An intelligent question and answer processing method and system based on a knowledge graph

Abstract

This invention discloses an intelligent question-answering method and system based on knowledge graphs, comprising the following steps: Step 1: Obtaining user questions and generating question-answering constraint units; Step 2: Locating the starting entity and candidate answer range, and extracting candidate question-answer subgraphs; Step 3: Configuring node semantic potential labels and relation edge evidence labels; Step 4: Performing edge weight calibration and generating a semantic recalibration graph; Step 5: Performing evidence role mapping and constructing an evidence role residual graph; Step 6: Performing dual-evidence path search and generating a candidate evidence path set; Step 7: Calculating path consistency residuals and updating the evidence role residual graph; Step 8: Determining the target evidence chain and generating answer text and evidence paths. This invention improves the accuracy, evidence independence, and path interpretability of intelligent question-answering results.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to an intelligent question-answering method and system based on knowledge graphs. Background Technology

[0002] With the increasing scale of knowledge graph applications and the growing demand for intelligent question answering, semantic parsing, graph retrieval, and evidence path reasoning technologies for complex business knowledge have received widespread attention. Existing knowledge graph question answering methods mainly rely on entity recognition, relation matching, path search, or query transformation to generate answers, but they generally suffer from the following problems in practical applications: User queries often contain ambiguous entity representations, relational orientations, and answer ranges. Existing semantic parsing results lack stable constraint mappings with knowledge graph nodes and relational edges, easily leading to initial entity positioning deviations and excessively broad candidate answer ranges. Traditional path search methods typically filter candidate paths based on graph structure distance or simple relation similarity, making it difficult to distinguish between common semantic nodes and real evidence sources. This can result in different path forms but duplicate evidence sources, affecting the independence of answer basis. For multi-hop question-answering scenarios, existing candidate path filtering technologies lack closed-loop verification between paths and question slots. The answer entities, relational directions, and limiting conditions in the path may be inconsistent with the user query, resulting in generated answer text that, while able to connect to graph nodes, lacks sufficient evidence path support, leading to poor accuracy and interpretability of the question-answering results.

[0003] Therefore, how to provide an intelligent question-answering processing method and system based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent question-answering processing method and system based on knowledge graphs. This invention generates question-answering constraint units through semantic parsing and slot alignment, and constructs an evidence role residual graph by combining node semantic potential energy labeling, relation edge evidence labeling, and an improved Suurballe algorithm. It completes dual-evidence path search, path consistency residual update, and target evidence chain determination in the knowledge graph, realizing the synchronous output of answer text and evidence path. It has the advantages of accurate answer positioning, independent evidence source, clear reasoning path, and strong interpretability of question-answering results.

[0005] According to an embodiment of the present invention, an intelligent question-answering method based on a knowledge graph includes the following steps: Step 1: Obtain the user's question, perform semantic parsing and slot alignment on the user's question, and generate question-answer constraint units; Step 2: Based on the question-answering constraint unit, locate the starting entity and the range of candidate answers in the knowledge graph, and extract the candidate question-answer subgraph covering the range of candidate answers around the starting entity; Step 3: Configure node semantic potential labels and relation edge evidence labels for the candidate question-answering subgraph based on the question-answering constraint unit; Step 4: Perform edge weight labeling on the candidate question-answering subgraph based on node semantic potential labeling and relation edge evidence labeling to generate a semantic relabeling graph; Step 5: Perform evidence role mapping on the nodes and relation edges in the semantic recalibration graph, generate evidence role mapping results, and construct the evidence role residual graph of the improved Suurballe algorithm based on the evidence role mapping results; Step 6: In the residual graph of evidence roles, perform a dual evidence path search from the starting entity to the range of candidate answers to generate a set of candidate evidence paths; Step 7: Semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual map based on the path consistency residual. Step 8: Determine the target evidence chain based on the updated evidence role residual map, generate the answer text and evidence path according to the target evidence chain, and output the intelligent question answering processing result.

[0006] Optionally, step one specifically includes: The system acquires user questions and generates question identifiers, associates user questions with question identifiers, performs character format standardization, statement boundary determination, and invalid character cleanup on user questions, and generates well-organized question text. Perform term segmentation and domain term normalization on regular question text to generate question term sequence. Based on the question term sequence, perform semantic dependency analysis and domain semantic mapping to generate semantic component record. Based on the semantic component record, entity pointing components and relation pointing components are extracted to form the question semantic parsing result. Based on the question semantic parsing result, the answer range pointing and constraint condition pointing are determined to generate question slot record; The semantic parsing results of the questions are associated and organized with the question slot records according to the question identifier, and the constraint unit identifier is configured to generate the question-answer constraint unit.

[0007] Optionally, step two specifically includes: Extract entity pointing components, answer range pointing and constraint condition pointing from the question and answer constraint unit, retrieve the knowledge graph composed of graph entity nodes, graph relation edges and graph attribute fields, perform text normalization and retrieval key generation on the entity pointing components, match the retrieval key with the node name field and node category field of the graph entity node to form entity matching records; Extract node name matching values ​​and node category matching values ​​from entity matching records, perform weighted aggregation on node name matching values ​​and node category matching values ​​to generate entity matching confidence values, determine the graph entity nodes whose entity matching confidence values ​​reach the entity threshold as the starting entities, and associate the starting entities with question answering constraint units; The answer range is parsed to form an answer category marker. The answer category marker is matched with the node category field of the graph entity node to form the answer category boundary. Then, the constraint condition is used to limit the value range of the graph attribute field and the connection range of the graph relationship edge to generate the candidate answer range. Starting from the initial entity, perform hierarchical restricted expansion along the graph relation edges in the knowledge graph, record the expansion level, the graph entity nodes reached, and the graph relation edges passed through. Collect the graph entity nodes that are within the range of candidate answers and the graph relation edges passed through to generate a candidate question-and-answer subgraph that covers the range of candidate answers.

