Data retrieval method and apparatus, electronic device, and readable storage medium

Abstract

The present disclosure relates to the technical field of data processing, and provides a data retrieval method and device, electronic equipment and readable storage medium. The method comprises: performing intention anchoring processing on to-be-queried data to obtain initial anchor point data, reasoning link data and retrieval keyword data; performing directional recall processing on preset index data based on the retrieval keyword data and the reasoning link data to obtain effective evidence data; performing reasoning generation processing on the effective evidence data to obtain intermediate conclusion data; performing collaborative calibration processing on the intermediate conclusion data, the initial anchor point data and the reasoning link data to obtain anchor point calibration data and dynamic reasoning chain data; and performing convergence aggregation processing on the dynamic reasoning chain data and the anchor point calibration data to obtain target retrieval result data, thereby improving the positioning accuracy in the initial stage of retrieval, improving the relevance and accuracy of evidence recall, enhancing the reasoning depth of the retrieval process, and improving the integrity and reliability of the retrieval result.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a data retrieval method, apparatus, electronic device, and readable storage medium. Background Technology

[0002] In Retrieval-Augmented Generation Systems (RAGs), user queries are typically matched directly with vectorized document blocks based on similarity, and answers are generated based on the retrieved fragments. However, when dealing with complex queries that require multi-step logical reasoning, the accuracy and completeness of the reasoning are often difficult to guarantee due to incoherent retrieved evidence and a lack of logical guidance. In existing technologies, the relevance of retrieval can be enhanced by introducing structured knowledge graphs, but their construction cost is high and it is difficult to cover open-domain knowledge. Alternatively, iterative retrieval combined with reasoning loops can be used to gradually refine the query, but initial retrieval bias can easily cause the reasoning path to deviate from the correct direction, and static scoring is difficult to adapt to the real-time evaluation requirements of evidence relevance during dynamic reasoning.

[0003] It is evident that existing technologies suffer from several problems: statically fixed retrieval conditions, linear separation between retrieval and reasoning processes, inability to resolve conflicts between evidence from different sources in multi-hop reasoning, resulting in large deviations in retrieval direction, broken logical chains, and insufficient dynamic assessment of evidence credibility in complex reasoning data retrieval. Summary of the Invention

[0004] In view of this, the present disclosure provides a data retrieval method, apparatus, electronic device, and readable storage medium to solve the problems in the prior art where static and fixed retrieval conditions lead to a linear separation between the retrieval and reasoning stages, and the inability to resolve conflicts between evidence from different sources in multi-hop reasoning, resulting in large deviations in retrieval direction, broken logical chains, and insufficient dynamic evaluation of evidence credibility in complex reasoning data retrieval.

[0005] A first aspect of this disclosure provides a data retrieval method, comprising: performing intent anchoring processing on data to be queried to obtain initial anchor data, inference link data, and search keyword data; performing targeted recall processing on preset index data based on the search keyword data and inference link data to obtain valid evidence data; performing inference generation processing on the valid evidence data to obtain intermediate conclusion data; performing collaborative calibration processing on the intermediate conclusion data, initial anchor data, and inference link data to obtain anchor calibration data and dynamic inference chain data; and performing convergence and aggregation processing on the dynamic inference chain data and anchor calibration data to obtain target retrieval result data.

[0006] In some embodiments, intermediate conclusion data, initial anchor data, and inference link data are subjected to collaborative calibration processing to obtain anchor calibration data and dynamic inference link data, including: performing semantic matching processing on intermediate conclusion data and initial anchor data to obtain semantic matching data; performing anchor supplementation processing on intermediate conclusion data based on semantic matching data to obtain anchor calibration data; and performing chain evolution processing on intermediate conclusion data, inference link data, and anchor calibration data to obtain dynamic inference link data.

[0007] In some embodiments, targeted recall processing of preset index data is performed based on retrieval keyword data and inference link data to obtain valid evidence data, including: performing index retrieval processing on preset index data based on retrieval keyword data to obtain candidate text block data; performing credibility filtering processing on candidate text block data to obtain text block filtered data; and performing stage matching processing on text block filtered data based on inference link data to obtain valid evidence data.

[0008] In some embodiments, the process of generating intermediate conclusion data by reasoning from valid evidence data includes: performing evidence analysis on the valid evidence data to obtain structured evidence data; performing logical deduction on the structured evidence data to obtain deduction result data; and performing conclusion solidification on the deduction result data to obtain intermediate conclusion data.

[0009] In some embodiments, the dynamic inference chain data and anchor calibration data are converged and aggregated to obtain target retrieval result data, including: performing link integrity detection processing on the dynamic inference chain data to obtain link status data; performing steady-state detection processing on the anchor calibration data to obtain anchor status data; and performing result aggregation processing on the link status data and anchor status data to obtain target retrieval result data.

[0010] In some embodiments, intent anchoring processing is performed on the query data to obtain initial anchor data, inference link data, and search keyword data, including: query parsing processing of the query data to obtain query element data; anchor construction processing of the query element data to obtain initial anchor data; and link prediction processing of the initial anchor data to obtain inference link data and search keyword data.

[0011] In some embodiments, intermediate conclusion data, inference link data, and anchor point calibration data are subjected to chain-like evolution processing to obtain dynamic inference chain data, including: performing logical link processing on intermediate conclusion data and anchor point calibration data to obtain logical path data; performing contradiction verification processing on logical path data to obtain path verification data; and performing branch convergence processing on path verification data based on inference link data to obtain dynamic inference chain data.

[0012] A second aspect of this disclosure provides a data retrieval apparatus, comprising: a first processing module for performing intent anchoring processing on query data to obtain initial anchor data, inference link data, and search keyword data; a second processing module for performing targeted recall processing on preset index data based on the search keyword data and inference link data to obtain valid evidence data; a third processing module for performing inference generation processing on the valid evidence data to obtain intermediate conclusion data; a fourth processing module for performing collaborative calibration processing on the intermediate conclusion data, initial anchor data, and inference link data to obtain anchor calibration data and dynamic inference chain data; and a fifth processing module for performing convergence and aggregation processing on the dynamic inference chain data and anchor calibration data to obtain target retrieval result data.

[0013] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0014] A fourth aspect of this disclosure provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0015] The beneficial effects of this disclosed embodiment compared with the prior art are as follows: By performing intent anchoring processing on the query data, initial anchor data, inference link data, and search keyword data are obtained; based on the search keyword data and inference link data, targeted recall processing is performed on the preset index data to obtain valid evidence data; inference generation processing is performed on the valid evidence data to obtain intermediate conclusion data; the intermediate conclusion data, initial anchor data, and inference link data are collaboratively calibrated to obtain anchor calibration data and dynamic inference chain data; the dynamic inference chain data and anchor calibration data are converged and aggregated to obtain target search result data. This improves the positioning accuracy in the initial stage of retrieval, enhances the relevance and accuracy of evidence recall, strengthens the inference depth of the retrieval process, improves the adaptive adjustment capability of the retrieval direction and the coherence of the evidence chain in multiple iterations, and improves the completeness and reliability of the search results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a schematic diagram illustrating an application scenario of an embodiment of this disclosure; Figure 2 This is a schematic flowchart of a data retrieval method provided in an embodiment of this disclosure; Figure 3 This is a flowchart illustrating another data retrieval method provided in this embodiment of the disclosure; Figure 4 This is a schematic diagram of the structure of a data retrieval device provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0018] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0019] It should be noted that the user information (including but not limited to terminal device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0020] A data retrieval method and apparatus according to embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram illustrating an application scenario of an embodiment of this disclosure. The application scenario may include terminal devices 1, 2, and 3, server 4, and network 5.

[0022] Terminal devices 1, 2, and 3 can be hardware or software. When terminal devices 1, 2, and 3 are hardware, they can be various electronic devices with displays and supporting communication with server 4, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 1, 2, and 3 are software, they can be installed in the aforementioned electronic devices. Terminal devices 1, 2, and 3 can be implemented as multiple software programs or software modules, or as a single software program or software module; this disclosure does not limit this. Furthermore, various applications can be installed on terminal devices 1, 2, and 3, such as data processing applications, instant messaging tools, social platform software, search applications, shopping applications, etc.

