Intelligent question answering method and device based on knowledge graph and computing device

Abstract

The application provides a knowledge graph-based intelligent question answering method and device and a computing device. The method comprises: acquiring and identifying a user question and context information, and generating a structured question representation; performing knowledge graph path retrieval according to the structured question representation, and obtaining candidate graph data; performing semantic retrieval according to the structured question representation and the candidate graph data, and obtaining an answer statement slot and a candidate original text evidence corresponding to the user question; constructing a binding relationship between the candidate graph data, the answer statement slot and the candidate original text evidence, and generating a candidate evidence package; calling a large language model, inputting the structured question representation and the candidate evidence package into the large language model, and outputting an initial draft of an answer corresponding to the user question; splitting the initial draft of the answer based on the answer statement slot, obtaining at least one answer unit, and performing credible verification on the at least one answer unit, and obtaining a target answer corresponding to the user question, so as to form an answer with higher accuracy, interpretability and traceability.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent question-answering method based on knowledge graphs. This application also relates to an intelligent question-answering device based on knowledge graphs, and a computing device. Background Technology

[0002] With the development of large language models and retrieval-enhanced generation technologies, intelligent question-answering systems have been widely used for tasks such as knowledge retrieval, technical support, learning guidance, scientific research question-answering, encyclopedia question-answering, and professional knowledge assistance. Users typically expect the system to directly provide natural language answers, but in professional scenarios, simply outputting the answer itself is insufficient; it is also necessary to explain which materials, entity relationships, and reasoning steps the answer is based on.

[0003] Existing intelligent question answering systems typically employ keyword retrieval, vector retrieval, knowledge graph question answering, or retrieval enhancement generation schemes. Keyword retrieval is controllable but struggles to understand complex natural language; vector retrieval can recall semantically similar text but easily overlooks entity consistency, relational direction, and temporal conditions; knowledge graph question answering enables structured queries and multi-hop reasoning, but users find it difficult to determine the source of the answer; large language models may organize similar facts, outdated content, or unretrieved information into seemingly reasonable answers.

[0004] Therefore, there is an urgent need for a technical approach for intelligent question answering tasks to solve the above problems. Summary of the Invention

[0005] In view of this, embodiments of this application provide a knowledge graph-based intelligent question answering method. This application also relates to a knowledge graph-based intelligent question answering device, a computing device, and a computer-readable storage medium, to solve the problem of inaccurate intelligent question answering in the prior art.

[0006] According to a first aspect of the embodiments of this application, a knowledge graph-based intelligent question answering method is provided, comprising: The system acquires and identifies user questions and contextual information to generate a structured question representation, which includes question intent, entities, relationships, attributes, limiting conditions, time, and answer type. Based on the structured question representation, a knowledge graph path retrieval is performed to obtain candidate graph data, wherein the candidate graph data includes entity nodes, edges, relations, attributes, time, and answer source; Semantic retrieval is performed based on the structured question representation and the candidate graph data to obtain the answer statement slots and candidate original text evidence corresponding to the user question, wherein the answer statement slots represent the slot information of the answer type statement required to answer the user question; Construct the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence to generate a candidate evidence package; The large language model is invoked, and the structured question representation and the candidate evidence package are input into the large language model to output the initial draft of the answer corresponding to the user question; The initial draft answer is split based on the answer declaration slot to obtain at least one answer unit, and the at least one answer unit is verified to obtain the target answer corresponding to the user question.

[0007] According to a second aspect of the embodiments of this application, a knowledge graph-based intelligent question-answering device is provided, comprising: The identification module is configured to acquire and identify user questions and contextual information, and generate a structured question representation, which includes question intent, entities, relationships, attributes, limiting conditions, time, and answer type. The knowledge graph path retrieval module is configured to perform knowledge graph path retrieval based on the structured question representation to obtain candidate knowledge graph data, wherein the candidate knowledge graph data includes entity nodes, edges, relations, attributes, time, and answer source; The semantic retrieval module is configured to perform semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question, wherein the answer statement slot represents the slot information of the answer type statement required to answer the user question; The evidence package construction module is configured to construct the binding relationship between the candidate graph data, the answer declaration slot and the candidate original text evidence, and generate a candidate evidence package; The answer generation module is configured to call a large language model, input the structured question representation and the candidate evidence package into the large language model, and output a draft answer corresponding to the user question; The answer credibility verification module is configured to split the initial answer based on the answer declaration slot to obtain at least one answer unit, and to perform credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question.

[0008] According to a third aspect of the embodiments of this application, a computing device is provided, including a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor executes the computer instructions to implement the steps of the knowledge graph-based intelligent question answering method.

[0009] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions, which, when executed by a processor, implement the steps of the knowledge graph-based intelligent question-answering method.

[0010] The knowledge graph-based intelligent question answering method provided in this application acquires and identifies user questions and contextual information to generate a structured question representation; performs knowledge graph path retrieval based on the structured question representation to obtain candidate knowledge graph data; performs semantic retrieval based on the structured question representation and candidate knowledge graph data to obtain answer statement slots and candidate original text evidence corresponding to the user question; constructs the binding relationship between candidate knowledge graph data, answer statement slots, and candidate original text evidence to generate a candidate evidence package; calls a large language model, inputs the structured question representation and candidate evidence package into the large language model, and outputs a draft answer corresponding to the user question; splits the draft answer based on the answer statement slots to obtain at least one answer unit, and performs credibility verification on at least one answer unit to obtain the target answer corresponding to the user question.

[0011] One embodiment of this application generates a structured question representation by identifying user questions and contextual information. This enables rapid knowledge graph path retrieval to obtain candidate graph data. A recall mechanism then performs semantic retrieval to obtain the answer statement slots and candidate original text evidence corresponding to the user question. Subsequently, the graph path, original text evidence, and answer statement slots are aligned and bound to form a traceable evidence chain. Finally, a large language model is invoked to output the answer to the user question. This achieves controlled generation of the large model within a traceable evidence package; content without bound evidence cannot be output as a definitive answer, unlike selecting the optimal candidate answer based on semantic similarity. Based on this, this application uses intelligent question answering tasks as the clear application boundary, knowledge graph paths as the factual skeleton, original text evidence as the basis, and answer type and constraint rules as the generation boundary. After generation, a declaration-level credibility verification is performed, resulting in stronger accuracy, interpretability, traceability, and controllable output capabilities compared to ordinary RAG solutions. Attached Figure Description

[0012] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the structure of an intelligent question-answering method based on knowledge graphs provided in an embodiment of this application; Figure 2 This is a flowchart of an intelligent question-answering method based on a knowledge graph, provided in one embodiment of this application; Figure 3This is a flowchart illustrating a knowledge graph-based intelligent question-answering method provided in another embodiment of this application. Figure 4 This is a schematic diagram of the structure of an intelligent question-answering device based on a knowledge graph, provided in one embodiment of this application; Figure 5 This is a structural block diagram of a computing device provided in one embodiment of this application. Detailed Implementation

[0014] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0015] The terminology used in one or more embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of one or more embodiments of this application. The singular forms “a,” “the,” and “the” used in one or more embodiments of this application and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of this application refers to and includes any or all possible combinations of one or more associated listed items.

[0016] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this application, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this application, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0017] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0018] First, the terms and concepts involved in one or more embodiments of this application will be explained.