[0008] Optionally, step three specifically includes: Extract the semantic parsing results of questions and the question slot records from the question-answering constraint unit, extract the graph entity nodes and graph relation edges from the candidate question-answering subgraph, map the semantic parsing results of questions to the graph entity nodes, and generate node semantic correspondence records; According to the node semantic correspondence record, the semantic correspondence state between the graph entity node and the question semantic parsing result is quantified to generate the node semantic potential value, and the node semantic potential value is written into the graph entity node to form the node semantic potential value tag. Map the question slot records to the graph relation edges to generate corresponding records for relation edge evidence; Based on the corresponding records of relation edge evidence, the evidence support status between the graph relation edge and the question slot record is quantified, relation edge evidence value is generated, and the relation edge evidence value is written into the graph relation edge to form relation edge evidence marker.

[0009] Optionally, step four specifically includes: Extract graph relation edges from candidate question-and-answer subgraphs, determine the starting graph entity nodes and the ending graph entity nodes connected by the graph relation edges, extract the starting semantic potential value and the ending semantic potential value from the node semantic potential energy markers, extract the relation edge evidence value from the relation edge evidence markers, and generate edge weight labeling records. Perform difference calculation on the initial semantic potential value and the terminal semantic potential value in the edge weight labeling record to generate the node potential offset value; perform evidence cost transformation on the relation edge evidence value to generate the evidence support cost value. Extract the basic connection cost from the graph relation edges, and perform a fusion calculation on the basic connection cost, node potential offset value and evidence support cost to generate semantic relabeling edge weights; Write the semantic relabeling edge weights into the corresponding graph relation edges, and reconstruct the candidate question-answer subgraph according to the connection relationship between the graph entity nodes, graph relation edges and semantic relabeling edge weights to generate the semantic relabeling graph.

[0010] Optionally, step five specifically includes: Extract graph entity nodes, graph relation edges, and semantic relabeling edge weights from the semantic relabeling graph. Associate the semantic carrying relationship of the graph entity nodes in the question-answering constraint unit, the range belonging relationship in the candidate answer range, and the connection position in the semantic relabeling graph to generate node evidence role records. The graph relation edges are associated with the slot support relationship of the question-and-answer constraint unit, the connection pointing relationship of the candidate answer range, and the semantic relabeling edge weight associated with the graph relation edges to generate relation edge evidence role records; Write the node evidence role record and the relation edge evidence role record into the semantic relabeling graph to generate the evidence role mapping result. Then, use the evidence role mapping result to convert the path mutual exclusion judgment object in the improved Suurballe algorithm from the graph connection structure into the evidence source conflict object. Based on the evidence role mapping results, mark the conflicting evidence sources and the common semantic carriers in the semantic recalibration graph, configure the residual occupancy status for the conflicting evidence sources, and configure the shared access status for the common semantic carriers. Residual search edges are generated based on residual occupancy status, shared access status, and semantic relabeling edge weights. This ensures that residual search edges associated with residual occupancy status form a reverse release relationship, while residual search edges associated with shared access status maintain a positive access relationship. The residual search edges are then aggregated to form the evidence role residual graph of the improved Suurballe algorithm.

[0011] Optionally, step six specifically includes: Extract the starting entity, candidate answer range, and residual search edge from the residual graph of evidence roles. Configure the starting entity as the search source node and the graph entity node in the candidate answer range as the search target node to establish a dual evidence path search task. Based on the semantic relabeling edge weights, the first round of minimum cost path search is performed in the evidence role residual graph to generate the first evidence path, and the evidence source conflict objects passed through by the first evidence path are marked as path occupancy objects. Adjust the residual search edges according to the residual occupancy status associated with the path occupancy object, switch the residual search edges associated with the path occupancy object to the reverse release relationship, keep the residual search edges associated with the semantic common carrier object to the forward passage relationship, and generate a search update residual graph. A second round of minimum cost path search is performed in the search update residual graph to generate a second evidence path, and the reverse release relationship is resolved on the first evidence path and the second evidence path to form a dual evidence path record; The dual-evidence path records are associated and organized according to the search target node, the order of the path and the path cost to generate a candidate evidence path set.

[0012] Optionally, step seven specifically includes: Extract dual-evidence path records from the candidate evidence path set, search for target nodes and path traversal order, arrange the graph entity nodes and graph relation edges in the dual-evidence path records according to the path traversal order, and generate path structure records. Extract the question semantic parsing results and question slot records from the question-answering constraint unit, match the path structure records with the question semantic parsing results in semantic position, match the path structure records with the question slot records in slot position, and generate path semantic alignment records; Based on the path semantic alignment record, quantification is performed on the answer consistency state between the search target node and the candidate answer range, the direction consistency state between the graph relation edge and the question semantic parsing result, and the constraint consistency state between the path structure record and the question slot record to generate path consistency residuals; The path structure record where the path consistency residual reaches the residual threshold is identified as a conflict path segment, and the conflict path segment is associated with the graph entity node and graph relation edge to generate a conflict segment record; Based on the conflict fragment records, the semantic relabeling edge weights and residual occupancy states of the residual search edges in the evidence role residual graph are adjusted to generate an updated evidence role residual graph.

[0013] Optionally, step eight specifically includes: Extract residual search edges, semantic relabeling edge weights and residual occupancy status from the updated evidence role residual graph, and extract dual evidence path records, search target nodes, path costs and path consistency residuals from the candidate evidence path set to generate target evidence chain screening records. Based on the target evidence chain, the path cost, path consistency residual and residual occupancy status corresponding to the dual evidence path records are jointly sorted to generate evidence chain sorting results. The dual-evidence path record that ranks first in the evidence chain sorting result is identified as the target evidence chain. Graph entity nodes, graph relation edges, and the order of the search target node and path are extracted from the target evidence chain to generate the target evidence chain record. Align the target evidence chain record with the question-and-answer constraint unit in slots, generate the answer body based on the search target node, and generate the evidence path based on the graph entity node, graph relation edge and path sequence. The answer subject and evidence path are linked and organized to generate answer text and evidence path. The answer text, evidence path, target evidence chain record and question-answer constraint unit are linked and output to form intelligent question-answering processing results.

