solidifying the candidate path branches into an inference chain when the candidate path branches meet the preset termination condition.
8. The knowledge-graph-based artificial intelligence model training and reasoning management system of claim 2, wherein the improved T-GNN model comprises an input construction layer, a time coding layer, a message passing layer, a time weight modulation layer, a node feature update layer, a staged timing reasoning layer, and a graph reasoning embedding generation layer: The input construction layer represents each node in the context subgraph as a node input feature vector, represents a relationship edge in the context subgraph as a relationship edge input feature vector, and constructs the sequential relationship of the nodes and the relationship edge in the inference chain as an inference chain structure mask; The Time coding layer codes Time labeling values in the node input feature vectors into Time embedding vectors by adopting a Time2Vec Time embedding encoder to form node initial Time sequence feature representations containing Time information; The message transfer layer uses a neighborhood message transfer calculation mode of a T-GNN model to perform message propagation processing on node initial time sequence characteristic representation, introduces a message transfer clipping mechanism of inference chain constraint, and limits nodes to only receive message input which comes from a current node in a front node in an inference chain and has consistent relation edge direction identification through an inference chain structure mask, wherein the nodes outside the inference chain or the messages corresponding to relation edges which do not meet the direction identification consistency constraint do not participate in message propagation; the time weight modulation layer receives the message vector output by the message transfer layer, introduces a context-aware time weight modulation mechanism, and generates a time weight coefficient by inputting a time annotation difference value between a message source node and a target node and a context state attribute vector of the target node into two layers of full-connection mapping functions for mapping, and normalizing a mapping result through a Softmax function; The node characteristic updating layer adopts a weighted summation aggregation mode based on attention weight, takes a time weight coefficient as attention weight, performs weighted summation on message vectors from different preamble nodes to obtain an aggregate message vector, and inputs the aggregate message vector and a previous round of time sequence characteristic representation of the node to the GRU gating updating unit together to generate a new round of time sequence characteristic representation of the node; The method comprises the steps that a problem target driven phased timing sequence reasoning mechanism is introduced by a phased timing sequence reasoning layer, first-stage timing sequence reasoning is executed based on a context subgraph, multiple rounds of integral timing sequence feature updating are executed on all nodes participating in reasoning in the context subgraph according to the processing flow of a message transmission layer, a time weight modulation layer and a node feature updating layer, node timing sequence representation is obtained, and second-stage timing sequence reasoning is executed according to a reasoning chain corresponding to a user problem on the basis of a first-stage timing sequence reasoning result; The graph inference embedding generation layer performs aggregation processing on node time sequence characteristic representations corresponding to the end nodes of the inference chain output by the hierarchical time sequence inference layer, and the aggregation processing generates graph inference embedding in a weighted summation mode based on attention weights.
9. The knowledge-graph-based artificial intelligence model training and reasoning management system of claim 2, wherein the question-answering generation model adopts a GLM autoregressive transducer decoding language model, receives semantic embedded vectors of graph reasoning embedding and user questions, and generates question-answering results.
10. The knowledge-based artificial intelligence model training and reasoning management system of claim 2, wherein the seventh step specifically comprises: constructing an inference credibility index based on graph inference embedding and question-answer results; when the inference reliability index is lower than a preset reliability threshold, triggering updating processing, executing context state attribute adjustment or structure supplementation on nodes and relationship sides corresponding to the current inference chain in the context knowledge graph, and executing parameter updating on the improved T-GNN model based on the updated context knowledge graph.

Description

Artificial intelligence model training and reasoning management system based on knowledge graph Technical Field The invention relates to the technical field of artificial intelligence and knowledge engineering, in particular to an artificial intelligence model training and reasoning management system based on a knowledge graph. Background With the continuous development of artificial intelligence technology in the fields of intelligent question-answering, knowledge management, auxiliary decision making and the like, knowledge reasoning and question-answering systems based on multi-source data fusion are receiving a great deal of attention. In the prior art, text, voice, image and structured business data are collected to analyze user questions, and a question and answer result is generated by combining a knowledge base or a knowledge map so as to improve information acquisition and decision support capability. However, in practical applications, there are still a number of disadvantages in the related art: On one hand, the existing system is dependent on a simple feature extraction or unified coding mode for processing multi-source perception data, lacks of system modeling on entities, events and time relations of the entities, events and the time relations of the events, is difficult to align effectively due to the fact that differences of different mode data in acquisition time, semantic granularity and structural forms are difficult to align, the consistency of contexts is insufficient, the stability of reasoning results is poor, meanwhile, most systems do not establish traceable relations between knowledge maps and original data, and the reasoning process is difficult to verify and trace. In addition, the coupling degree between the graph reasoning model and the question-answer generating model is low, the question-answer generating process is difficult to be effectively guided by a reasoning result, and a feedback updating mechanism based on reasoning credibility is generally lacking, so that the knowledge graph and model parameters cannot be continuously optimized along with the using process, and the long-term applicability and intelligent evolution capability of the system are limited. Therefore, how to provide an artificial intelligence model training and reasoning management system based on a knowledge graph is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide an artificial intelligent model training and reasoning management system based on a knowledge graph, which fully combines multisource perception data analysis, context knowledge graph construction and reversible mapping, semantic intention driven causal reasoning chain generation and an improved T-GNN model graph reasoning method, and details the whole process from user problem analysis to graph reasoning embedding generation, question-answer output and feedback closed loop updating, and has the advantages of clear reasoning structure, accurate semantic alignment, reliable and controllable results and strong system continuous evolution capability. According to the embodiment of the invention, the artificial intelligent model training and reasoning management system based on the knowledge graph comprises the following modules: the data analysis module is used for collecting multi-source perception data related to the question-answering task to form an entity set and an event set; The map construction module is used for constructing a context knowledge map based on the entity set and the event set; The context mapping module is used for establishing a reversible mapping relation between the context knowledge graph and the original multi-source perception data; The problem analysis module is used for carrying out semantic analysis on the problem input by the user, resolving the problem into a plurality of semantic intention components, mapping the semantic intention components to corresponding nodes, relations or sub-graph structures in the contextual knowledge graph, and executing up-down Wen Zitu clipping to obtain up-down Wen Zitu; The inference chain generation module is used for executing path search under time sequence constraint in the context subgraph and generating an inference chain meeting the consistency of time sequence and causal direction; The graph reasoning module is used for inputting the reasoning chain and the corresponding context subgraph into the improved T-GNN model to generate graph reasoning embedding, and inputting the graph reasoning embedding and the user problem together into the question and answer generating model to generate a question and answer result; And the feedback updating module is used for collecting feedback information of the user on the question and answer result, forming an inference reliability index, and updating the context knowledge graph and the improved T-GNN model parameters according to the inference reliability index. According to the