[0019] Intelligent question answering refers to the technical process by which an artificial intelligence system receives a user's natural language question and then combines a knowledge base, knowledge graph, retrieval model, and generation model to output an answer, supporting evidence, and confidence hints.

[0020] Domain knowledge graph: A structured knowledge network built around concepts, entities, attributes, relationships, events, times, source documents, and evidence fragments in a specific question-and-answer domain.

[0021] Chain of evidence: A combination of a graph path, original text fragments, source locations, entity relationships, time conditions, and reasoning steps that support a certain answer or key statement.

[0022] Evidence chain alignment: Establish a traceable mapping relationship between nodes, edges, attributes, and reasoning steps in the graph path and specific evidence locations in the original text, structured data, or multimodal materials.

[0023] Declaration-level confidence: A confidence score is calculated for each key fact statement in the answer to determine whether the statement can be output directly, requires further explanation, or should be transferred to human review.

[0024] It's important to note the overlaps, compliance avoidance, and further technical challenges of existing conventional technical approaches: The main process of existing conventional knowledge graph question answering solutions can generally be summarized as follows: "Constructing an initial knowledge graph from multi-source heterogeneous data—parse the question and execute a single-hop query—expanding the path to multiple hops if the query is not satisfied—calling a large language model for semantic matching and selecting the answer—correcting the knowledge graph's closed-loop feedback." This process belongs to the general architecture of knowledge graph-assisted large model question answering. It typically involves segmenting documents or knowledge fragments and writing them into a vector database. After a user asks a question, several fragments are retrieved based on semantic similarity, and then a large language model generates a natural language answer. Some solutions add keyword retrieval, source links, or simple citation tags. Alternatively, they typically first identify the entities and relationships in the question, then convert the natural language question into a graph query or structured query, obtaining the answer based on the entity relationships in the graph. While advantageous in structured factual question answering, it lacks sufficient support for unstructured data, complex textual evidence, cross-paragraph descriptions, and natural language answer generation. The above-mentioned approach is effective for single-knowledge-source and simple fact-based question answering. However, in intelligent question answering scenarios involving multiple knowledge sources, complex factual relationships, multi-hop reasoning, and the need for interpretable and traceable answers, it remains difficult to reliably guarantee the accuracy, consistency, and verifiability of the answers. Without further restrictions on evidence chain alignment and declaration-level trust verification, it is easily misunderstood as a general retrieval-enhanced question answering process.

[0025] The challenge of this application is not "how to find a high-scoring graph path or a semantically similar answer," but rather "how to prove that every factual statement in the generated answer is supported by traceable evidence units and graph paths." Therefore, this application limits its technical focus to evidence unitization, evidence chain alignment, statement-level credibility verification, and evidence constraint generation.

[0026] This application avoids the conventional design of directly determining the candidate answer with the highest semantic similarity as the optimal answer. Instead, it performs evidence gating before generation, executes evidence package constraints during generation, and performs declaration-level verification after generation. Only content that completes the four-layer binding of "question-path-evidence-declaration" is allowed to be output as a deterministic answer.

[0027] This application further addresses the problem of the decoupling between the graph path and the original evidence: even if a multi-hop path exists in the graph, if its nodes, edges, attributes, and time conditions cannot be aligned with the positions of the original evidence item by item, the path cannot directly enter the generation context, but can only enter the supplementary recall, conflict marking, or manual review process.

[0028] Based on this, this application addresses intelligent question-answering technology in the field of artificial intelligence, particularly applicable to scenarios such as open-domain question answering, vertical-domain question answering, retrieval-enhanced question answering, knowledge-based question answering, knowledge graph question answering, question-answering robots, and multi-source knowledge fusion question answering. These scenarios share the common characteristic that answers not only require natural language expression but also need to explain which knowledge facts, original text evidence, and reasoning paths support the answer; conclusions cannot be drawn solely based on text similarity or the parametric memory of large language models.

[0029] The core technical problem this application aims to solve is: how to align the evidence chain of user questions, graph paths, original evidence, and generated answers during intelligent question answering, so that the AI ​​system is constrained by verifiable evidence when generating each key answer statement, and provides reliable prompts when evidence is insufficient, conflicting, or the reasoning path is incomplete. The following is a summary of the technical problems that can be solved: 1. Address the issues of lack of traceable evidence, ambiguous citation locations, and the potential for factual confusion or "illusion" in intelligent question answering; 2. This addresses the issue that vector retrieval enhancement generation schemes primarily rely on semantic similarity, which may recall content with similar topics but mismatched entities, relationships, times, or answer types; 3. To address the issue that while knowledge graph question answering can express structured relationships, it is difficult to bind them item by item to the original text evidence, resulting in answers lacking natural language evidence. 4. When solving complex problems that require multi-hop entity relationship reasoning, attribute constraint judgment, and cross-source knowledge fusion, existing question-answering systems struggle to form a complete chain of evidence; 5. Resolve the issue of lacking declaration-level validation after answer generation, making it impossible to distinguish between "conclusions supported by evidence," "speculated completions," and "conflicting evidence." 6. Address the issue that user error correction, answer acceptance, manual review, and subsequent interaction results are difficult to use to update graph weights, evidence ranking, and generation strategies.

[0030] To address the aforementioned issues, this application proposes a knowledge graph evidence chain alignment and declaration-level credibility verification technology. This technology processes natural language questions, candidate answers, domain knowledge graphs, original knowledge fragments, structured facts, source metadata, and question-answering feedback. Its core is not to perform single-hop / multi-hop queries on the knowledge graph and then select answers based on semantic similarity, but rather to transform "whether the answer is credible" into "whether a computable evidence chain exists, supported by question intent, entity relationships, graph paths, original evidence, temporal conditions, and generated declarations." This can be summarized as employing a technical path of "knowledge fragment standardization—question-answer ontology modeling—question parsing and decomposition—graph path retrieval—semantic evidence recall—evidence chain alignment—evidence-constrained generation—declaration-level credibility verification—feedback evolution," ensuring that the intelligent question-answering output is always jointly constrained by the knowledge graph, original evidence, and reasoning constraints. The purpose of this application is further clarified as follows: to construct an intelligent question answering method with the core of computable, traceable, constrainable, and verifiable evidence chains, so that the system can exclude non-alignable evidence in the recall stage, restrict the introduction of facts without evidence in the generation stage, and provide credibility and risk warnings for each answer statement in the output stage, thereby distinguishing it from existing solutions that rely solely on knowledge graph path expansion and semantic similarity answer selection.

[0031] This application provides a knowledge graph-based intelligent question answering method. This application also relates to a knowledge graph-based intelligent question answering device, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0032] Figure 1 A schematic diagram of the structure of a knowledge graph-based intelligent question answering method provided according to an embodiment of this application is shown.

[0033] The knowledge graph-based intelligent question-answering method provided in this embodiment can be executed by electronic devices, local servers, private clouds, or hybrid cloud environments, and this application does not specifically limit it.

[0034] It should be noted that the main application scenarios of this application include: users asking natural language questions about professional materials; intelligent customer service systems needing to generate evidence-based answers from multi-source knowledge bases; learning tutoring systems needing to explain the source of answers and reasoning steps; research material question answering needing to locate evidence from papers, tables, and terminological relationships; and encyclopedic or technical knowledge question answering needing to handle entity ambiguity, relational direction, and temporal conditions. All of the above tasks can be abstracted into a joint judgment of question intent, entities, relations, attributes, time, and answer type.