[0014] According to an embodiment of the present invention, an intelligent question-answering system based on a knowledge graph includes: The question constraint module is used to obtain user questions, perform semantic parsing and slot alignment on user questions, and generate question-answer constraint units. The subgraph extraction module is used to locate the starting entity and the range of candidate answers in the knowledge graph, and extract candidate question-answer subgraphs that cover the range of candidate answers around the starting entity; The tagging configuration module is used to configure node semantic potential tags and relation edge evidence tags for candidate question-answer subgraphs; The edge weight labeling module is used to perform edge weight labeling on the candidate question-answering subgraph based on the semantic potential energy labeling of nodes and the relation edge evidence labeling, and generate a semantic relabeling graph. The residual graph construction module is used to perform evidence role mapping on nodes and relation edges in the semantic recalibration graph to generate an evidence role residual graph. The path search module is used to perform dual evidence path search in the evidence role residual graph and generate a set of candidate evidence paths; The residual update module is used to semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual graph. The question-and-answer output module is used to determine the target evidence chain based on the updated evidence role residual map, generate answer text and evidence path, and output the intelligent question-and-answer processing results.

[0015] The beneficial effects of this invention are: This invention performs semantic parsing and slot realignment on user questions to form question-and-answer constraint units that constrain the starting entity, the range of candidate answers, and the path retrieval process. It then extracts candidate question-and-answer subgraphs covering the range of candidate answers from the knowledge graph around the starting entity, enabling stable correspondences between the semantic information of the question and the entity nodes and relational edges of the graph. By configuring node semantic potential labels and relational edge evidence labels on the candidate question-and-answer subgraphs, and then performing edge weight calibration, a semantic recalibration graph is generated. This allows the path search process to no longer solely rely on graph connection distances but is simultaneously constrained by question semantics, answer range, and evidence support status. This reduces the risk of entity localization errors, overly broad candidate path expansion, and irrelevant paths entering the question-and-answer process, thereby improving the accuracy of answer retrieval.

[0016] This invention further constructs an evidence role residual graph for an improved Suurballe algorithm through evidence role mapping results. This transforms the path mutual exclusion judgment object from a normal graph connection structure into an evidence source conflict object, enabling dual-evidence path search to retain the traversability of semantically common carrier objects while limiting the repeated occupation of evidence source conflict objects, thereby enhancing the independence between evidence paths. By semantically aligning the candidate evidence path set with the question-answering constraint unit, calculating the path consistency residual, and updating the evidence role residual graph, it is possible to perform rollback correction on path segments with inconsistent answer entities, relational directions, and constraint conditions. This ensures that the final determined target evidence chain simultaneously satisfies the requirements of path cost, evidence source, and question semantic consistency, thereby improving the reliability, interpretability, and applicability of intelligent question-answering results in complex multi-hop question-answering scenarios. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent question-answering method based on knowledge graphs proposed in this invention; Figure 2 This is a schematic diagram illustrating the construction of the evidence role residual graph for an intelligent question-answering processing method based on knowledge graphs proposed in this invention. Figure 3 This is a schematic diagram illustrating the generation of the target evidence chain in an intelligent question-answering method based on knowledge graphs proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A knowledge graph-based intelligent question answering method includes the following steps: Step 1: Obtain the user's question, perform semantic parsing and slot alignment on the user's question, and generate question-answer constraint units; Step 2: Based on the question-answering constraint unit, locate the starting entity and the range of candidate answers in the knowledge graph, and extract the candidate question-answer subgraph covering the range of candidate answers around the starting entity; Step 3: Configure node semantic potential labels and relation edge evidence labels for the candidate question-answering subgraph based on the question-answering constraint unit; Step 4: Perform edge weight labeling on the candidate question-answering subgraph based on node semantic potential labeling and relation edge evidence labeling to generate a semantic relabeling graph; Step 5: Perform evidence role mapping on the nodes and relation edges in the semantic recalibration graph, generate evidence role mapping results, and construct the evidence role residual graph of the improved Suurballe algorithm based on the evidence role mapping results; Step 6: In the residual graph of evidence roles, perform a dual evidence path search from the starting entity to the range of candidate answers to generate a set of candidate evidence paths; Step 7: Semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual map based on the path consistency residual. Step 8: Determine the target evidence chain based on the updated evidence role residual map, generate the answer text and evidence path according to the target evidence chain, and output the intelligent question answering processing result.

[0020] In this embodiment, step one specifically includes: The system acquires user questions and assigns question identifiers to them. It establishes a correspondence between user questions and question identifiers, processes the character encoding, full-width and half-width forms, uppercase and lowercase forms, and punctuation forms in user questions in a unified manner, determines the sentence boundaries according to the sentence pause position and syntactic connection relationship, and removes whitespace placeholder characters, control characters, and repeated separator characters to generate a regular question text. The text of regular questions is segmented into terms according to the domain dictionary and semantic segmentation rules. Terms with the same domain meaning are mapped to a unified term expression to generate a sequence of question terms. The master-slave relationship, modification relationship and pointing relationship between terms are analyzed based on the sequence of question terms. The analysis results are mapped to the entity category and relation category of the knowledge graph to generate semantic component records. Entity pointing components are extracted based on the word positions corresponding to entity categories in the semantic component records, and relation pointing components are extracted based on the word positions corresponding to relation categories in the semantic component records. The entity pointing components and relation pointing components are combined to form the semantic parsing result of the question. The answer range pointing and constraint condition pointing are determined based on the semantic parsing result of the question, and the question slot record is generated. Establish a correspondence between the semantic parsing results of the questions and the question slot records according to the question identifiers, and configure constraint unit identifiers for the correspondences, so that the question identifiers, semantic parsing results of the questions, question slot records and constraint unit identifiers form a unified record structure, and generate question-answer constraint units.