[0023] Server 4 can be a server that provides various services, such as a backend server that receives requests sent by terminal devices with which it has established communication connections. This backend server can receive and analyze the requests sent by the terminal devices and generate processing results. Server 4 can be a single server, a server cluster consisting of several servers, or a cloud computing service center. This disclosure embodiment does not limit this.

[0024] It should be noted that server 4 can be either hardware or software. When server 4 is hardware, it can be various electronic devices that provide various services to terminal devices 1, 2, and 3. When server 4 is software, it can be multiple software programs or software modules that provide various services to terminal devices 1, 2, and 3, or it can be a single software program or software module that provides various services to terminal devices 1, 2, and 3. This disclosure does not limit the scope of the embodiments.

[0025] Network 5 can be a wired network using coaxial cable, twisted pair, and fiber optic connection, or it can be a wireless network that enables interconnection of various communication devices without wiring, such as Bluetooth, Near Field Communication (NFC), and Infrared. This disclosure does not limit the scope of the network.

[0026] Users can establish a communication connection with server 4 via network 5 through terminal devices 1, 2, and 3 to receive or send information. Specifically, server 4 can obtain the data to be queried through terminal devices 1, 2, and 3, and obtain initial anchor data, inference link data, and search keyword data by performing intent anchoring processing on the data to be queried; perform targeted recall processing on preset index data based on search keyword data and inference link data to obtain valid evidence data; perform inference generation processing on valid evidence data to obtain intermediate conclusion data; perform collaborative calibration processing on intermediate conclusion data, initial anchor data, and inference link data to obtain anchor calibration data and dynamic inference chain data; and perform convergence and aggregation processing on dynamic inference chain data and anchor calibration data to obtain target search result data.

[0027] It should be noted that the specific types, quantities, and combinations of terminal devices 1, 2, and 3, server 4, and network 5 can be adjusted according to the actual needs of the application scenario, and this disclosure embodiment does not impose any restrictions on this.

[0028] Figure 2 This is a schematic flowchart of a data retrieval method provided in an embodiment of this disclosure. Figure 2 Data retrieval methods can be derived from Figure 1 The server executes the command. For example... Figure 2 As shown, the data retrieval method includes: S201, perform intent anchoring processing on the query data to obtain initial anchor data, inference link data and search keyword data.

[0029] Specifically, the data to be queried can be the user's input question or instruction text, which can be used to trigger the entire data retrieval process as the starting point for intent anchoring processing; the data to be queried can include the user's original input content such as the question description and search target in natural language form; the data to be queried can be received through the input interface or obtained through the user interaction interface.

[0030] Furthermore, intent anchoring can be a data processing procedure that transforms unstructured user queries into structured retrieval guidance data. This intent anchoring can be used to parse and predict the query data to extract core semantic elements and plan the retrieval path. Based on preset parsing rules, this intent anchoring can extract semantic elements and predict inference paths from the query data to obtain initial anchor data, inference link data, and retrieval keyword data.

[0031] The initial anchor data can be structured data representing core semantic elements extracted from the data to be queried. This initial anchor data can serve as the initial basis and calibration benchmark for subsequent retrieval and reasoning, and can be used to anchor the retrieval target and provide semantic constraints. The initial anchor data can contain structured elements such as entity identifiers, relationship types, and attribute descriptions. The initial anchor data can be the structured output obtained after extracting semantic elements from the data to be queried through intent anchoring processing.

[0032] Inference link data can be the predicted logical path data required to reach the retrieval target from the initial anchor point. This inference link data can be used to pre-plan the macro path of retrieval and inference, guiding the evidence screening and intermediate conclusion generation direction of staged retrieval. This inference link data can contain path description information such as the sequence of inference stages and the logical jump relationship between stages. This inference link data can be the path planning output obtained after the intent of the query data is predicted through intent anchoring processing.

[0033] The search keyword data can be a set of terms that can be used for index queries, which is transformed from the initial anchor data. This search keyword data can be used to recall candidate evidence related to the current reasoning step from the index data. The search keyword data can include query terms such as entity names, relation descriptions, and attribute keywords. The search keyword data can also be query input data obtained by expanding the initial anchor data with keywords through intent anchoring processing.

[0034] Furthermore, the intent anchoring process can specifically include parsing the query data to obtain query element data; constructing anchor points on the query element data to obtain initial anchor point data; and performing link prediction on the initial anchor point data to obtain inference link data and search keyword data.

[0035] This application's embodiments transform unstructured user queries into structured search guidance data through intent anchoring processing, achieving precise anchoring of query intent; provide clear semantic constraints and calibration benchmarks for subsequent searches through initial anchor data; pre-plan search paths through inference link data to avoid blind searches; and provide direct query input for index retrieval through search keyword data. This improves the positioning accuracy in the initial stage of retrieval, enhances the adaptability of search direction to query intent, and improves the coherence and goal-orientation of subsequent search and inference processes.

[0036] For example, in intelligent question answering, the query data could be "What art is another poet, a friend of a certain poet, good at?" The query data can be processed to anchor intent, and query parsing can be performed to identify query element data such as the poet entity and the friend relationship. Then, anchor point construction processing can be performed on the query element data to extract initial anchor point data containing the poet entity and the friend relationship. Link prediction processing can be performed on the initial anchor point data to predict the reasoning link data from confirming the friend relationship to finding the area of ​​expertise. At the same time, search keyword data containing relevant poet names and relationship descriptions can be generated to guide the subsequent targeted recall processing to retrieve relevant evidence from the index data.

[0037] S202, based on the search keyword data and reasoning link data, perform targeted recall processing on the preset index data to obtain valid evidence data.

[0038] Specifically, the preset index data can be a pre-built and stored collection containing text block data and its associated structured data and vector representation data. This preset index data can serve as a retrieval source for targeted recall processing to support online retrieval operations. The preset index data can contain associated information such as text block data, triple data, vector representation data, and / or scoring data. The preset index data can be structured storage data obtained by cleaning, segmenting, structurally annotating, vectorizing, and scoring the original document collection through offline index building processing.

[0039] Furthermore, targeted recall processing can be a data processing procedure that performs targeted and scope-controlled retrieval and filtering from index data based on retrieval intent and stage constraints. This targeted recall processing can be used to match and filter preset index data by combining retrieval keyword data and inference link data. This targeted recall processing can perform index matching on preset index data based on retrieval keyword data, and perform stage relevance filtering on the matching results based on inference link data to obtain valid evidence data.

[0040] Among them, the effective evidence data can be a set of text fragments that are relevant to the current reasoning stage and meet the quality standards, which are obtained through targeted recall processing. This effective evidence data can be used to support the generation of intermediate conclusions. The effective evidence data can include text fragments that are semantically related to the current reasoning stage and their associated credibility data. The effective evidence data can be the filtered output data obtained after index retrieval, credibility filtering and stage matching of the preset index data through targeted recall processing.

[0041] Furthermore, the targeted recall process may specifically include performing index retrieval processing on preset index data based on search keyword data to obtain candidate text block data; performing credibility filtering processing on the candidate text block data to obtain text block filtered data; and performing stage matching processing on the text block filtered data based on inference link data to obtain valid evidence data.

[0042] This application embodiment obtains candidate text block data by indexing the index data based on search keyword data, obtains text block filtered data by credibility filtering of the candidate text block data, and obtains valid evidence data by performing stage matching on the text block filtered data based on inference link data. This enhances the collaborative screening capability of index retrieval and inference link stage constraints, improves the quality filtering and relevance matching accuracy of candidate evidence, and provides a structured, high-quality evidence foundation for the generation of subsequent intermediate conclusions.

[0043] For example, in knowledge-based question answering, based on search keyword data containing poet entities and friend relationships, and inference link data indicating the process from confirming friend relationships to finding areas of expertise, the preset index data can be indexed and retrieved based on the search keyword data to recall candidate text blocks containing relevant poet names and friend relationships. Then, the candidate text block data can be filtered based on credibility, and the filtered text block data can be obtained by filtering based on dimensions such as source credibility and semantic integrity. The filtered text block data can be matched in stages based on the inference link data to extract text fragments that explicitly mention the confirmation of friend relationships as valid evidence data.

[0044] S203, perform reasoning and generation processing on valid evidence data to obtain intermediate conclusion data.