[0035] See Figure 1 , Figure 1 The diagram illustrates the functional units of this application, including a knowledge source processing unit, a question-answering ontology and graph construction unit, a question parsing and decomposition unit, an evidence retrieval and evidence chain alignment unit, a controlled generation unit, and a credible verification and feedback unit.

[0036] Specifically, the Knowledge Source Processing Unit: This unit is used to access textual materials, structured data, web page fragments, question-and-answer logs, tables, image descriptions, and multimodal parsing results, performing layout parsing, chapter recognition, fragment segmentation, field mapping, deduplication, source annotation, time stamping, and quality assessment. Each knowledge fragment retains its source, location, time, entity candidates, relation candidates, and quality score. The Question-Answer Ontology and Graph Construction Unit: This unit is used to define entity types, relation types, attribute fields, question intent, answer types, and evidence templates in the intelligent question-answering scenario. For example, entity types include people, organizations, terms, papers, products, locations, events, and fault phenomena; relation types include raising, containing, belonging to, causing, using, occurring at, comparing, citing, and originating from. The Question Parsing and Decomposition Unit: This unit is used to parse user questions and multi-turn contexts, identify question intent, core entities, candidate relations, limiting conditions, time conditions, and answer types, and generate a set of sub-tasks with dependencies in complex question-answering scenarios. Evidence Recall and Evidence Chain Alignment Unit: This unit first searches for candidate paths in the knowledge graph based on the structured question representation, then recalls original text fragments from the semantic evidence index, and aligns the graph paths with the original text evidence according to entities, relations, attributes, time, source, and answer statements. Evidence that fails to align or has significant conflicts is not directly used for generation. Controlled Generation Unit: This unit inputs the structured question, evidence package, graph path, constraint rules, and output template into the AI ​​model, requiring the model to output answers, justifications, reasoning paths, applicable conditions, and risk warnings only based on aligned evidence. For content with insufficient evidence, the output is "Insufficient to judge" or "Supplementary information required." Credibility Verification and Feedback Unit: This unit performs factual consistency checks, evidence coverage checks, graph path completeness checks, time validity checks, and conflict checks on the generated results. User error correction, answer adoption, evidence clicks, and manual review conclusions are mapped back to entities, relations, evidence fragments, and prompt strategies to update the graph quality score and evidence ranking weights.

[0037] In summary, the purpose of this application is to provide an intelligent question-answering technology based on knowledge graph evidence chain alignment. This technology abstracts multi-source knowledge materials into knowledge fragments, entity nodes, relation edges, attribute conditions, source evidence, and time information, enabling the intelligent question-answering system to obtain verifiable graph evidence paths before generating answers, and then undergo declaration-level credibility verification after answer generation. Furthermore, this application aims to achieve the following objectives: first, improve the accuracy of evidence recall by locating key knowledge based on entities, relations, attributes, and answer types; second, improve the credibility of generated results by binding each key answer declaration to an evidence number and graph path; third, improve the ability to handle complex problems by supporting multi-hop reasoning, cross-source knowledge fusion, and evidence conflict identification; fourth, improve output controllability by prompting "insufficient to judge" or initiating manual review when evidence is insufficient or conflicts cannot be resolved; and fifth, improve continuous maintenance capabilities by mapping user feedback and review results back to graph weights, evidence quality scores, and prompting strategies.

[0038] Figure 2 The flowchart illustrates an intelligent question-answering method based on a knowledge graph according to an embodiment of this application, which specifically includes the following steps: It should be noted that this embodiment relates to an intelligent question-answering technology based on knowledge graph evidence chain alignment. This technology abstracts multi-source knowledge data into knowledge fragments, entity nodes, relation edges, attribute conditions, source evidence, and time information. After receiving a user's natural language question, it first performs question intent recognition and structured question parsing. Then, it obtains candidate evidence through graph path retrieval, semantic evidence recall, and candidate answer type constraints. Subsequently, it aligns the graph path with the original text evidence into an evidence chain, and under the constraint of the evidence package, it calls an AI model to generate the question-answering answer. Finally, it reduces the risk of illusion and improves the interpretability of the answer through declaration-level consistency verification, confidence scoring, conflict identification, and feedback update mechanisms. Compared with solutions that rely solely on vector libraries or a single large language model, this invention can form a closed loop of "question parsing—graph constraint—evidence recall—evidence chain alignment—controlled generation—credible verification—feedback evolution" in intelligent question-answering scenarios where multi-source knowledge coexists, questions require multi-hop reasoning, and answers require traceable evidence, thereby improving the accuracy, stability, and verifiability of the generated answer.

[0039] Step 202: Obtain and identify user questions and contextual information, and generate a structured question representation, which includes question intent, entities, relationships, attributes, limiting conditions, time, and answer type.

[0040] Specifically, after obtaining the user's question and related contextual information, the system first identifies the question intent, entities, relationships, attributes, constraints, time, and answer type in the input. It should be noted that for more accurate identification, multiple rounds of context can be obtained before converting the information into a structured question representation, as shown in Formula 1 below: T_Q =<Intent_Q, E_Q, Rel_Q, Obj_Q, Cond_Q, Time_Q, AnsType_Q, Ctx_Q> Formula 1 Wherein, T_Q represents the structured question representation of the Q question; Intent_Q represents the question intent, which can be categorized as definition, cause, step, comparison, list, numerical, or source; E_Q represents the entity involved in the question; Rel_Q represents the relationship involved in the question; Obj_Q represents the question object (which can also be understood as the question attribute); Cond_Q represents the limiting condition; Time_Q represents the time condition; AnsType_Q represents the answer type; and Ctx_Q represents the multi-turn dialogue context.

[0041] For example, if a user asks "What are the main improvements of the method proposed in a certain paper compared to the baseline model?", this embodiment needs to identify the paper, method name, and comparison object to facilitate subsequent tracing along the knowledge graph path such as paper—method—module—experimental metric—dataset, and to retrieve the paper text and tabular evidence.

[0042] Furthermore, for complex problems, the method further constructs a set of subtasks. Continuing with the previous example, the question "What are the advantages of the method proposed in a certain paper compared to baseline methods, and from which dataset did the experimental data come?" can be broken down into subtasks such as paper entity recognition, method entity recognition, baseline method recognition, comparison dimension recognition, experimental indicator retrieval, dataset relationship retrieval, and final answer integration. See Formula 2 below for details: Decomp(Q) = {tau_1, tau_2, ..., tau_n}, Dep = {(tau_i, tau_j) | tau_imust be solved before tau_j} Formula 2 Where Decomp(Q) represents the set of subtasks obtained by decomposing problem Q; \tau_1, \tau_2, ..., \tau_n represent the individual subtasks obtained by decomposing problem Q, and n represents the number of subtasks; Dep represents the set of dependencies between subtasks; \tau_i, \tau_j indicate that there is an execution dependency between subtasks \tau_i and subtasks \tau_j, that is, the processing result of subtasks \tau_i is a prerequisite for executing subtasks \tau_j, and subtasks \tau_i must be completed before subtasks \tau_j can be executed.