[0021] In this embodiment, step two specifically includes: Extract entity pointing components, answer range pointing, and constraint condition pointing from the question-and-answer constraint unit. Retrieve the knowledge graph composed of graph entity nodes, graph relation edges, and graph attribute fields. Perform text normalization on the entity pointing components to unify the character form, term form, and domain expression form in the entity pointing components. Then, generate search keys according to the normalized entity pointing components. Write the search keys into the node name field matching channel and the node category field matching channel respectively. Then, match the search keys with the node name field and node category field of the graph entity nodes item by item to form an entity matching record containing the matching node identifier, node name matching result, and node category matching result. Extract node name matching value and node category matching value from entity matching records. Perform product calculation on node name matching value and name weight, and perform product calculation on node category matching value and category weight. Then sum the product calculation results to generate entity matching confidence value. Compare entity matching confidence value with entity threshold. Graph entity nodes whose entity matching confidence value reaches entity threshold are determined as starting entities. Associate starting entities with question answering constraint units according to question identifiers. Perform answer category parsing on the answer range pointer, extract category terms and category limiting terms from the answer range pointer, and consolidate the category terms and category limiting terms into answer category tags. Match the answer category tags with the node category fields of the graph entity nodes to form answer category boundaries. Then extract attribute limiting content and relation limiting content from the constraint condition pointer. Use attribute limiting content to limit the value range of graph attribute fields, and use relation limiting content to limit the connection range of graph relation edges to generate candidate answer ranges. Starting with the initial entity as the starting point for expansion, hierarchical restricted expansion is performed along the graph relation edges in the knowledge graph. In each expansion level, the expansion level, the reached graph entity node, and the passed graph relation edges are recorded. The range of the reached graph entity node and the candidate answer range are matched. The reached graph entity nodes and passed graph relation edges that are within the candidate answer range are collected and formed into a candidate question-and-answer subgraph that covers the candidate answer range according to the expansion level and connection order.

[0022] In this embodiment, step three specifically includes: Extract the semantic parsing results of questions and the question slot records from the question-answering constraint unit, extract the graph entity nodes and graph relation edges from the candidate question-answering subgraph, match the entity pointing components and relation pointing components in the semantic parsing results with the node name field and node category field of the graph entity nodes, and establish a mapping relationship according to the connection position of the graph entity nodes in the candidate question-answering subgraph to generate node semantic correspondence records. Based on the node semantic correspondence record, extract the semantic matching value, slot coverage value and connection position value between the graph entity node and the question semantic parsing result. Perform normalization processing on the semantic matching value, slot coverage value and connection position value, and then perform weighted calculation according to the preset weight to generate the node semantic potential value. Write the node semantic potential value into the graph entity node to form the node semantic potential value tag. Map the answer range and constraint points in the question slot records to graph relation edges. Establish corresponding relationships according to the graph entity nodes connected by the graph relation edges, the relation type of the graph relation edges, and the connection order of the graph relation edges in the candidate question and answer subgraphs, and generate relation edge evidence corresponding records. Based on the corresponding records of relation edge evidence, extract the relation support value, endpoint coverage value, and connection order value between the graph relation edge and the question slot record. Perform normalization processing on the relation support value, endpoint coverage value, and connection order value, and then perform weighted calculation according to preset weights to generate relation edge evidence value. Write the relation edge evidence value into the graph relation edge to form relation edge evidence marker.

[0023] In this embodiment, step four specifically includes: Extract graph relation edges one by one from the candidate question-and-answer subgraph. Determine the starting graph entity node and the ending graph entity node according to the connection direction of the graph relation edges. Then, extract the starting semantic potential value from the node semantic potential energy marker corresponding to the starting graph entity node, extract the ending semantic potential value from the node semantic potential energy marker corresponding to the ending graph entity node, and extract the relation edge evidence value from the relation edge evidence marker corresponding to the graph relation edge. Then, associate and organize the graph relation edges, starting graph entity nodes, ending graph entity nodes, starting semantic potential values, ending semantic potential values, and relation edge evidence values ​​to generate edge weight labeling records. Perform difference calculation on the initial semantic potential value and the terminal semantic potential value in the edge weight labeling record, and form the node potential offset value by subtracting the initial semantic potential value from the terminal semantic potential value. Perform evidence cost transformation on the relation edge evidence value and map the relation edge evidence value to the evidence support cost corresponding to the path search cost. The basic connection cost is extracted from the graph relation edge, and the basic connection cost is used as the original passage cost of the graph relation edge. The node potential energy offset value is used as the semantic direction adjustment amount of the graph relation edge, and the evidence support cost value is used as the evidence support adjustment amount of the graph relation edge. The basic connection cost, node potential energy offset value and evidence support cost value are fused and calculated to generate semantic recalibrated edge weights. The semantic relabeling edge weights are written into the corresponding graph relation edges, so that the graph relation edges carry the semantic relabeling edge weights to participate in path search. The candidate question-answer subgraph is reconstructed according to the connection relationship between the graph entity nodes, graph relation edges and semantic relabeling edge weights, and a semantic relabeling graph is generated.