[0045] Specifically, the reasoning generation process can be a data processing procedure that performs semantic understanding and logical integration based on valid evidence data to generate structured reasoning conclusions. This reasoning generation process can be used to perform in-depth analysis and logical deduction on valid evidence data, and output intermediate conclusion data that represents the new facts or relationships derived from the current reasoning step. This reasoning generation process can also perform semantic analysis and logical deduction on valid evidence data based on preset reasoning rules to obtain intermediate conclusion data.

[0046] Among them, intermediate conclusion data can be structured data generated based on valid evidence data, representing new facts or relationships derived from the current reasoning step. This intermediate conclusion data can be used to trigger anchor point calibration and search keyword updates. This intermediate conclusion data can include structured natural language expressions such as entity relationship statements and attribute inference results. This intermediate conclusion data can be reasoning output data obtained after semantic understanding and logical integration of valid evidence data through reasoning generation processing.

[0047] Furthermore, the reasoning generation process can specifically include performing evidence analysis on valid evidence data to obtain structured evidence data; performing logical deduction on the structured evidence data to obtain deduction result data; and performing conclusion solidification on the deduction result data to obtain intermediate conclusion data.

[0048] This application embodiment enhances the ability to transform valid evidence into structured evidence data by parsing valid evidence data; it also improves the collaborative efficiency of logical deduction and conclusion solidification; and it provides structured intermediate reasoning results for subsequent anchor point calibration and keyword updates.

[0049] For example, in knowledge-based question answering, valid evidence data can be processed through evidence analysis to identify structured evidence data about the relationship between Person A and Person B in the text fragment; then, logical deduction can be performed on the structured evidence data to deduce the deduction result data that Person A and Person B have a friendship relationship by combining multiple interaction records; the deduction result data can be processed to solidify the conclusion and generate standardized intermediate conclusion data, indicating that Person A's friends include Person B.

[0050] S204, perform collaborative calibration processing on intermediate conclusion data, initial anchor point data and inference chain data to obtain anchor point calibration data and dynamic inference chain data.

[0051] Specifically, collaborative calibration processing can be a data processing procedure that dynamically calibrates and evolves the initial anchor data and inference chain data based on intermediate conclusion data. This collaborative calibration processing can be used to evaluate the semantic consistency between intermediate conclusions and initial query intent, and supplement new anchors to expand the inference chain when the consistency is insufficient. This collaborative calibration processing can perform semantic matching on intermediate conclusion data and initial anchor data based on preset matching rules, and dynamically supplement the anchor set based on the matching results to obtain anchor calibration data and dynamic inference chain data.

[0052] The anchor calibration data can be a set of anchor points that has been dynamically adjusted and expanded through collaborative calibration processing. This anchor calibration data can be used to represent the core entities and relationships that are activated and associated during the reasoning process. The anchor calibration data can include associated anchor points such as new entities or attributes identified in the initial anchor data and intermediate conclusion data. The anchor calibration data can be a dynamically updated output obtained by performing semantic matching evaluation and anchor point supplementation on the initial anchor data through collaborative calibration processing.

[0053] Among them, the dynamic inference chain data can be structured data that records all inference steps and their logical relationships from the initial query to the retrieval target through collaborative calibration processing. This dynamic inference chain data can be used to fully present the derivation process of the retrieval results. This dynamic inference chain data can contain chain structure information such as logical path nodes and inference relationships between nodes. This dynamic inference chain data can be the evolutionary output obtained after linking intermediate conclusion data to the inference chain and performing conflict verification through collaborative calibration processing.

[0054] Furthermore, the collaborative calibration process may specifically include performing semantic matching processing on intermediate conclusion data and initial anchor point data to obtain semantic matching data; performing anchor point supplementation processing on intermediate conclusion data based on semantic matching data to obtain anchor point calibration data; and performing chain evolution processing on intermediate conclusion data, inference link data and anchor point calibration data to obtain dynamic inference chain data.

[0055] This application embodiment obtains semantically matched data by semantically matching intermediate conclusion data and initial anchor data; anchor points are added to the intermediate conclusion data based on the semantically matched data to obtain anchor calibration data; and the intermediate conclusion data, inference link data, and anchor calibration data are subjected to chain evolution and conflict verification to obtain dynamic inference chain data. This enhances the semantic consistency assessment capability between intermediate conclusions and initial query intent, improves the collaborative efficiency of dynamic anchor point addition and inference chain evolution during the inference process, and provides a structured traceability basis for the complete derivation process of retrieval results.

[0056] For example, in knowledge-based question answering, semantic matching processing can be performed on intermediate conclusion data and initial anchor data to evaluate the degree of matching between Person B as a new entity and the initial query intent, thus obtaining semantic matching data. Then, anchor supplementation processing can be performed on intermediate conclusion data based on semantic matching data, such as adding Person B as an associated anchor to obtain anchor calibration data. Intermediate conclusion data, inference link data, and anchor calibration data can be processed in a chain-like evolution, such as linking the friend relationship from Person A to Person B to the inference chain and verifying path consistency to obtain dynamic inference chain data.

[0057] S205, the dynamic inference chain data and anchor point calibration data are converged and aggregated to obtain the target retrieval result data.

[0058] Specifically, convergence and aggregation processing can be a data processing procedure that focuses on paths, resolves conflicts, and integrates results based on dynamic inference chain data and anchor calibration data. This convergence and aggregation processing can be used to filter and converge logical paths generated by multiple rounds of inference, and aggregate the intent guidance provided by anchor calibration data to generate target retrieval results. This convergence and aggregation processing can perform path parsing and branch evaluation on dynamic inference chain data based on preset aggregation rules, and perform semantic relevance verification based on anchor calibration data to obtain target retrieval result data.

[0059] Specifically, the target retrieval result data can be structured data representing the retrieval answer and its supporting evidence chain, which is integrated through convergence aggregation processing. This target retrieval result data can be used to present the complete derivation result from the initial query to the retrieval target. The target retrieval result data can include the target answer statement and structured information such as the reasoning path and evidence association supporting the answer. The target retrieval result data can be obtained by focusing the path and resolving conflicts of the dynamic reasoning chain data through convergence aggregation processing, and aggregating the anchor point calibration data to ensure its intent orientation.

[0060] Furthermore, the convergence and aggregation process may specifically include performing link integrity detection on the dynamic inference chain data to obtain link status data; performing steady-state detection on the anchor point calibration data to obtain anchor point status data; and aggregating the link status data and anchor point status data to obtain target retrieval result data.

[0061] This application embodiment obtains link status data by performing link integrity detection on dynamic inference chain data; obtains anchor point status data by performing steady-state detection on anchor point calibration data; aggregates the link status data and anchor point status data, prunes off-branch and resolves conflicts, and integrates the main path inference nodes and associated evidence to obtain target retrieval result data. This enhances the focusing convergence and conflict resolution capabilities of multi-round inference paths; improves the calibration accuracy of anchor point intent guidance for target retrieval results; and provides a structured presentation foundation for the retrieval answer and its supporting evidence chain.

[0062] For example, in knowledge-based question answering, the dynamic reasoning chain data can be processed to detect the integrity of the chain, identifying the main path covering the search target from the initial query and the branch paths deviating from the anchor point semantics to obtain the chain state data; then, the anchor point calibration data can be processed to detect the steady state, evaluating the convergence of the anchor point set to obtain the anchor point state data; the chain state data and anchor point state data can be aggregated, deviating branches can be pruned and path conflicts can be resolved, and then the reasoning nodes and associated evidence on the main path can be integrated to obtain the target retrieval result data.

[0063] According to the technical solution provided in this disclosure, intent anchoring processing is performed on the data to be queried. Initial anchor data, inference link data, and search keyword data are obtained through query parsing, anchor point construction, and link prediction. Based on the search keyword data and inference link data, targeted recall processing is performed on preset index data. Valid evidence data is obtained through index retrieval, credibility filtering, and stage matching. Inference generation processing is performed on the valid evidence data. Intermediate conclusion data is obtained through evidence parsing, logical deduction, and conclusion solidification. The intermediate conclusion data, initial anchor data, and inference link data are then collaboratively calibrated. Through semantic matching, anchor point supplementation, and chain evolution, anchor point calibration data and dynamic inference chain data are obtained. These two data are then converged and aggregated. Through link integrity detection, anchor point steady-state detection, and result aggregation, target retrieval results are obtained. This enhances the ability to guide unstructured queries towards structured retrieval; improves the collaborative filtering accuracy of index retrieval and inference chain stage constraints; enhances the ability to transform valid evidence into structured inference conclusions; improves the collaborative efficiency of dynamic anchor point supplementation and inference chain evolution during the inference process; and enhances the focusing convergence and conflict resolution capabilities of multi-round inference paths.