[0043] It should be noted that if the task analysis results lack key conditions, such as the missing question object, time range, comparison dimensions, or answer type, the method can decide whether to ask follow-up questions based on the risk level. In low-risk questions and answers, answers with hypothetical conditions can be output; in high-risk questions and answers or those with significant evidence conflicts, priority should be given to prompting supplementary conditions or entering manual review.

[0044] In summary, this method is beneficial for binding each improvement point to the original paragraph or experimental table number when outputting the answer; if different versions of the paper give different indicators, a conflict prompt will be output, thus enabling the complete and accurate output of the answer to the user's question.

[0045] Step 204: Perform knowledge graph path retrieval based on the structured question representation to obtain candidate graph data, wherein the candidate graph data includes entity nodes, edges, relations, attributes, time, and answer source.

[0046] Knowledge graphs are used to represent concepts, entities, attributes, relationships, events, source documents, evidence fragments, and temporal conditions in the question-answering domain. Graph path retrieval represents the process of finding relevant evidence paths within a graph structure, starting from a question entity, relationship, or answer type, based on relationships and constraints.

[0047] Furthermore, to ensure the controllability of subsequent graph construction and AI generation, this application first performs unified modeling on textual materials, structured data, question-and-answer logs, web page materials, tables, image descriptions, and multimodal parsing results. Specifically, before performing knowledge graph path retrieval based on the structured question representation to obtain candidate graph data, the process also includes: First, acquire multi-source data, standardize the multi-source data, and abstract it into at least one knowledge fragment with attribute information, including source identifier, location identifier, time information, and quality score; A knowledge graph is constructed based on at least one knowledge fragment with attribute information.

[0048] The multi-source data can be understood as multimodal data such as textual materials, structured data, question-and-answer logs, web page materials, tables, and image descriptions. A domain knowledge graph is represented as a structured knowledge network built around concepts, entities, attributes, relationships, events, time, and evidence fragments within a specific question-and-answer domain.

[0049] Specifically, after acquiring multi-source data, the data is standardized, abstracting multi-source knowledge into knowledge fragments, entity nodes, relation edges, attribute conditions, source evidence, and time information. In practical applications, it mainly targets the processing objects of intelligent question answering, namely text data, structured data, question-answering logs, and multimodal data. The core technology processing includes, but is not limited to, operations such as multi-source data segmentation, deduplication, source labeling, entity candidate recognition, and time and quality marking. The main output content includes knowledge fragments and metadata.

[0050] As an example: Let the i-th knowledge fragment be k_i, which consists of the main text content, source identifier, location identifier, entity set, relation set, time information, quality score, and vector representation, as shown in Formula 3 below: k_i =<x_i, src_i, pos_i, ent_i, rel_i, time_i, q_i, emb_i> Formula 3 Where x_i is the main text content or structured record; src_i is the source document or data source; pos_i is the page number, paragraph, line number, table cell, or webpage anchor; ent_i is the set of entities identified in the fragment; rel_i is the set of candidate relations; time_i is the time, version, or event occurrence condition; q_i is the data quality score; and emb_i is the fragment vector representation.

[0051] For multiple sources of the same knowledge fact, this application does not simply retain the most similar text, but rather retains the source's authority, time validity, content completeness, and historical feedback information. If the user's question explicitly includes a time condition, the method can select valid evidence at that time according to that time condition; if no time is specified, the latest evidence that has been verified for multi-source consistency will be given priority. The timeliness score for a certain knowledge fragment can be found in Formula 4 below: Fresh(k_i, t_q) = I(time_i matches t_q) exp(-delta max(0, t_now- t_update_i)) Formula 4 Where Fresh(k_i,t_q) represents the timeliness score of knowledge fragment k_i relative to query time t_q; I( ) is the indicator function; delta is the time decay coefficient; t_now represents the current time when performing knowledge retrieval or timeliness calculation; t_update_i represents the most recent update time of knowledge fragment k_i.

[0052] Furthermore, after standardizing the multi-source data, a domain knowledge graph can be constructed. Specifically, the domain knowledge graph can be represented by the following formula 5: G = (V, E, A, S, T, R, B) Formula 5 Wherein, G represents the domain knowledge graph; V represents the set of nodes such as concepts, entities, attribute values, events, question intents, answer types, and evidence fragments; E represents the set of edges relating nodes; A represents the set of attributes of nodes or edges; S represents the set of sources such as source documents, data tables, web pages, question-and-answer records, or multimodal data; T represents time, version, or temporal conditions; R represents inference constraints, relational constraints, and answer format constraints; and B represents the binding relationship between graph nodes, edges, and the original evidence locations.

[0053] Through the aforementioned knowledge graph structure, this application can transform the question "Why is a certain intelligent question-answering answer valid?" into a computational question of "Which question entities, candidate relations, graph paths, original text fragments, time conditions, and reasoning constraints jointly support the answer?" Compared with traditional retrieval schemes, this application not only determines whether the texts are similar, but also whether the entities are consistent, whether the relations match, whether the attribute conditions are met, whether the answer type is correct, and whether there is more reliable conflict evidence.

[0054] Step 206: Perform semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question, wherein the answer statement slot represents the slot information of the answer type statement required to answer the user question.

[0055] In existing technologies, the output answer is determined by scoring the semantic similarity of candidate answers. However, this cannot guarantee that every key factual statement in the answer is directly implied by the evidence. It is still possible to output semantically related but factually inconsistent content as a definitive answer. To address this, semantic retrieval is performed based on the structured question representation and candidate graph data corresponding to the user question to obtain the information of each answer statement slot corresponding to the user question, as well as the candidate original text evidence corresponding to each answer statement slot.

[0056] Further, the step of performing semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question includes: Based on the structured problem representation and the candidate graph data, a similarity calculation is performed, and the graph data with a similarity that meets a preset threshold is used as the recalled knowledge fragment; Determine the answer statement slot and candidate original text evidence corresponding to the user question in the recalled knowledge fragment.

[0057] Specifically, in the entity fusion process, similarity calculation is mainly used to recall knowledge fragments related to structured questions from the knowledge graph data. Within these recalled knowledge fragments, the corresponding answer type statement for the user's question and the supporting evidence under that answer statement are identified; this can be understood as recalling candidate original text evidence from the graph data. During entity fusion, the method needs to determine whether two candidate entities represent the same object. To avoid incorrectly merging objects with the same name but different identities, the method comprehensively considers name similarity, context vector similarity, attribute overlap, and graph neighbor similarity. Entity merging is only performed when the comprehensive similarity exceeds the merging threshold and there are no type conflicts, time conflicts, or key attribute conflicts. The fusion similarity calculation can refer to the following formula 6: Sim_ent(a,b)=eta1 Sim_name(a,b)+eta2 Sim_emb(a,b)+eta3 Sim_attr(a,b)+eta4 Sim_nei(a,b) Formula 6 Where Sim_ent(a,b) is the fusion similarity between entity a and entity b; Sim_name is the name similarity; Sim_emb is the context vector similarity; Sim_attr is the attribute similarity; Sim_nei is the graph neighbor structure similarity; eta1 to eta4 are weight coefficients.