[0024] In this embodiment, step five specifically includes: Extract graph entity nodes, graph relation edges, and semantic relabeling edge weights from the semantic relabeling graph. Assign each graph entity node to the semantic carrying relationship in the question-answering constraint unit, and associate each graph entity node with the range belonging relationship in the candidate answer range. Then record the connection position of each graph entity node in the semantic relabeling graph, so that the semantic carrying relationship, range belonging relationship, and connection position corresponding to the same graph entity node form the same node role judgment basis, and generate node evidence role record; Each graph relation edge is mapped to the slot support relation in the question-and-answer constraint unit, and each graph relation edge is mapped to the connection pointing relation in the candidate answer range. Then, the semantic relabeling edge weight associated with each graph relation edge is written into the same relation edge record, so that the slot support relation, connection pointing relation and semantic relabeling edge weight corresponding to the same graph relation edge form the same relation edge role judgment basis, and a relation edge evidence role record is generated. The node evidence role record is bound to the corresponding graph entity node, and the relation edge evidence role record is bound to the corresponding graph relation edge. These are then written into the semantic relabeling graph to generate the evidence role mapping result. The evidence role mapping result is used to change the mutual exclusion judgment basis of the improved Suurballe algorithm, so that the path search process no longer directly uses the overlap of graph entity nodes and the overlap of graph relation edges as the mutual exclusion basis, but uses the evidence source conflict reflected by the node evidence role record and the relation edge evidence role record as the mutual exclusion basis. Thus, the path mutual exclusion judgment object is changed from the graph connection structure to the evidence source conflict object. Based on the evidence role mapping results, the graph entity nodes and graph relation edges in the semantic recalibration graph that bear the support of evidence sources and are at risk of source reuse are marked as evidence source conflict objects. The graph entity nodes and graph relation edges in the semantic recalibration graph that bear the semantics of the question, the range of the answer, and the connection transition and are allowed to be traversed by the same path are marked as semantic common bearing objects. The evidence source conflict objects are configured with residual occupancy status, and the semantic common bearing objects are configured with shared passage status. Residual search edges are generated based on residual occupancy status, shared access status, and semantic relabeling edge weights. Semantic relabeling edge weights are retained in residual search edges associated with residual occupancy status, and reverse release relationships are established, enabling conflicting evidence sources that have been occupied by an evidence path to form a backtracking release in residual search. Semantic relabeling edge weights are retained in residual search edges associated with shared access status, and forward access relationships are established, enabling semantically common carrier objects to maintain access in dual evidence path search. The residual search edges are aggregated to form the evidence role residual graph of the improved Suurballe algorithm.

[0025] This invention introduces an evidence role mapping mechanism into the residual graph construction stage of the Suurballe algorithm. It improves the traditional algorithm's approach of using overlapping graph entity nodes and graph relation edges as the basis for mutual exclusion judgment. Instead, it uses the evidence source conflicts reflected in the node evidence role records and relation edge evidence role records as the basis for mutual exclusion judgment. This prevents the algorithm from mechanically excluding reasonably shared semantically common carrier objects in the path, and instead focuses on restricting graph entity nodes and graph relation edges that bear evidence source support and pose a risk of source reuse. Simultaneously, residual search edges are generated through residual occupancy status, shared access status, and semantically recalibrated edge weights. A reverse release relationship is established among the residual search edges associated with the residual occupancy status, and a reverse release relationship is established in the shared access status. By establishing a positive passage relationship in the residual search edges of state association, conflicting evidence sources that have been occupied by an evidence path can be released through a backtracking process in the residual search. This allows semantically common carrier objects to remain accessible in dual-evidence path searches. The improvement avoids the problem of two evidence paths appearing different but having duplicate evidence sources, and also avoids the problem of incorrectly excluding valid paths due to the sharing of common semantic nodes. It ensures that dual-evidence paths remain independent at the evidence source level, continuous at the semantic carrier level, and balance low cost and high credibility at the path search level. This improves the reliability of the target evidence chain, the support strength of the answer text, the interpretability of the question-and-answer results, and the stability of path retrieval in complex multi-hop question-and-answer scenarios.

[0026] In this embodiment, step six specifically includes: Extract the starting entity, candidate answer range, and residual search edge from the residual graph of evidence roles. Configure the starting entity as the search source node and configure each graph entity node in the candidate answer range as the search target node to establish a dual evidence path search task. Organize the search source node, search target node, residual search edge, and semantic relabeling edge weight into path search input records. Based on the semantic relabeling edge weight, the first round of minimum cost path search is performed in the evidence role residual graph. The semantic relabeling edge weight is accumulated along the residual search edge. The minimum cumulative cost connection sequence from the search source node to the search target node is selected as the first evidence path, and the evidence source conflict objects passed through by the first evidence path are marked as path occupancy objects. Adjust the residual search edges according to the residual occupancy status associated with the path occupancy object, switch the residual search edges associated with the path occupancy object to a reverse release relationship, so that the evidence source conflict object of the first evidence path occupancy forms a release constraint in the second round of search, keep the residual search edges associated with the semantic common carrier object in a positive passage relationship, and generate a search update residual graph. The second round of minimum cost path search is performed in the search update residual graph. The semantic relabeling edge weight is accumulated along the residual search edge in the search update residual graph to generate the second evidence path. The reverse release relationship resolution is performed on the residual search edges in the first evidence path and the second evidence path that are opposite in direction and correspond to the conflicting object of the same evidence source to form a dual evidence path record. The dual evidence path records are associated and organized according to the search target node, the order of the path and the path cost. The correspondence between the first evidence path, the second evidence path, the search target node and the path cost is recorded to generate a candidate evidence path set.

[0027] In this embodiment, step seven specifically includes: Extract dual-evidence path records, search target nodes, and path traversal order from the candidate evidence path set. Arrange the graph entity nodes and graph relationship edges in the dual-evidence path records according to the path traversal order, keep the graph relationship edges between adjacent graph entity nodes in the corresponding connection positions, and establish a correspondence between the arrangement results and the search target nodes to generate path structure records. Extract the question semantic parsing results and question slot records from the question-answering constraint unit. Make semantic position correspondence between the graph entity nodes in the path structure record and the entity pointing components in the question semantic parsing results. Make semantic position correspondence between the graph relation edges in the path structure record and the relation pointing components in the question semantic parsing results. Then make slot position correspondence between the search target node and the path connection position in the path structure record and the question slot record to generate a path semantic alignment record. Based on the path semantic alignment record, the matching degree of the search target node falling into the candidate answer range is calculated to form the answer consistency state. The matching degree between the connection direction of the graph relation edge and the relation pointing direction in the question semantic parsing result is calculated to form the direction consistency state. The coverage degree of the path structure record on the answer range pointing and constraint condition pointing in the question slot record is calculated to form the constraint consistency state. Then, residual convergence is performed on the answer consistency state, direction consistency state and constraint consistency state to generate path consistency residual. The path consistency residual is compared with the residual threshold. The path structure record where the path consistency residual reaches the residual threshold is identified as a conflict path segment. The corresponding graph entity node and graph relation edge are extracted according to the position of the conflict path segment in the path structure record. The conflict path segment, graph entity node and graph relation edge are associated to generate a conflict segment record. Based on the conflict fragment records, the corresponding residual search edge is located in the evidence role residual graph. The semantic relabeling edge weight of the residual search edge is adjusted incrementally, the residual occupancy status associated with the residual search edge is updated, and the connection relationship between the residual search edges is maintained, generating an updated evidence role residual graph.