[0064] In some embodiments, intermediate conclusion data, initial anchor data, and inference link data are subjected to collaborative calibration processing to obtain anchor calibration data and dynamic inference link data, including: performing semantic matching processing on intermediate conclusion data and initial anchor data to obtain semantic matching data; performing anchor supplementation processing on intermediate conclusion data based on semantic matching data to obtain anchor calibration data; and performing chain evolution processing on intermediate conclusion data, inference link data, and anchor calibration data to obtain dynamic inference link data.

[0065] Specifically, semantic matching processing can be a data processing process that calculates the semantic relevance and assesses the consistency of two types of data based on preset matching rules. This semantic matching processing can be used to quantify the degree of semantic deviation between intermediate conclusion data and initial anchor data, providing a basis for decision-making for subsequent anchor supplementation. This semantic matching processing can also perform vector mapping and similarity calculation on intermediate conclusion data and initial anchor data based on preset semantic association rules to obtain semantic matching data.

[0066] The semantic matching data can be a quantitative result data that represents the degree of semantic association between intermediate conclusion data and initial anchor data, obtained through semantic matching processing. This semantic matching data can be used to trigger dynamic adjustment of the anchor set. The semantic matching data can include evaluation information such as semantic association degree value and deviation status identifier. The semantic matching data can be a quantitative evaluation result obtained by performing semantic association calculation on intermediate conclusion data and initial anchor data.

[0067] Furthermore, the semantic matching process may specifically include extracting semantic elements from intermediate conclusion data to obtain conclusion element data; extracting anchor elements from initial anchor data to obtain anchor element data; and performing association mapping between the conclusion element data and the anchor element data to obtain semantic matching data.

[0068] Furthermore, anchor point supplementation processing can be a data processing procedure that supplements the anchor point set with new entities or attributes identified in the intermediate conclusion data based on the semantic matching results. This anchor point supplementation processing can be used to include newly discovered semantic elements into the anchor point set when the semantic matching degree is lower than a preset threshold, thereby realizing the dynamic expansion of anchor points. This anchor point supplementation processing can perform threshold comparison and new element identification on the semantic matching data based on preset supplementation rules to obtain anchor point calibration data.

[0069] Furthermore, the anchor point supplementation process may specifically include performing threshold comparison processing on semantic matching data to obtain deviation state data; performing new feature identification processing on intermediate conclusion data based on the deviation state data to obtain new feature data; and performing set merging processing on the new feature data and the initial anchor point data to obtain anchor point calibration data.

[0070] Furthermore, chain evolution processing can be a data processing procedure that links intermediate conclusion data to the inference chain and performs conflict verification based on a pre-defined logical framework of inference link data. This chain evolution processing can be used to connect discrete intermediate conclusions into a coherent inference path and resolve potential conflicts. This chain evolution processing can also insert nodes and establish logical edges in intermediate conclusion data based on the stage sequence of inference link data to obtain dynamic inference chain data.

[0071] Furthermore, the chain-like evolution processing can specifically include performing stage parsing processing on the inference link data to obtain stage sequence data; performing node insertion processing on the intermediate conclusion data based on the stage sequence data to obtain chain node data; and performing conflict verification processing on the chain node data based on the anchor calibration data to obtain dynamic inference chain data.

[0072] For example, in knowledge-based question answering, semantic matching can be performed based on intermediate conclusion data representing Person A's friends, including Person B, and initial anchor data containing Person A and their friend relationships. The new entity Person B from the intermediate conclusion can be extracted and semantically associated with Person A from the initial anchor, resulting in semantic matching data representing the degree of association between Person B as a new entity and the initial query intent. Furthermore, anchor supplementation processing can be performed on the intermediate conclusion data based on this semantic matching data. Since Person B is a new entity not included in the initial anchor, it is added as an associated anchor, resulting in anchor calibration data containing both Person A and Person B. The intermediate conclusion data, the inference chain data indicating the process from confirming friend relationships to finding areas of expertise, and the anchor calibration data can be processed in a chain-like evolution. The friend relationship from Person A to Person B can be inserted as a new node into the inference chain, and path consistency can be verified, resulting in dynamic inference chain data recording the complete inference steps.

[0073] According to the technical solution provided in this disclosure, semantic matching processing is performed on intermediate conclusion data and initial anchor point data. Semantic matching data is obtained through conclusion element extraction, anchor point element extraction, and association mapping. Anchor point supplementation processing is performed on intermediate conclusion data based on semantic matching data. Anchor point calibration data is obtained through threshold comparison, new element identification, and set merging. Intermediate conclusion data, inference link data, and anchor point calibration data are subjected to chain evolution processing. Dynamic inference chain data is obtained through stage parsing, node insertion, and conflict verification. This enhances the quantitative ability of semantic consistency assessment between intermediate conclusions and initial anchor points; improves the accuracy of dynamic supplementation of new anchor points during inference; and realizes automatic linking and conflict resolution between intermediate conclusions and inference chains.

[0074] In some embodiments, targeted recall processing of preset index data is performed based on retrieval keyword data and inference link data to obtain valid evidence data, including: performing index retrieval processing on preset index data based on retrieval keyword data to obtain candidate text block data; performing credibility filtering processing on candidate text block data to obtain text block filtered data; and performing stage matching processing on text block filtered data based on inference link data to obtain valid evidence data.

[0075] Specifically, index retrieval processing can be a data processing procedure that performs matching searches in index data based on query terms to recall relevant candidate text blocks. This index retrieval processing can be used to filter a set of text fragments that may be related to the current search keywords from batch index data. This index retrieval processing can also perform entity matching and relation mapping on the search keyword data based on a preset index structure to obtain candidate text block data.

[0076] Among them, the candidate text block data can be a set of text fragments containing relevant information that are initially recalled from the preset index data through index retrieval processing. The candidate text block data can be used to provide the basis for candidate evidence to be screened. The candidate text block data can contain multiple text fragments and their associated metadata information. The candidate text block data can be the preliminary recall result obtained after matching and searching the search keyword data through index retrieval processing.

[0077] Furthermore, the index retrieval process may specifically include performing term parsing on the search keyword data to obtain query element data; performing entity matching on the preset index data based on the query element data to obtain entity matching data; performing relation mapping on the preset index data based on the query element data to obtain relation mapping data; and merging and deduplicating the entity matching data and the relation mapping data to obtain candidate text block data.

[0078] In addition, credibility filtering can be a data processing procedure that performs quality screening on candidate text blocks based on preset credibility assessment rules. This credibility filtering can be used to remove low-quality and highly redundant noise information from the initially recalled candidate text blocks and retain evidence candidates that meet the credibility standards. This credibility filtering can also evaluate the source authority and verify the semantic integrity of candidate text block data based on preset scoring dimensions to obtain filtered text block data.

[0079] The text block filtering data can be a set of text fragments that meet the credibility standard and are semantically complete after credibility filtering. This text block filtering data can be used to provide qualified evidence candidates. The text block filtering data can include text fragments that have passed the credibility assessment and their associated score information. The text block filtering data can be the filtered output obtained by performing multi-dimensional quality screening on the candidate text block data.

[0080] Furthermore, the credibility filtering process may specifically include: evaluating the source authority of candidate text block data to obtain source score data; performing semantic integrity verification on the candidate text block data to obtain integrity score data; and performing comprehensive filtering on the source score data and the integrity score data to obtain text block filtered data.

[0081] Furthermore, the stage matching process can be a data processing procedure that performs relevance screening on filtered text blocks based on the current inference stage indicated by the inference link data. This stage matching process can be used to extract evidence fragments directly related to the goal of the current inference stage from qualified evidence candidates. This stage matching process can also perform semantic alignment and relevance determination on the filtered text block data based on the stage goal description in the inference link data to obtain valid evidence data.

[0082] Furthermore, the stage matching process may specifically include performing stage target parsing on the inference link data to obtain stage target data; performing semantic alignment on the text block filtering data based on the stage target data to obtain alignment result data; and performing relevance determination on the alignment result data to obtain valid evidence data.