[0058] In practical applications, knowledge graphs do not directly write all extracted triples into the graph database, but are first constrained by the question-answering ontology. The question-answering ontology defines the legal entity types, relation types, attribute fields, question templates, answer types, and evidence templates in intelligent question-answering scenarios. For example, a person can have their birthplace, works, and tenure; a paper can propose methods, use datasets, and report metrics; equipment can have fault phenomena, causes, and handling steps; concepts can have hierarchical, subordinate, compositional, causal, or comparative relationships. The question-answering ontology can be referenced in the following formula 7: O = {C, Rel, Attr, Cons, Temp, AnsType} Formula 7 Where O represents the question-answering ontology; C is the set of entity types; Rel is the set of relation types; Attr is the set of attribute types; Cons is the set of constraint rules; Temp is the set of question parsing templates, relation extraction templates, evidence alignment templates, and output templates; and AnsType is the set of answer types.

[0059] In the example of intelligent question answering, entity nodes can be people, organizations, terms, papers, products, locations, events, or fault phenomena. In the knowledge graph, these are represented as nodes in V of Formula 5, with aliases, types, attributes, and source tags. Therefore, entity nodes can serve as the basic objects for question identification and path retrieval. Relational edges can be represented as proposing, belonging to, containing, causing, applicable to, comparing, using, or occurring in. In the knowledge graph, these are represented as directed or undirected edges in E of Formula 5, with confidence and evidence binding. Therefore, relational edges can express the factual relationships required for answer reasoning. Attribute nodes... Attribute nodes can be time, numerical value, indicator, version, unit, condition, or range. In a knowledge graph, they can be expressed as attribute values ​​bound to entities or relationships, used to define answer boundaries and eliminate ambiguity. Evidence nodes can be original text paragraphs, table cells, web page fragments, image descriptions, or structured records. In a knowledge graph, they can be expressed as evidence sources bound to nodes / edges, used to support graph relationships and the final answer. Answer type nodes can be definitions, reasons, steps, comparisons, lists, numerical values, sources, or conclusions. In a knowledge graph, they can be expressed as associations with the question intent and output template, mainly used to constrain the retrieval direction and generation format.

[0060] It should be noted that the mixed evidence recall is a key aspect that distinguishes this application from ordinary RAG schemes. Ordinary RAGs often rely primarily on text similarity, easily recalling segments that are semantically similar but inconsistent in entity, relation, time, or answer type. This application first uses structured question representations to determine candidate entities and candidate relation paths in the knowledge graph, then combines dense vectors, sparse keywords, and structured indexes to supplement the original text evidence, and finally obtains the candidate evidence set through a unified scoring function. The candidate graph path set can be referenced in the following formula 8: P_Q={p|p=<v_0,e_1,v_1,...,e_l,v_l> ,l<=H,type(e_j)inRel_allow,MatchType(p,AnsType_Q)=1} Formula 8 Where P_Q is the set of candidate graph paths recalled for question Q; H is the maximum number of hops; Rel_allow is the set of relation types allowed to participate in the current question-answering task; MatchType(p,AnsType_Q)=1 indicates that path p can support the target answer type. The graph path score is based on the following formula 9: S_path(p)=alpha1 Match(p,T_Q)+alpha2 Auth(p)+alpha3 Fresh(p)+alpha4 Completeness(p)-alpha5 Len(p)-alpha6 Conflict(p) Formula 9 Where S_path(p) is the graph path score of candidate path p; Match is the matching degree with the structured question; Auth is the source reliability; Fresh is the timeliness; Completeness is the path integrity; Len is the path length penalty; Conflict is the conflict penalty; alpha1 to alpha6 are weight coefficients.

[0061] Simultaneously, the method recalls text fragments semantically relevant to the question or subtask from the text evidence index and cross-validates the text recall results with the graph candidate paths. If the key entities, relational directions, time conditions, or answer types in the text fragment are inconsistent with the graph path, the ranking score of that fragment is reduced.

[0062] vec(q,d_i)=cos(Emb(q),Emb(d_i))

[0063] S_recall(p,d_i)=mu1 S_path(p)+mu2 S_vec(q,d_i)+mu3 Align(p,d_i)+mu4 Fresh(d_i)-mu5 Risk(d_i) Formula 10

[0064] Where S_recall(p,d_i) is the comprehensive recall score of the candidate evidence pair consisting of the graph path p and the text fragment d_i; Align represents the alignment degree between the path and the text evidence; Risk represents the risk such as expiration, conflict, low quality, or unreliable source; mu1, mu2, mu3, mu4, and mu5 are the weight coefficients corresponding to each scoring item, which are used to adjust the influence of the graph path matching score, vector semantic similarity, alignment degree between the path and the text, timeliness score, and risk score on the comprehensive recall score, respectively.

[0065] Step 208: Construct the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence to generate a candidate evidence package.

[0066] In practical applications, existing multi-hop dynamic search technologies mainly focus on path expansion efficiency and path scoring. They fail to form an evidence package composed of graph paths, text fragments, structured fields, metadata, excluded evidence, and conflicting evidence. As a result, the evidence package contains unusable evidence content. In response, this application splits the answer matched to the user question into statements and completes the binding between statements, graphs, and evidence to generate answer content with original evidence statements.

[0067] While existing knowledge graph question-answering schemes can form reasoning paths and associated evidence, they typically do not require each entity, relation, attribute, and time condition in the graph path to establish a traceable binding with the location of the original text evidence. This can still lead to the problem of "the graph is searchable, but the original text cannot prove it." To address this issue, this application provides an original text evidence chain alignment method. The method involves constructing the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence to generate a candidate evidence package, including: Based on the slot declared in the answer, extract the map data to be aligned from the candidate map data; Based on the answer statement slot, extract the original text evidence to be aligned from the candidate original text evidence; Calculate the alignment score between the target graph data and the target original text evidence. The alignment score represents the consistency relationship between the attribute content of the knowledge fragment corresponding to each answer statement slot and the attribute content of the extracted original text evidence. Based on the alignment score, the candidate graph data, the answer statement slot, and the candidate original text evidence, a candidate evidence package is constructed.

[0068] Specifically, based on the answer declaration slot corresponding to the user's question, the unit content to be aligned is extracted from the candidate graph data and candidate original text evidence, such as comparing the same entity, the same relationship, or the same attribute. The alignment score can be calculated to determine the matching degree between the content, thereby excluding mismatched text fragments, which will not be input into the big oracle model as evidence. From the perspective of model input, the accuracy of the answer has already been improved.

[0069] As an example, the constructed candidate evidence package fields may include the `task_id` field, meaning the task number for this question-and-answer session, such as Q-AIQA-001; the `question_parse` field, meaning the question intent, entity, relation, condition, and answer type, such as a definition-type question; an entity being an algorithm; and an answer type being "principle explanation"; the `graph_paths` field, meaning the set of graph paths supporting the answer, such as algorithm—proposed paper—key modules—experimental indicators; the `text_evidence` field, meaning the original text fragments or structured records aligned with the graph paths, such as paragraph 3 of the paper or the indicator row in Table 2; the `claim_bindings` field, meaning the binding relationship between the answer claim and the evidence number, such as claim S1 binding E1 and E3; the `excluded_evidence` field, meaning the evidence that was excluded or downgraded and the reasons, such as entities with the same name but different types; or the source being too old; and the `risk_flags` field, meaning risk flags such as conflict, expiration, lack of evidence, and low reliability, such as numerical indicators having two inconsistent sources.