[0028] In this embodiment, step eight specifically includes: Extract residual search edges, semantic relabeling edge weights, and residual occupancy status from the updated evidence role residual graph. Extract dual evidence path records, search target nodes, path costs, and path consistency residuals from the candidate evidence path set. Establish a correspondence between the residual search edges, semantic relabeling edge weights, residual occupancy status, and path passage order in the dual evidence path records. Then associate the search target nodes, path costs, and path consistency residuals with the corresponding dual evidence path records to generate target evidence chain screening records. Based on the target evidence chain, the path cost corresponding to the dual evidence path record is normalized, the residual value is normalized, the residual occupancy status is converted into the occupancy constraint value, and then the normalized result of the value, the normalized result of the residual value and the occupancy constraint value are jointly compared. The dual evidence path records are arranged according to the comparison result to generate the evidence chain ranking result. The dual-evidence path record that ranks first in the evidence chain sorting result is identified as the target evidence chain. Graph entity nodes, graph relation edges, and search target nodes are extracted from the target evidence chain according to the path traversal order. The graph entity nodes, graph relation edges, search target nodes, and path traversal order are then arranged into the same record structure to generate the target evidence chain record. Align the target evidence chain record with the question-and-answer constraint unit in slots to form a correspondence between the search target node and the answer range in the question-and-answer constraint unit, and to form a correspondence between the graph entity node and the graph relation edge and the question semantic parsing result in the question-and-answer constraint unit. Generate the answer subject based on the search target node, and generate the evidence path based on the graph entity node, graph relation edge and path traversal order. The answer subject and evidence path are associated and organized. The correspondence between the answer subject, evidence path and target evidence chain record is established according to the question identifier in the question-answer constraint unit. The answer text and evidence path are generated, and the answer text, evidence path, target evidence chain record and question-answer constraint unit are associated and output to form intelligent question-answering processing results.

[0029] A knowledge graph-based intelligent question-answering system includes: The question constraint module is used to obtain user questions, perform semantic parsing and slot alignment on user questions, and generate question-answer constraint units. The subgraph extraction module is used to locate the starting entity and the range of candidate answers in the knowledge graph, and extract candidate question-answer subgraphs that cover the range of candidate answers around the starting entity; The tagging configuration module is used to configure node semantic potential tags and relation edge evidence tags for candidate question-answer subgraphs; The edge weight labeling module is used to perform edge weight labeling on the candidate question-answering subgraph based on the semantic potential energy labeling of nodes and the relation edge evidence labeling, and generate a semantic relabeling graph. The residual graph construction module is used to perform evidence role mapping on nodes and relation edges in the semantic recalibration graph to generate an evidence role residual graph. The path search module is used to perform dual evidence path search in the evidence role residual graph and generate a set of candidate evidence paths; The residual update module is used to semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual graph. The question-and-answer output module is used to determine the target evidence chain based on the updated evidence role residual map, generate answer text and evidence path, and output the intelligent question-and-answer processing results.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to an intelligent question-and-answer scenario for an enterprise operations and maintenance knowledge base. The knowledge graph consists of equipment entities, component entities, fault entities, work order entities, handling measure entities, graph relationship edges, and graph attribute fields, covering equipment files, fault records, inspection records, maintenance records, and standard handling documents. In actual use, operations and maintenance personnel typically ask questions in natural language, such as "Which component should be checked after a certain model of equipment experiences continuous alarms?", "Is there a correlation between a certain type of fault and the most recent maintenance record?", and "What is the source of evidence for a certain handling measure?" These questions often simultaneously contain entity abbreviations, implicit relationships, answer ranges, and limiting conditions. Ordinary knowledge graph question-and-answer processes are prone to incorrectly linking equipment aliases to nearby nodes or only returning superficially related answers based on the shortest path, resulting in answers lacking independent evidence support.

[0031] In this scenario, after the anonymized maintenance question is input into the system of this invention, the system first performs semantic parsing and slot alignment on the user's question to form question-answer constraint units. Then, based on the question-answer constraint units, it locates the starting entity and the range of candidate answers in the knowledge graph and extracts candidate question-answer subgraphs around the starting entity. After the candidate question-answer subgraphs are formed, the system configures node semantic potential energy labels for graph entity nodes and relational edge evidence labels for graph relational edges. It also forms a semantic recalibration graph through edge weight labeling, so that relational edges that are semantically close to the question and have strong evidence support can obtain lower search costs in path search. Subsequently, the system performs evidence role mapping on the nodes and relational edges in the semantic recalibration graph, converting the ordinary graph connection structure into evidence source conflict objects and semantic common carrier objects, and then constructs an evidence role residual graph of the improved Suurballe algorithm. Compared with conventional solutions, this invention does not simply find two shortest paths, but restricts the repeated use of the same evidence source while preserving the passage of common semantic nodes, thereby obtaining dual evidence paths with independent evidence sources. After the candidate evidence path set is generated, the system semantically aligns the path structure with the question-and-answer constraint unit, calculates the path consistency residual, updates the residual graph for path segments with inconsistent answer directions, constraints, and evidence sources, and finally determines the target evidence chain, and simultaneously outputs the answer text and evidence path.