[0083] For example, in knowledge-based question answering, multiple text fragments containing relevant poet names and descriptions of interactions can be recalled as candidate text blocks through entity matching and relationship mapping. These candidate text blocks can then undergo credibility filtering, where low-quality fragments can be eliminated based on dimensions such as source authority and semantic integrity, resulting in filtered text block data. Furthermore, the filtered text block data can be subjected to stage matching based on inference link data indicating the current stage as confirming a friendship relationship, extracting text fragments that clearly describe the friendship between Person A and Person B as valid evidence data.

[0084] According to the technical solution provided in this disclosure, candidate text block data is obtained through term parsing, entity matching, relation mapping, and deduplication. The candidate text block data undergoes credibility filtering, and filtered text block data is obtained through source authority assessment, semantic integrity verification, and comprehensive screening. Based on inference link data, the filtered text block data undergoes stage matching processing. Through stage target parsing, semantic alignment, and relevance determination, valid evidence data is obtained. This enhances the collaborative screening capability of index retrieval and inference link stage constraints; improves the quality filtering and relevance matching accuracy of candidate evidence; and provides a structured, high-quality evidentiary foundation for the generation of subsequent intermediate conclusions.

[0085] In some embodiments, the process of generating intermediate conclusion data by reasoning from valid evidence data includes: performing evidence analysis on the valid evidence data to obtain structured evidence data; performing logical deduction on the structured evidence data to obtain deduction result data; and performing conclusion solidification on the deduction result data to obtain intermediate conclusion data.

[0086] Specifically, evidence analysis processing can be a data processing process that extracts structured information and converts the format of unstructured text evidence. This evidence analysis processing can be used to transform text-based evidence data into machine-readable and logically clear structured data. Based on preset analysis rules, this evidence analysis processing can perform entity recognition and relationship extraction on valid evidence data to obtain structured evidence data.

[0087] Among them, structured evidence data can be data that represents evidence information in a structured form and is obtained through evidence analysis and processing. This structured evidence data can be used to provide inference material with clear logical relationships. This structured evidence data can contain structured elements such as entity identifiers, relationship types, and attribute descriptions. This structured evidence data can be the format conversion output obtained after entity recognition and relationship extraction of valid evidence data.

[0088] Furthermore, the evidence analysis process may specifically include performing entity recognition processing on the valid evidence data to obtain entity identifier data; performing relation extraction processing on the valid evidence data to obtain relation type data; performing attribute extraction processing on the valid evidence data to obtain attribute description data; and performing structured assembly processing on the entity identifier data, the relation type data, and the attribute description data to obtain structured evidence data.

[0089] Furthermore, logical deduction processing can be a data processing procedure that performs logical operations and relationship deductions on structured evidence data based on preset deduction rules. This logical deduction processing can be used to derive new implicit conclusions based on existing structured evidence through logical chains. This logical deduction processing can also perform path deduction and fact inference on entities and relationships in structured evidence data based on preset deduction rules to obtain deduction result data.

[0090] Among them, the deduction result data can be new information or intermediate inference data derived from existing evidence through the representation obtained by logical deduction processing. The deduction result data can be used to provide deduction output to be standardized. The deduction result data can include deduction information such as fact inference results and relation deduction conclusions. The deduction result data can be the logical operation result obtained by applying preset deduction rules to structured evidence data and performing multi-step deduction.

[0091] Furthermore, the logical deduction process may specifically include performing path parsing on the structured evidence data to obtain reasoning path data; performing fact inference processing based on the reasoning path data to obtain inference result data; and performing relation deduction processing on the inference result data to obtain deduction result data.

[0092] In addition, the conclusion solidification process can be a data processing procedure that formats and standardizes the temporary results generated by logical deduction. This conclusion solidification process can be used to ensure the stability and usability of the deduction results. Based on preset solidification rules, the conclusion solidification process can perform format conversion and duplicate verification on the deduction result data to obtain intermediate conclusion data.

[0093] Furthermore, the conclusion solidification process may specifically include format conversion of the deduction result data to obtain standardized conclusion data; duplicate verification of the standardized conclusion data to obtain deduplicated conclusion data; and semantic association verification of the deduplicated conclusion data to obtain intermediate conclusion data.

[0094] For example, in knowledge-based question answering, text fragments containing descriptions of the social relationships of historical figures can be processed as valid evidence data for evidence analysis to identify two entities, Person A and Person B, and their relationship, resulting in structured evidence data containing entity identifiers and relationship types. This structured evidence data can then be logically deduced, and based on Person A and Person B's multiple trips together, a deduction result can be derived that they have a friendship. This deduction result data can then be processed to solidify the conclusion, converting the inference into a standardized entity relationship statement and performing semantic association verification to obtain intermediate conclusion data representing that Person A's friends include Person B.

[0095] According to the technical solution provided in this disclosure, structured evidence data is obtained through entity recognition, relation extraction, attribute extraction, and structured assembly. Logical deduction processing is then performed on the structured evidence data, and deduction result data is obtained through path parsing, fact inference, and relation deduction. Conclusion solidification processing is then performed on the deduction result data, and intermediate conclusion data is obtained through format conversion, duplication verification, and semantic association verification. This enhances the ability to transform unstructured textual evidence into structured deduction material; improves the automation level of logical inference based on evidence; and enhances the standardization and stability of deduction conclusions.

[0096] In some embodiments, the dynamic inference chain data and anchor calibration data are converged and aggregated to obtain target retrieval result data, including: performing link integrity detection processing on the dynamic inference chain data to obtain link status data; performing steady-state detection processing on the anchor calibration data to obtain anchor status data; and performing result aggregation processing on the link status data and anchor status data to obtain target retrieval result data.

[0097] Specifically, the link integrity detection process can be a data processing procedure that evaluates and analyzes the logical coherence and step completeness of the inference path. This link integrity detection process can be used to check whether there are logical breaks or missing key steps in the inference chain and to evaluate the reliability of the inference path. Based on preset detection rules, this link integrity detection process can perform adjacent node coherence verification and endpoint coverage determination on dynamic inference chain data to obtain link status data.

[0098] The link status data can be status indicator data representing the current integrity and reliability of the dynamic inference chain, obtained through link integrity detection processing. This link status data can be used to provide a basis for judging the quality of the inference path. The link status data can include evaluation information such as logical integrity identifier and step coverage status. The link status data can be the evaluation result obtained after performing adjacent node association verification and endpoint coverage detection on the dynamic inference chain data.

[0099] Furthermore, the link integrity detection process may specifically include performing node parsing processing on the dynamic inference chain data to obtain path node data; performing adjacent association verification processing on the path node data to obtain continuity data; performing endpoint coverage detection processing on the path node data to obtain coverage status data; and performing comprehensive evaluation processing on the continuity data and the coverage status data to obtain link status data.

[0100] Furthermore, steady-state detection processing can be a data processing procedure for evaluating and analyzing the stability and changing trends of the anchor point set. This steady-state detection processing can be used to determine whether the anchor point set has become stable or is still changing in multiple iterations. Based on preset detection rules, this steady-state detection processing can perform change statistics and set similarity calculation on the anchor point calibration data to obtain anchor point state data.

[0101] Among them, the anchor point state data can be state index data representing the current stability of the anchor point calibration data obtained through steady-state detection processing. This anchor point state data can be used to provide a basis for judging the degree of focus of the query intent. The anchor point state data can include evaluation information such as stability indicators and descriptions of change trends. The anchor point state data can be the stability evaluation result obtained after performing multiple rounds of change statistics and set consistency evaluation on the anchor point calibration data.

[0102] Furthermore, the steady-state detection processing can specifically include performing historical backtracking processing on the anchor point calibration data to obtain historical anchor point data; performing change statistical processing on the historical anchor point data and the current anchor point calibration data to obtain change amount data; and performing stability determination processing on the change amount data to obtain anchor point state data.

[0103] Furthermore, the result aggregation processing can be a data processing procedure that comprehensively evaluates the quality of the inference path and the focus of the query intent based on link status data and anchor point status data, and generates target retrieval results. This result aggregation processing can be used to comprehensively evaluate the reliability and accuracy of the inference process and decide on the target retrieval results. This result aggregation processing can perform weighted aggregation and convergence determination on link status data and anchor point status data based on preset aggregation rules to obtain target retrieval result data.