[0070] In practical applications, evidence chain alignment (representing the mapping between nodes, edges, and reasoning steps in the graph and specific evidence locations in the original material) is used to ensure that each piece of evidence input to the AI ​​model corresponds simultaneously to both the structured graph path and the location in the original material. The method extracts entity, relation, attribute, answer type, and time information from the graph path, and then extracts entity mentions, relational expressions, field values, source locations, and contextual sentences from text fragments or structured records. Finally, it calculates the alignment score between the two, as shown in Formula 11 below: Align(p,d_i)=rho1 Jaccard(Ent(p),Ent(d_i))+rho2 RelMatch(p,d_i)+rho3 AttrMatch(p,d_i)+rho4 TimeMatch(p,d_i)+rho5 SrcMatch(p,d_i) Formula 11 Wherein, Ent represents the entity set; RelMatch represents the relational expression matching degree; AttrMatch represents the attribute or numerical matching degree; TimeMatch represents the time condition matching degree; SrcMatch represents source consistency; and rho1 to rho5 are alignment weights.

[0071] It should be noted that, in the evidence chain alignment stage, the present application not only calculates the similarity between paths and text fragments, but also establishes a three-level alignment matrix of "graph path p - evidence unit u - answer statement c". Before generation, candidate statement slots are formed first, such as definition, cause, value, source, step, comparison conclusion, etc.; each statement slot must be bound to at least one set of evidence units that meets a threshold, and in this regard, the candidate evidence package can refer to the following Formula 12: M(p,u,c)=w1 EntAlign(p,u,c)+w2 RelAlign(p,u,c)+w3 AttrAlign(p,u,c)+w4 TimeAlign(p,u,c)+w5 NLI(u,c)-w6 Conflict(u,c) Formula 12 Wherein, NLI(u,c) indicates whether the evidence unit u entails the statement c; Conflict(u,c) indicates the contradiction risk between the evidence and the statement. w1 is the entity alignment weight coefficient, which is used to control the influence degree of entity consistency on the comprehensive alignment score; w2 is the relation alignment weight coefficient, which is used to control the influence degree of consistency of relation type, relation semantics and relation direction on the comprehensive alignment score; w3 is the attribute alignment weight coefficient, which is used to control the influence degree of consistency of attributes and attribute values on the comprehensive alignment score; w4 is the time alignment weight coefficient, which is used to control the influence degree of consistency among query time, evidence valid time and time conditions involved in the statement on the comprehensive alignment score; w5 is the semantic entailment weight coefficient, which is used to control the influence degree of the semantic support degree of the evidence unit for the candidate statement on the comprehensive alignment score; w6 is the conflict risk weight coefficient, which is used to control the penalty degree on the comprehensive alignment score when there is factual contradiction, semantic conflict or condition inconsistency between the evidence unit and the candidate statement. If max M(p,u,c)<tau_align, the statement c is marked as an unsupported statement and is not allowed to be output as a deterministic conclusion.

[0072] In addition, the present application also provides a solution in terms of conflict recognition, and the method compares the conclusions of the same entity, the same relation or the same attribute in different evidences. If one source indicates that an entity has a certain attribute, and another source gives an opposite attribute value, it is further judged whether it is a real conflict according to source reliability, update time, context range and evidence coverage. If the conflict cannot be automatically resolved, the conflict description is retained in the output result and the confidence level is reduced.

[0073] Further, the step of constructing a candidate evidence package based on the alignment score binding the candidate graph data, the answer statement slot, and the candidate original text evidence includes: If the alignment score is determined to be lower than a preset alignment score threshold, the candidate graph data corresponding to the alignment score and the candidate original text evidence are determined as excluded answer content. Record the answer declaration slot, the content of the excluded answer, and the reason for excluding the content of the excluded answer.

[0074] Specifically, text fragments whose alignment scores calculated in the above manner are below the threshold will not directly enter the generation context, but will enter the supplementary retrieval or manual review process. Candidate graph data and candidate original text evidence with low scores can be recorded as excluded answer content, along with the corresponding exclusion reasons.

[0075] In practical applications, excluded evidence is not discarded directly, but the reasons for exclusion are recorded, such as entities with the same name but different meanings, inconsistent relationship directions, expired time, low source priority, conflict with high-credibility sources, or inability to support the target answer type.

[0076] By recording Excl_Q, this application can explain to human reviewers or subsequent maintenance why certain semantically similar content was not used to generate answers, avoiding possible erroneous splicing when filtering based solely on semantic matching scores.

[0077] Step 210: Invoke the large language model, input the structured question representation and the candidate evidence package into the large language model, and output the initial draft of the answer corresponding to the user question.

[0078] Specifically, the step of invoking the large language model, inputting the structured question representation and the candidate evidence package into the large language model, and outputting a draft answer corresponding to the user question includes: The large language model is invoked, and the structured question representation, candidate evidence package, constraints, and output answer template are input into the large language model. The structured question representation includes the identified answer entities, the candidate evidence package includes candidate answer types, graph paths, and original text evidence that completes the alignment score calculation, the constraints include exclusion constraints for original text evidence that does not meet the alignment score conditions, and the output answer template includes risk level categories. Obtain a draft answer output by the large language model, wherein each answer unit in the draft answer is bound to at least one evidence number and risk level.

[0079] Specifically, the structured question representation corresponding to the user's question, candidate evidence package, constraints, and output answer template are all input into the Big Prophet model. This allows the model to output answers corresponding to the decomposed structured questions, with the answer content having bound original evidence text in the candidate evidence package, ensuring traceability. Simultaneously, the model outputs answer content to the user according to the answer template, satisfying the corresponding constraints. To present accurate answer information to the user, the initial draft of the output answer also displays the binding relationship to the original evidence text; at least one evidence number and its corresponding risk level are displayed in the answer.

[0080] In practical applications, during the answer generation stage, the input to the AI ​​model is not an open-ended question, but a controlled context containing structured questions, evidence packages, graph paths, constraint rules, and output templates. The method organizes the input into seven parts: "Question Analysis—Identified Entities—Candidate Answer Types—Graph Paths—Aligned Evidence—Prohibited Items—Output Format." It requires the model to output "Insufficient to Judge" or "Requires Supplementary Information" for information not present in the evidence package, rather than automatically completing it. For example, the final candidates can be represented as shown in Formula 13: A = argmax_A[logP_LLM(A|Q,B,R)+kappa Support(A,B) - lambda Risk(A,R) ] Formula 13 Among them, A P_LLM represents the final candidate answer; P_LLM represents the probability that the AI ​​model generates answer A under the conditions of question Q, evidence package B, and constraint R; Support(A,B) represents the degree to which the answer is supported by the evidence package; Risk(A,R) represents the risk that the answer violates the constraints or lacks evidence; kappa and lambda are adjustment coefficients.

[0081] In the preferred embodiment, the output must include a direct answer, explanation of the basis, reasoning path, applicable conditions, risk warning, and traceable evidence number; each key conclusion must be associated with at least one evidence number. If a conclusion cannot be associated with an evidence number, it is marked as speculative content and deleted, downgraded, or transferred to manual review based on the risk level.

[0082] It should be noted that the description of the large language model in this application can be understood as a description of the use of other AI models. The concept includes, but is not limited to, large language models, text vector models, sparse retrieval models, entity relation extraction models, ranking models, implication judgment models, classification models, and graph reasoning models.