[0032] The comparative scheme employs a conventional knowledge graph question-answering process, specifically involving entity linking and relation similarity matching of user questions, performing shortest path expansion centered on the starting entity, and generating answers by ranking candidate paths according to path length, relation similarity, and node matching degree. The comparative scheme does not configure node semantic potential tags, construct evidence role residual graphs, or calculate path consistency residuals. Both the present invention and the comparative scheme use the same set of anonymized questions and the same knowledge graph. The question set includes questions related to equipment fault location, maintenance measure query, alarm cause tracking, and evidence path explanation. Manual annotation results are used as the evaluation benchmark, as shown in Table 1. Table 1. Comparison of the effectiveness of evidence chain question-and-answer sessions.

[0033] As shown in Table 1, the present invention, with a slight increase in average response time, improves the initial entity localization accuracy from 87.6% to 94.8% and the answer accuracy from 82.9% to 92.6%, indicating that the question-answering constraint unit, node semantic potential labeling, and relation edge evidence labeling can effectively reduce entity localization bias and irrelevant path interference. The average number of candidate paths decreased from 18.7 per question to 9.4 per question, indicating that the semantic recalibration graph can compress the candidate path search space. The evidence source duplication rate decreased from 31.5% to 8.7%, indicating that the evidence role residual graph can avoid dual evidence paths depending on the same fact source. The path direction inconsistency rate decreased from 14.2% to 4.1%, indicating that the path consistency residual update can correct conflicts in relation direction and constraint conditions.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent question-answering method, characterized in that, Includes the following steps: Step 1: Obtain the user's question, perform semantic parsing and slot alignment on the user's question, and generate question-answer constraint units; Step 2: Based on the question-answering constraint unit, locate the starting entity and the range of candidate answers in the knowledge graph, and extract the candidate question-answer subgraph covering the range of candidate answers around the starting entity; Step 3: Configure node semantic potential labels and relation edge evidence labels for the candidate question-answering subgraph based on the question-answering constraint unit; Step 4: Perform edge weight labeling on the candidate question-answering subgraph based on node semantic potential labeling and relation edge evidence labeling to generate a semantic relabeling graph; Step 5: Perform evidence role mapping on the nodes and relation edges in the semantic recalibration graph, generate evidence role mapping results, and construct the evidence role residual graph of the improved Suurballe algorithm based on the evidence role mapping results; Step 6: In the residual graph of evidence roles, perform a dual evidence path search from the starting entity to the range of candidate answers to generate a set of candidate evidence paths; Step 7: Semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual map based on the path consistency residual. Step 8: Determine the target evidence chain based on the updated evidence role residual map, generate the answer text and evidence path according to the target evidence chain, and output the intelligent question answering processing result.

2. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step one specifically involves: The system acquires user questions and generates question identifiers, associates user questions with question identifiers, performs character format standardization, statement boundary determination, and invalid character cleanup on user questions, and generates well-organized question text. Perform term segmentation and domain term normalization on regular question text to generate question term sequence. Based on the question term sequence, perform semantic dependency analysis and domain semantic mapping to generate semantic component record. Based on the semantic component record, entity pointing components and relation pointing components are extracted to form the question semantic parsing result. Based on the question semantic parsing result, the answer range pointing and constraint condition pointing are determined to generate question slot record; The semantic parsing results of the questions are associated and organized with the question slot records according to the question identifier, and the constraint unit identifier is configured to generate the question-answer constraint unit.

3. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step two specifically involves: Extract entity pointing components, answer range pointing and constraint condition pointing from the question and answer constraint unit, retrieve the knowledge graph composed of graph entity nodes, graph relation edges and graph attribute fields, perform text normalization and retrieval key generation on the entity pointing components, match the retrieval key with the node name field and node category field of the graph entity node to form entity matching records; Extract node name matching values ​​and node category matching values ​​from entity matching records, perform weighted aggregation on node name matching values ​​and node category matching values ​​to generate entity matching confidence values, determine the graph entity nodes whose entity matching confidence values ​​reach the entity threshold as the starting entities, and associate the starting entities with question answering constraint units; The answer range is parsed to form an answer category marker. The answer category marker is matched with the node category field of the graph entity node to form the answer category boundary. Then, the constraint condition is used to limit the value range of the graph attribute field and the connection range of the graph relationship edge to generate the candidate answer range. Starting from the initial entity, perform hierarchical restricted expansion along the graph relation edges in the knowledge graph, record the expansion level, the graph entity nodes reached, and the graph relation edges passed through. Collect the graph entity nodes that are within the range of candidate answers and the graph relation edges passed through to generate a candidate question-and-answer subgraph that covers the range of candidate answers.

4. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step three specifically involves: Extract the semantic parsing results of questions and the question slot records from the question-answering constraint unit, extract the graph entity nodes and graph relation edges from the candidate question-answering subgraph, map the semantic parsing results of questions to the graph entity nodes, and generate node semantic correspondence records; According to the node semantic correspondence record, the semantic correspondence state between the graph entity node and the question semantic parsing result is quantified to generate the node semantic potential value, and the node semantic potential value is written into the graph entity node to form the node semantic potential value tag. Map the question slot records to the graph relation edges to generate corresponding records for relation edge evidence; Based on the corresponding records of relation edge evidence, the evidence support status between the graph relation edge and the question slot record is quantified, relation edge evidence value is generated, and the relation edge evidence value is written into the graph relation edge to form relation edge evidence marker.

5. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step four specifically involves: Extract graph relation edges from candidate question-and-answer subgraphs, determine the starting graph entity nodes and the ending graph entity nodes connected by the graph relation edges, extract the starting semantic potential value and the ending semantic potential value from the node semantic potential energy markers, extract the relation edge evidence value from the relation edge evidence markers, and generate edge weight labeling records. Perform difference calculation on the initial semantic potential value and the terminal semantic potential value in the edge weight labeling record to generate the node potential offset value; perform evidence cost transformation on the relation edge evidence value to generate the evidence support cost value. Extract the basic connection cost from the graph relation edges, and perform a fusion calculation on the basic connection cost, node potential offset value and evidence support cost to generate semantic relabeling edge weights; Write the semantic relabeling edge weights into the corresponding graph relation edges, and reconstruct the candidate question-answer subgraph according to the connection relationship between the graph entity nodes, graph relation edges and semantic relabeling edge weights to generate the semantic relabeling graph.

6. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step five specifically involves: Extract graph entity nodes, graph relation edges, and semantic relabeling edge weights from the semantic relabeling graph. Associate the semantic carrying relationship of the graph entity nodes in the question-answering constraint unit, the range belonging relationship in the candidate answer range, and the connection position in the semantic relabeling graph to generate node evidence role records. The graph relation edges are associated with the slot support relationship of the question-and-answer constraint unit, the connection pointing relationship of the candidate answer range, and the semantic relabeling edge weight associated with the graph relation edges to generate relation edge evidence role records; Write the node evidence role record and the relation edge evidence role record into the semantic relabeling graph to generate the evidence role mapping result. Then, use the evidence role mapping result to convert the path mutual exclusion judgment object in the improved Suurballe algorithm from the graph connection structure into the evidence source conflict object. Based on the evidence role mapping results, mark the conflicting evidence sources and the common semantic carriers in the semantic recalibration graph, configure the residual occupancy status for the conflicting evidence sources, and configure the shared access status for the common semantic carriers. Residual search edges are generated based on residual occupancy status, shared access status, and semantic relabeling edge weights. This ensures that residual search edges associated with residual occupancy status form a reverse release relationship, while residual search edges associated with shared access status maintain a positive access relationship. The residual search edges are then aggregated to form the evidence role residual graph of the improved Suurballe algorithm.

7. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step six specifically involves: Extract the starting entity, candidate answer range, and residual search edge from the residual graph of evidence roles. Configure the starting entity as the search source node and the graph entity node in the candidate answer range as the search target node to establish a dual evidence path search task. Based on the semantic relabeling edge weights, the first round of minimum cost path search is performed in the evidence role residual graph to generate the first evidence path, and the evidence source conflict objects passed through by the first evidence path are marked as path occupancy objects. Adjust the residual search edges according to the residual occupancy status associated with the path occupancy object, switch the residual search edges associated with the path occupancy object to the reverse release relationship, keep the residual search edges associated with the semantic common carrier object to the forward passage relationship, and generate a search update residual graph. A second round of minimum cost path search is performed in the search update residual graph to generate a second evidence path, and the reverse release relationship is resolved on the first evidence path and the second evidence path to form a dual evidence path record; The dual-evidence path records are associated and organized according to the search target node, the order of the path and the path cost to generate a candidate evidence path set.

8. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step seven specifically involves: Extract dual-evidence path records from the candidate evidence path set, search for target nodes and path traversal order, arrange the graph entity nodes and graph relation edges in the dual-evidence path records according to the path traversal order, and generate path structure records. Extract the question semantic parsing results and question slot records from the question-answering constraint unit, match the path structure records with the question semantic parsing results in semantic position, match the path structure records with the question slot records in slot position, and generate path semantic alignment records; Based on the path semantic alignment record, quantification is performed on the answer consistency state between the search target node and the candidate answer range, the direction consistency state between the graph relation edge and the question semantic parsing result, and the constraint consistency state between the path structure record and the question slot record to generate path consistency residuals; The path structure record where the path consistency residual reaches the residual threshold is identified as a conflict path segment, and the conflict path segment is associated with the graph entity node and graph relation edge to generate a conflict segment record; Based on the conflict fragment records, the semantic relabeling edge weights and residual occupancy states of the residual search edges in the evidence role residual graph are adjusted to generate an updated evidence role residual graph.

9. The intelligent question-answering method based on knowledge graphs according to claim 1, characterized in that, Step eight specifically involves: Extract residual search edges, semantic relabeling edge weights and residual occupancy status from the updated evidence role residual graph, and extract dual evidence path records, search target nodes, path costs and path consistency residuals from the candidate evidence path set to generate target evidence chain screening records. Based on the target evidence chain, the path cost, path consistency residual and residual occupancy status corresponding to the dual evidence path records are jointly sorted to generate evidence chain sorting results. The dual-evidence path record that ranks first in the evidence chain sorting result is identified as the target evidence chain. Graph entity nodes, graph relation edges, and the order of the search target node and path are extracted from the target evidence chain to generate the target evidence chain record. Align the target evidence chain record with the question-and-answer constraint unit in slots, generate the answer body based on the search target node, and generate the evidence path based on the graph entity node, graph relation edge and path sequence. The answer subject and evidence path are linked and organized to generate answer text and evidence path. The answer text, evidence path, target evidence chain record and question-answer constraint unit are linked and output to form intelligent question-answering processing results.

10. A knowledge graph-based intelligent question-answering system, comprising executing the knowledge graph-based intelligent question-answering method described in any one of claims 1 to 9, characterized in that, include: The question constraint module is used to obtain user questions, perform semantic parsing and slot alignment on user questions, and generate question-answer constraint units. The subgraph extraction module is used to locate the starting entity and the range of candidate answers in the knowledge graph, and extract candidate question-answer subgraphs that cover the range of candidate answers around the starting entity; The tagging configuration module is used to configure node semantic potential tags and relation edge evidence tags for candidate question-answer subgraphs; The edge weight labeling module is used to perform edge weight labeling on the candidate question-answering subgraph based on the semantic potential energy labeling of nodes and the relation edge evidence labeling, and generate a semantic relabeling graph. The residual graph construction module is used to perform evidence role mapping on nodes and relation edges in the semantic recalibration graph to generate an evidence role residual graph. The path search module is used to perform dual evidence path search in the evidence role residual graph and generate a set of candidate evidence paths; The residual update module is used to semantically align the candidate evidence path set with the question-answering constraint unit, calculate the path consistency residual, and update the evidence role residual graph. The question-and-answer output module is used to determine the target evidence chain based on the updated evidence role residual map, generate answer text and evidence path, and output the intelligent question-and-answer processing results.

Images