[0104] Furthermore, the result aggregation process may specifically include performing path quality assessment on the link status data to obtain path score data; performing intent focus assessment on the anchor point status data to obtain focus score data; performing weighted comprehensive processing on the path score data and the focus score data to obtain comprehensive score data; and performing convergence determination processing based on the comprehensive score data to obtain target retrieval result data.

[0105] For example, in knowledge-based question answering, a link integrity check can be performed on dynamic reasoning chain data containing information from Person A to Person B's friend relationship and then to Person B's area of ​​expertise. This checks whether the entity associations between adjacent nodes are coherent and whether they cover the search target for the area of ​​expertise, resulting in link status data that indicates the quality of the reasoning path meets the standards. Then, steady-state detection can be performed on anchor point calibration data containing Person A, friend relationships, Person B, and their area of ​​expertise attributes. This involves statistically analyzing the changes in the anchor point set over multiple iterations and evaluating its stability, resulting in anchor point status data indicating that the anchor points have become stable. Finally, this link status data and anchor point status data can be aggregated, and by combining the path quality and anchor point stability evaluation results, target search result data containing Person B's expertise in calligraphy and related evidence can be generated.

[0106] According to the technical solution provided in this disclosure, the link integrity of the dynamic reasoning chain data is detected through node parsing, adjacent association verification, endpoint coverage detection, and comprehensive evaluation to obtain link status data. The anchor point calibration data is then subjected to steady-state detection processing, and anchor point status data is obtained through historical backtracking, change statistics, and stability determination. The link status data and anchor point status data are then aggregated, and the target retrieval result data is obtained through path quality evaluation, intent focus evaluation, weighted synthesis, and convergence determination. This enhances the collaborative evaluation capability of the logical coherence of the reasoning path and the stability of the anchor points; improves the comprehensive judgment accuracy of path quality and intent focus during the retrieval result generation process; and provides a structured presentation foundation for the retrieval answer and its supporting evidence chain.

[0107] In some embodiments, intent anchoring processing is performed on the query data to obtain initial anchor data, inference link data, and search keyword data, including: query parsing processing of the query data to obtain query element data; anchor construction processing of the query element data to obtain initial anchor data; and link prediction processing of the initial anchor data to obtain inference link data and search keyword data.

[0108] Specifically, query parsing can be a data processing process that performs structured analysis and element extraction on unstructured text queries. This query parsing can be used to transform the query data in natural language form into a set of structured elements that can be identified and processed in subsequent steps. Based on preset parsing rules, this query parsing can perform semantic analysis and key information extraction on the query data to obtain query element data.

[0109] Among them, the query element data can be a set of discrete information units extracted from the data to be queried through query parsing processing, which can be used to represent the core intent of the query. The query element data can be used to provide the basic semantic components for constructing the initial anchor point. The query element data can contain structured elements such as key entity identifiers, relationship descriptions and attribute features. The query element data can be the structured parsing output obtained after semantic analysis and key information extraction of the data to be queried.

[0110] Furthermore, the query parsing process may specifically include segmenting the query data to obtain word metadata; performing semantic annotation on the word metadata to obtain semantic tag data; and performing element extraction on the semantic tag data to obtain query element data.

[0111] Furthermore, the anchor point construction process can be a data processing procedure that identifies and constructs structured information nodes representing the core semantic elements of the query based on the query element data. This anchor point construction process can be used to transform discrete query elements into fixed reference points and starting points for the reasoning process. This anchor point construction process can perform entity recognition and relation assembly on the query element data based on preset construction rules to obtain initial anchor point data.

[0112] The initial anchor data can be structured data identified and constructed from the query element data through anchor construction processing, serving as the starting point for inference. This initial anchor data can be used to anchor the retrieval target and provide semantic constraints. The initial anchor data can contain structured elements such as entity identifiers, relationship types, and attribute descriptions. The initial anchor data can be a structured output obtained by performing entity identification and relationship assembly on the query element data.

[0113] Furthermore, the anchor point construction process may specifically include performing entity recognition processing on the query element data to obtain entity identifier data; performing relationship recognition processing on the query element data to obtain relationship type data; performing attribute recognition processing on the query element data to obtain attribute description data; and performing structured assembly processing on the entity identifier data, the relationship type data, and the attribute description data to obtain initial anchor point data.

[0114] Furthermore, the link prediction processing can be a data processing process that predicts the reasoning path and logical sequence from the starting point to the retrieval target based on the initial anchor data. This link prediction processing can be used to pre-plan the reasoning direction and provide a guiding framework for phased retrieval and dynamic evolution. Based on preset prediction rules, this link prediction processing can perform path prediction and phase division on the initial anchor data to obtain reasoning link data and retrieval keyword data.

[0115] Furthermore, the link prediction processing can specifically include performing path prediction processing on the initial anchor point data to obtain predicted path data; performing stage division processing on the predicted path data to obtain stage sequence data; performing keyword generation processing on the stage sequence data to obtain search keyword data; and performing association and encapsulation processing on the stage sequence data and the search keyword data to obtain inference link data.

[0116] For example, in a knowledge-based question-and-answer platform, when receiving user input for the query "What art is another poet, a friend of a certain poet, good at?", the system can perform query parsing processing on this data. Through word segmentation and semantic annotation, it can extract query element data such as poet entities, friend relationships, and areas of expertise. Then, it can perform anchor point construction processing on this query element data to identify the two poet entities and the friend relationship, and assemble them into initial anchor point data containing entity and relationship descriptions. Finally, it can perform link prediction processing on this initial anchor point data to predict the reasoning path from confirming the friend relationship to finding the area of ​​expertise, and generate retrieval keyword data containing relevant poet names and relationship descriptions, as well as reasoning link data indicating the sequence of reasoning stages.

[0117] According to the technical solution provided in this disclosure, query element data is obtained by parsing the query data, performing word segmentation, semantic annotation, and element extraction. Anchor point construction is then performed on the query element data, and initial anchor point data is obtained through entity recognition, relationship recognition, attribute recognition, and structured assembly. Link prediction processing is then performed on the initial anchor point data, and inference link data and retrieval keyword data are obtained through path prediction, stage division, keyword generation, and association encapsulation. This enhances the ability to transform unstructured natural language queries into structured semantic elements; improves the accuracy of identifying and constructing core query anchor points; and achieves the collaborative generation of inference path prediction and retrieval keywords.

[0118] In some embodiments, intermediate conclusion data, inference link data, and anchor point calibration data are subjected to chain-like evolution processing to obtain dynamic inference chain data, including: performing logical link processing on intermediate conclusion data and anchor point calibration data to obtain logical path data; performing contradiction verification processing on logical path data to obtain path verification data; and performing branch convergence processing on path verification data based on inference link data to obtain dynamic inference chain data.

[0119] Specifically, logical linking can be a data processing procedure that performs entity matching and relationship mapping on intermediate conclusion data and anchor calibration data based on preset linking rules to construct logical paths. This logical linking can be used to associate intermediate conclusions with calibrated anchors to construct a logical relationship network from the initial query to the current inference step. This logical linking can also match entities in intermediate conclusion data and anchor calibration data based on preset linking rules and establish connection relationships to obtain logical path data.

[0120] The logical path data can be a data structure representing the logical relationship between intermediate conclusions and calibrated anchor points, obtained through logical linking. This logical path data can be used to provide the logical path skeleton to be verified. The logical path data can contain path structure information such as node identifiers and connection relationships between nodes. The logical path data can be an associated output obtained by performing entity matching and relationship mapping on intermediate conclusion data and anchor point calibration data.

[0121] Furthermore, the logical link processing can specifically include performing node extraction processing on intermediate conclusion data to obtain conclusion node data; performing node extraction processing on anchor point calibration data to obtain anchor point node data; performing entity matching processing on the conclusion node data and the anchor point node data to obtain matching result data; and performing relation mapping processing based on the matching result data to obtain logical path data.

[0122] Furthermore, the contradiction verification process can be a data processing procedure that performs consistency checks and conflict identification on the information carried by different paths or nodes in the logical path based on preset verification rules. This contradiction verification process can be used to ensure the internal logical consistency of the logical path and eliminate information conflicts caused by different sources of evidence. This contradiction verification process can perform consistency comparison and conflict detection on the node information in the logical path data based on preset verification rules to obtain path verification data.