[0083] Step 212: Based on the answer declaration slot, split the initial draft of the answer to obtain at least one answer unit, and perform a credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question.

[0084] Specifically, after the answer is generated, this application further splits the answer into a set of statements C={c1,c2,...,cn}, and performs checks on each statement for evidence coverage, graph path completeness, implication consistency, time validity, conflict status, and source reliability. The answer-level confidence score does not use the overall semantic similarity of candidate answers, but rather a weighted aggregation of statement-level confidence scores or a minimum value constraint. The output can include four parts: the first part is the direct answer for the user; the second part is a list of evidence supporting the answer; the third part is the reasoning path, such as "problem entity—candidate relationship—source fragment—answer statement"; and the fourth part is a risk warning, such as "There is ambiguity in entities with the same name; an entity has been selected based on the context" or "The logarithmic indicators of the two sources are inconsistent; manual review is recommended."

[0085] In high-risk or highly specialized question-and-answer scenarios, if the confidence level is below a preset threshold, the method does not directly output a definitive answer. Instead, it outputs candidate answers, points of difference, and requests for supplementary information to reduce the risk of AI errors being directly adopted.

[0086] It should be noted that existing technical solutions typically lack comprehensive verification of declaration-level confidence, evidence coverage, graph path completeness, and conflict penalty. Low-confidence content may be mistakenly identified as a definitive answer. Therefore, this application requires credibility verification of each answer unit in the initial draft of the answer, and calculates path completeness, evidence coverage, implied consistency, time validity, conflict penalty, and source reliability for each declaration.

[0087] Further, the step of performing credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question includes: Calculate the confidence level corresponding to at least one answer unit; If the confidence level is greater than a first preset threshold, output at least one answer unit corresponding to the user question and the confidence level; If the confidence level is less than or equal to the first preset threshold and greater than the second preset threshold, at least one answer unit and the risk level corresponding to the answer unit shall be output. If the confidence level is less than or equal to the second preset threshold, output supplementary conditions or manual review prompts corresponding to the user's question.

[0088] In practical applications, the trusted verification layer scores each key statement in the generated results. Key statements can be such as "a method was proposed by a certain paper," "a certain indicator reached a certain value on a certain dataset," "a certain fault is usually caused by a certain reason," or "a certain concept belongs to a certain higher-level concept," etc. The method calculates the statement-level confidence and the answer-level confidence separately. In high-risk question answering, the lowest statement confidence can be used as the output constraint; the comprehensive confidence can be referenced in the following formula 14: C(A) = sigmoid(theta1) C_path+theta2 C_evi+theta3 C_rule+theta4 C_model+theta5 C_align-theta6 C_conf-theta7 C_time) Formula 14 Where C(A) is the overall confidence score of answer A; C_path is the graph path completeness; C_evi is the evidence coverage; C_rule is the constraint consistency; C_model is the model self-consistency; C_align is the evidence chain alignment; C_conf is the conflict penalty; C_time is the time penalty; theta1 is the graph path completeness weight coefficient, used to control the influence of the graph path completeness on the overall confidence score of the answer; theta2 is the evidence coverage weight coefficient, used to control the influence of the degree of evidence coverage of the answer statement on the overall confidence score; theta3 is the constraint consistency weight coefficient, used to control the influence of the degree of consistency between the answer and the question constraints on the overall confidence score; theta4 is the model self-consistency weight coefficient, used to control the influence of the degree of consistency of different inferences or generated results of the model itself on the overall confidence score; theta5 is the evidence chain alignment weight coefficient, used to control the influence of the degree of alignment between the graph path, evidence unit and answer statement on the overall confidence score; theta6 is the model self-consistency weight coefficient, used to control the influence of the degree of alignment between the graph path, evidence unit and answer statement on the overall confidence score; theta6 is the model self-consistency weight coefficient, used to control the influence of the model self-consistency weight coefficient; theta5 is the evidence chain alignment weight coefficient, used to control the influence of the degree of alignment between the graph path, evidence unit and answer statement on the overall confidence score; theta6 is the model self-consistency weight coefficient, used to control the influence of the model self-consistency weight coefficient; theta6 is the model self-consistency weight coefficient, used to control the influence of the model self-consistency weight coefficient; theta7 is the model self-consistency weight coefficient, used to control the influence of the model self-consistency weight coefficient; theta8 is the model self-consistency weight coefficient, used to control the influence of the degree of alignment between the graph path, evidence unit and answer statement on the overall confidence score; theta6 is the model self-consistency weight coefficient, used to control the influence of the The conflict penalty weight coefficient is used to control the degree of negative impact of evidence conflict or contradiction between evidence and answer on the overall confidence level; theta7 is the timeliness penalty weight coefficient, used to control the degree of negative impact of expired evidence or mismatch of time conditions on the overall confidence level.

[0089] It should be noted that the completeness of the graph path indicates whether the key entities and relationships in the answer can form a connected path; the evidence coverage indicates whether the key statements are supported by original text fragments or structured records; the conflict penalty indicates whether different sources give opposite or incompatible conclusions; and the timeliness penalty indicates whether the evidence is outdated or replaced by updated content.

[0090] When the confidence level is higher than the high threshold, the answer and evidence chain are output directly; when the confidence level is in the middle range, candidate answers, risk warnings and supplementary explanations are output; when the confidence level is lower than the low threshold, no definitive answer is output, but supplementary conditions are prompted or manual review is initiated.

[0091] The key point of this application is that the method does not only output a natural language answer, but also simultaneously outputs a chain of evidence and a confidence level. The chain of evidence may include graph paths, original text fragments, source locations, entity relationships, attribute conditions, time information, relevant constraints, and excluded evidence. The confidence level is used to indicate to the user whether the answer can be directly adopted, or whether supplementary conditions or manual review are required.

[0092] Furthermore, the method also includes: Record user feedback events, wherein the feedback events include explicit feedback events and implicit feedback events. The explicit feedback events include user likes, copy operations, evidence clicks, correction operations, or manual review feedback events. The implicit feedback events include whether the user continues to ask follow-up questions and whether the same question appears repeatedly. The feedback event is mapped to the knowledge graph, and the graph data in the knowledge graph is updated; wherein, the update operation includes incremental updates to the entity relationships, evidence source weights, evidence ranking, and generation strategies of the graph data.

[0093] In practical applications, to ensure the method continuously adapts to knowledge changes, this application establishes a feedback-driven knowledge evolution mechanism. This feedback-driven graph evolution is not a static knowledge base, but rather updates weights based on user corrections and review results, allowing the technical solution to continuously optimize with knowledge updates and user feedback. User acceptance of answers, corrections, follow-up questions, evidence clicks, manual review results, and subsequent question-and-answer performance are all recorded as feedback events and mapped to specific entities, relationships, evidence fragments, answer templates, and output strategies. Feedback can be divided into explicit and implicit feedback. Explicit feedback includes likes, corrections, selection of correct evidence, and manual review opinions; implicit feedback includes whether the user continues to ask follow-up questions, clicks on evidence, copies the answer, and whether the same question recurs. The method maps feedback to entities, relationships, evidence fragments, and prompt strategies, updating the graph quality score, source weights, and model invocation strategies accordingly. As an example, the feedback mechanism can be implemented by calculating evidence weights, as shown in Formula 15 below: w_e^{t+1}=clip(w_e^t+beta Feedback(e)-gamma Age(e)-zeta Error(e),0,1) Formula 15 Where w_e^t represents the weight of the relation edge or evidence e in round t; Feedback(e) represents positive feedback; Age(e) represents time decay; Error(e) represents the number of errors or manual rejections; beta, gamma, and zeta are update coefficients.