[0123] The path verification data can be logical path data marked with potential conflict points or resolved conflicts, obtained through conflict verification processing. This path verification data can be used to provide a set of paths that have passed preliminary purification and verification. The path verification data can include verification information such as path nodes and conflict status identifiers.

[0124] Furthermore, the conflict verification process may specifically include performing node parsing on the logical path data to obtain node information data; performing consistency comparison on the node information data to obtain comparison result data; performing conflict detection on the comparison result data to obtain conflict status data; and performing conflict resolution on the conflict status data to obtain path verification data.

[0125] In addition, branch convergence processing can be a data processing process that filters and prunes the validated paths based on a pre-set logical roadmap of inference link data. This branch convergence processing can be used to ensure that the generated inference chain conforms to the inference intent prediction and to remove inference branches that deviate from the core query intent. This branch convergence processing can also perform step alignment and branch filtering on the path validation data based on the stage sequence in the inference link data to obtain dynamic inference chain data.

[0126] Furthermore, the branch convergence process may specifically include performing stage parsing on the inference link data to obtain stage sequence data; performing step alignment processing on the path verification data based on the stage sequence data to obtain aligned path data; and performing branch filtering processing on the aligned path data to obtain dynamic inference link data.

[0127] For example, in knowledge-based question answering, logical linking can be performed based on intermediate conclusion data representing Person A's friends, including Person B, and anchor calibration data containing Person A, the friend relationship, and Person B. Person B can be used as a new node to establish a friend relationship with Person A, resulting in logical path data. Then, contradiction verification processing can be performed on this logical path data to check for conflicting information about Person B's attributes in the path and resolve it according to the credibility of the evidence, resulting in path verification data. Based on the reasoning link data indicating the process from confirming the friend relationship to finding the area of ​​expertise, branch convergence processing can be performed on this path verification data, retaining the main path that is highly relevant to the predicted intention and removing deviating branches, resulting in dynamic reasoning chain data that records the complete reasoning steps.

[0128] According to the technical solution provided in this disclosure, by logically linking intermediate conclusion data and anchor point calibration data, logical path data is obtained through node extraction, entity matching, and relationship mapping. The logical path data undergoes contradiction verification processing, and path verification data is obtained through node parsing, consistency comparison, conflict detection, and conflict resolution. Based on the inference link data, branch convergence processing is performed on the path verification data. Through stage parsing, step alignment, and branch filtering, dynamic inference chain data is obtained. This enhances the accuracy of constructing the logical association between intermediate conclusions and anchor points; improves the automation level of consistency verification and conflict resolution within the path; and achieves efficient alignment and convergence of inference branches with preset intentions.

[0129] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0130] Figure 3 This is a schematic diagram of another data retrieval method provided in an embodiment of this disclosure. For example... Figure 3 As shown, the data retrieval method includes: 1. Offline Index Building: This section aims to build an index structure that supports fine-grained retrieval and credibility assessment.

[0131] Data cleaning and segmentation: Input: Original document collection D = {d1, d2, ..., d...} n}, where d n This can be the nth original document input.

[0132] Processing: Remove HyperText Markup Language (HTML) tags, garbled characters, and other noise. Then, use a sliding window algorithm to segment the document into fixed-length text chunks, denoted as C = {c1, c2, ..., c...}. m}, where c m It can be the m-th fixed-length segment obtained from the original document.

[0133] Structured triplet annotation: Each text block c is annotated using an information extraction model (IEModel). i Perform entity relation extraction to generate a set of triples T. i ={(h,r,t)|h,t∈Entities,r∈Relations}.

[0134] For example, `(Li Bai, friend, Du Fu)` can be extracted from the text "Li Bai's best friend is Du Fu".

[0135] Vectorized encoding: using a pre-trained vector model for each text block c i Encode to obtain a vector representation v i ∈ d Where d can be a specific positive integer representing the dimension of the vector. It can be a real number.

[0136] Formula: v i =Encoder(c i ).

[0137] Construction of an evidence credibility scoring system: A scoring model S is constructed based on three dimensions. trust : Source credibility (S) source ): Authority score based on document source (0-1).

[0138] Semantic integrity (S) semantic ): Determines whether a text block contains a complete triple.

[0139] Redundancy (S) redundancy): Calculated based on information entropy or repetition rate.

[0140] Comprehensive scoring formula:

[0141] Where α, β, and γ can be weighting coefficients. The score for each text block is stored in the index.

[0142] Inverted index construction: A bidirectional index of "triple-text block" is established, which supports quick location of text block by triple and also supports reverse lookup of triple by text block, realizing the integration of structured and unstructured information.

[0143] 2. Online closed-loop reasoning: This stage completes multi-hop question answering through a closed-loop mechanism of "calibration-retrieval-reasoning-evolution".

[0144] Initial anchor point generation and inference intent prediction: Input: User query (data to be queried) Q.

[0145] deal with: Anchor point extraction: Extract the initial "entity-relationship-attribute" triples from Q as the anchor point set A0.

[0146] Intent prediction: Using a finely tuned lightweight large language model (LLM) to predict the inference logic link L intent (e.g., core entity) core ) to related entities related (to attribute).

[0147] Initialization: Initialize the intermediate conclusion set E= Dynamic inference chain RC= Search for the keyword set K0=Keywords(A0).

[0148] Phased search: Input: Current search keyword set K t .

[0149] deal with: Index retrieval: based on K t Retrieve the set of candidate text blocks C in the triplet-text block index. cand .

[0150] Credibility filtering: Filter out S trust (c) Text blocks greater than or equal to τ (e.g., 0.6) are considered valid evidence C. valid .

[0151] On-demand retrieval: Based on L intentAt the current stage, only evidence fragments relevant to the current reasoning process are retained to reduce noise.

[0152] Intermediate conclusion reasoning and retrieval feedback: Input: Current evidence C valid .

[0153] deal with: Intermediate conclusion generation: C valid Input a lightweight LLM and generate temporary intermediate conclusions e t (e.g., "Li Bai's friends included Du Fu").

[0154] Feedback mechanism: based on e t Update the next search keyword set K t+1 .

[0155] formula:

[0156] Output: The updated keyword set K t+1 Feedback is sent to the search module.

[0157] Anchor point real-time calibration: deal with: Calculate intermediate conclusion e t The semantic matching degree Sim(e) with the initial anchor point A0 t ,A0).

[0158] Dynamic calibration: If Sim(e t If A0) is less than δ (adaptive threshold), then e t The new entity or attribute in the context is added as the associated anchor point A. new .

[0159] Cyclic triggering: If the anchor point changes significantly, a new round of retrieval is triggered.

[0160] Dynamic inference chain evolution and conflict resolution: deal with: Chain evolution: e t Link to the reasoning chain RC to form a logical path from n to n+1 (such as "Li Bai to his friend to Du Fu" to the supplement "Du Fu to his expertise in calligraphy").

[0161] Conflict resolution: If there are different pieces of evidence c a c b Calculate the overall score for conflicting descriptions of the same fact:

[0162] Select the evidence with the highest score and discard conflicting items.

[0163] Branch pruning: based on Lintent Eliminate reasoning branches that deviate from the intended prediction.

[0164] Termination of judgment and answer generation: Termination conditions: Intermediate conclusions cover L intent The endpoint (e.g., finding the target answer).

[0165] Reaching the maximum number of reasoning steps T max (5 steps by default).

[0166] Generation: Input the final reasoning chain (RC) and the optimal evidence set into the LLM to generate the final answer and evidence credibility report (target retrieval result data).

[0167] According to the technical solution provided in this disclosure, by calibrating anchor points in real time based on intermediate inference results and combining lightweight models to predict inference logic links, the semantic drift problem in long-link multi-hop inference is effectively solved, ensuring that retrieval and inference always focus on the correct direction; improving the accuracy of evidence recall in multi-hop problems and reducing invalid retrieval; ensuring that the inference chain logic is rigorous and the facts are credible, and having high deployment feasibility.

[0168] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0169] Figure 4 This is a schematic diagram of a data retrieval device provided in an embodiment of this disclosure. Figure 4 As shown, the data retrieval device includes: The first processing module 401 is used to perform intent anchoring processing on the data to be queried to obtain initial anchor data, inference link data and search keyword data. The second processing module 402 is used to perform targeted recall processing on the preset index data based on the search keyword data and reasoning link data to obtain valid evidence data; The third processing module 403 is used to perform reasoning and generation processing on valid evidence data to obtain intermediate conclusion data. The fourth processing module 404 is used to perform collaborative calibration processing on intermediate conclusion data, initial anchor point data and inference chain data to obtain anchor point calibration data and dynamic inference chain data. The fifth processing module 405 is used to perform convergence and aggregation processing on the dynamic inference chain data and anchor point calibration data to obtain target retrieval result data.

[0170] According to the technical solution provided in this disclosure, intent anchoring processing is performed on the data to be queried. Initial anchor data, inference link data, and search keyword data are obtained through query parsing, anchor point construction, and link prediction. Based on the search keyword data and inference link data, targeted recall processing is performed on preset index data. Valid evidence data is obtained through index retrieval, credibility filtering, and stage matching. Inference generation processing is performed on the valid evidence data. Intermediate conclusion data is obtained through evidence parsing, logical deduction, and conclusion solidification. The intermediate conclusion data, initial anchor data, and inference link data are then collaboratively calibrated. Through semantic matching, anchor point supplementation, and chain evolution, anchor point calibration data and dynamic inference chain data are obtained. These two data are then converged and aggregated. Through link integrity detection, anchor point steady-state detection, and result aggregation, target retrieval results are obtained. This enhances the ability to guide unstructured queries towards structured retrieval; improves the collaborative filtering accuracy of index retrieval and inference chain stage constraints; enhances the ability to transform valid evidence into structured inference conclusions; improves the collaborative efficiency of dynamic anchor point supplementation and inference chain evolution during the inference process; and enhances the focusing convergence and conflict resolution capabilities of multi-round inference paths.

[0171] In some embodiments, the fourth processing module 404 is specifically used to: perform semantic matching processing on intermediate conclusion data and initial anchor data to obtain semantic matching data; perform anchor supplementation processing on intermediate conclusion data based on semantic matching data to obtain anchor calibration data; and perform chain evolution processing on intermediate conclusion data, inference link data and anchor calibration data to obtain dynamic inference chain data.

[0172] In some embodiments, the second processing module 402 is specifically used to: perform index retrieval processing on preset index data based on retrieval keyword data to obtain candidate text block data; perform credibility filtering processing on the candidate text block data to obtain text block filtered data; and perform stage matching processing on the text block filtered data based on inference link data to obtain valid evidence data.

[0173] In some embodiments, the third processing module 403 is specifically used to perform evidence parsing processing on the valid evidence data to obtain structured evidence data; perform logical deduction processing on the structured evidence data to obtain deduction result data; and perform conclusion solidification processing on the deduction result data to obtain intermediate conclusion data.

[0174] In some embodiments, the fifth processing module 405 is specifically used to perform link integrity detection processing on the dynamic inference chain data to obtain link status data; perform steady-state detection processing on the anchor point calibration data to obtain anchor point status data; and perform result aggregation processing on the link status data and anchor point status data to obtain target retrieval result data.

[0175] In some embodiments, the first processing module 401 is specifically used to perform query parsing processing on the query data to obtain query element data; perform anchor point construction processing on the query element data to obtain initial anchor point data; and perform link prediction processing on the initial anchor data to obtain inference link data and search keyword data.

[0176] In some embodiments, the intermediate conclusion data, inference link data, and anchor point calibration data are subjected to chain-like evolution processing to obtain dynamic inference chain data. Specifically, the intermediate conclusion data and anchor point calibration data are logically linked to obtain logical path data; the logical path data is subjected to contradiction verification processing to obtain path verification data; and the path verification data is subjected to branch convergence processing based on the inference link data to obtain dynamic inference chain data.

[0177] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure.

[0178] Figure 5 This is a schematic diagram of the electronic device 5 provided in an embodiment of this disclosure. Figure 5 As shown, the electronic device 5 of this embodiment includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the various method embodiments described above. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the various device embodiments described above.

[0179] Electronic device 5 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 5 may include, but is not limited to, processor 501 and memory 502. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 5 and does not constitute a limitation on electronic device 5. It may include more or fewer components than shown, or different components.

[0180] The processor 501 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0181] The memory 502 can be an internal storage unit of the electronic device 5, such as a hard disk or RAM of the electronic device 5. The memory 502 can also be an external storage device of the electronic device 5, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 5. The memory 502 can also include both internal and external storage units of the electronic device 5. The memory 502 is used to store computer programs and other programs and data required by the electronic device.

[0182] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0183] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0184] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A data retrieval method, characterized in that, include: The query data is subjected to intent anchoring processing to obtain initial anchor data, inference link data and search keyword data; Based on the search keyword data and the inference link data, the preset index data is targeted for recall to obtain valid evidence data; The valid evidence data is subjected to reasoning and generation processing to obtain intermediate conclusion data; The intermediate conclusion data, the initial anchor point data, and the inference chain data are subjected to collaborative calibration processing to obtain anchor point calibration data and dynamic inference chain data. The dynamic inference chain data and the anchor point calibration data are converged and aggregated to obtain the target retrieval result data.

2. The data retrieval method according to claim 1, characterized in that, The step of performing collaborative calibration processing on the intermediate conclusion data, the initial anchor point data, and the inference chain data to obtain anchor point calibration data and dynamic inference chain data includes: Semantic matching processing is performed on the intermediate conclusion data and the initial anchor point data to obtain semantic matching data; Based on the semantic matching data, the intermediate conclusion data is subjected to anchor point supplementation processing to obtain the anchor point calibration data. The intermediate conclusion data, the inference link data, and the anchor point calibration data are subjected to chain-like evolution processing to obtain the dynamic inference chain data.

3. The data retrieval method according to claim 1, characterized in that, The process of performing targeted recall on preset index data based on the search keyword data and the inference link data to obtain valid evidence data includes: Based on the search keyword data, the preset index data is indexed and retrieved to obtain candidate text block data; The candidate text block data is subjected to confidence filtering to obtain text block filtered data; Based on the inference link data, the text block filtering data is subjected to stage matching processing to obtain the valid evidence data.

4. The data retrieval method according to claim 1, characterized in that, The process of reasoning and generating intermediate conclusion data from the valid evidence data includes: The valid evidence data is subjected to evidence parsing processing to obtain structured evidence data; The structured evidence data is subjected to logical deduction processing to obtain deduction result data; The deduction results data are processed to solidify the conclusions, resulting in the intermediate conclusion data.

5. The data retrieval method according to claim 1, characterized in that, The process of converging and aggregating the dynamic inference chain data and the anchor point calibration data to obtain target retrieval result data includes: The dynamic inference chain data is subjected to link integrity detection processing to obtain link status data; The anchor point calibration data is processed by steady-state detection to obtain anchor point state data; The link status data and the anchor point status data are aggregated to obtain the target retrieval result data.

6. The data retrieval method according to claim 1, characterized in that, The intent anchoring process is performed on the query data to obtain initial anchor data, inference link data, and search keyword data, including: The query data is parsed to obtain the query element data; The query element data is processed to construct anchor points, resulting in the initial anchor point data; The initial anchor point data is processed by link prediction to obtain the inference link data and the search keyword data.

7. The data retrieval method according to claim 2, characterized in that, The step of performing chain-like evolution processing on the intermediate conclusion data, the inference link data, and the anchor point calibration data to obtain the dynamic inference chain data includes: The intermediate conclusion data and the anchor point calibration data are logically linked to obtain logical path data. The logical path data is subjected to contradiction verification processing to obtain path verification data; Based on the inference link data, the path verification data is processed by branch convergence to obtain the dynamic inference link data.

8. A data retrieval device, characterized in that, include: The first processing module is used to perform intent anchoring processing on the query data to obtain initial anchor data, inference link data and search keyword data; The second processing module is used to perform targeted recall processing on the preset index data based on the search keyword data and the inference link data to obtain valid evidence data; The third processing module is used to perform reasoning and generation processing on the valid evidence data to obtain intermediate conclusion data. The fourth processing module is used to perform collaborative calibration processing on the intermediate conclusion data, the initial anchor point data, and the inference chain data to obtain anchor point calibration data and dynamic inference chain data. The fifth processing module is used to perform convergence and aggregation processing on the dynamic inference chain data and the anchor point calibration data to obtain target retrieval result data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images