[0094] When a certain type of question repeatedly experiences low confidence or user corrections, the method can generate a knowledge gap report, prompting knowledge maintainers to supplement information, improve ontology relationships, or revise extraction templates. When a piece of evidence is consistently adopted without correction, its weight can be increased, making it easier to recall in similar intelligent question-answering tasks.

[0095] The creative enhancements that this application needs to emphasize include the following: 1. Clear technical objective: This invention does not simply improve the question-answering hit rate through dynamic path search and semantic matching, but rather ensures the verifiability of the answer through evidence chain alignment and declaration-level verification.

[0096] 2. Different data structures: This invention introduces data structures such as evidence units, evidence packages, declaration slots, excluded evidence, and audit logs, so that the original evidence, graph path, and generated answer form a closed loop.

[0097] 3. Different recall mechanisms: This invention does not simply expand multi-hop paths, but rather reduces the recall of semantically similar but factually incorrect cases by integrating a hybrid recall and gating system using graphs, vectors, keywords, structured fields, metadata, and feedback signals.

[0098] 4. Different generation control: This invention limits the large model to controlled generation within the evidence package. Content not bound to evidence cannot be output as a definitive answer, which is different from selecting the optimal candidate answer based on semantic similarity.

[0099] 5. Different verification granularity: This invention calculates the credibility of each key statement in the generated answer, and the output results can be subdivided into direct output, cautious output, deletion and rewriting, and manual review, rather than just giving the overall optimal answer.

[0100] In summary, the technical solution of this application addresses the problem of structured factual relationships in intelligent question answering through question-answering ontology and knowledge graph representation; it addresses the problem of evidence localization through hybrid recall and evidence chain alignment; it addresses the problem of natural language answer organization through controlled generation; it addresses the problems of verifiability and acceptability through declaration-level trust verification; and it addresses the problem of continuous adaptation after knowledge changes through feedback updates. Overall, this application does not simply treat the knowledge graph as an external database for AI querying, nor does it merely add several retrieval fragments before the input of a large language model. Instead, it forms a methodological process around the intelligent question answering process: "knowledge fragment standardization—question-answering ontology modeling—structured question parsing—joint recall of graph paths and semantic evidence—evidence chain alignment—evidence-constrained generation—declaration-level trust verification—feedback evolution."

[0101] The following is in conjunction with the appendix Figure 3 , Figure 3 The diagram illustrates a processing flowchart of a knowledge graph-based intelligent question-answering method according to another embodiment of this application, which specifically includes the following steps: Step 302: User question input.

[0102] Step 304: Standardize knowledge fragments.

[0103] Specifically, this includes processing multi-source materials into evidence units to generate evidence units with source location, entity relationship, attribute time, and evidence hash. It also involves collecting and standardizing textual data, structured data, question-and-answer logs, and multimodal parsing results to form knowledge fragments with source, location, time, entity candidates, and quality scores.

[0104] Step 306: Question-answering ontology and graph construction.

[0105] Specifically, this includes constructing a domain ontology for intelligent question answering, defining entity types, relation types, attribute fields, question intent, answer types, evidence templates, and output templates; building a knowledge graph based on evidence units, and establishing binding indexes between graph nodes / edges and evidence units. Then, knowledge is extracted and integrated, writing entities, relations, attributes, and source evidence from knowledge fragments into the knowledge graph, and integrating or retaining alias entities, duplicate entities, and conflicting attributes.

[0106] Step 308: Problem Analysis.

[0107] Specifically, it parses user questions and multi-turn contexts, identifies question intent, entities, relationships, objects, limiting conditions, time conditions, and answer types, and generates a structured question representation.

[0108] Step 310: Hybrid Recall.

[0109] The hybrid recall in this application is not simply a concatenation of graph path results and vector retrieval results, but rather a simultaneous execution of graph path recall R_g, dense vector recall R_d, sparse keyword recall R_s, structured field recall R_f, metadata / time filtering R_m, and historical feedback recall R_h, with entity, relation, attribute, time, and answer type as gating conditions for fusion.

[0110] Specifically, the process involves performing graph path retrieval and semantic evidence recall. Based on structured questions, relevant nodes, relationships, and multi-hop paths are searched, while aligned original text evidence is recalled. This includes simultaneously performing graph path recall, dense vector recall, sparse keyword recall, structured field recall, metadata filtering, and feedback recall. The hybrid recall in this application is not simply concatenating graph path results with vector retrieval results; rather, it simultaneously performs graph path recall (R_g), dense vector recall (R_d), sparse keyword recall (R_s), structured field recall (R_f), metadata / time filtering (R_m), and historical feedback recall (R_h), fusing them with entity, relationship, attribute, time, and answer type as gating conditions. Candidate results that are semantically similar to the question but have inconsistent entities, opposite relationship directions, unmet time conditions, or cannot be bound to original text evidence are not allowed to enter the generation context.

[0111] Step 312: Align the chain of evidence.

[0112] Specifically, the process involves aligning and sorting evidence chains, binding nodes and edges in the graph path to original text fragments, structured fields, or multimodal evidence, and sorting them according to reliability, timeliness, path completeness, semantic relevance, and conflict risk. Entity, relation, attribute, time, answer type, and evidence binding gating are applied to candidate results to eliminate unalignable evidence. The process also includes constructing evidence packages, generating binding relationships between graph paths, evidence units, and declaration slots, and recording excluded evidence and the reasons for exclusion. This invention first decomposes the original data into evidence units, rather than simply extracting standardized semantic triples. Each evidence unit retains the original text excerpt, source location, entity mention, relation description, attribute value, time condition, and evidence hash, ensuring that subsequent graph edges, candidate paths, and answer declarations can all be traced back to specific evidence. The evidence package can be represented by the following formula 16: u_i =<eid_i, src_i, pos_i, quote_i, ent_i, rel_i, attr_i, time_i,auth_i, fresh_i, hash_i, emb_i> Formula 16 Where eid_i is the evidence unit number; quote_i is the value of the original text excerpt or structured field that can be displayed; hash_i is used to verify that the evidence has not been tampered with; auth_i and fresh_i represent the authority and timeliness of the source, respectively. Relational edges e in the graph are not directly considered as generating facts; only when Bind(e,u_i)=1 is the relational edge qualified to enter the evidence package.

[0113] Step 314: Controlled generation.

[0114] Specifically, the AI ​​model is invoked to generate a controlled answer. The structured question, candidate evidence, graph path, constraint rules, and output format requirements are input into the model to generate a draft answer. That is, the AI ​​model is invoked to generate a draft answer under the evidence package and prohibitive constraints. During the generation phase, the large language model is limited to an evidence organizer. The input context is organized according to "structured question, available evidence units, graph path, declaration slots, prohibitions, output format, and review threshold." Prohibited items include at least: not using facts outside the evidence package, not merging evidence from different time periods, not changing the direction of the relationship, not stating one side of conflicting evidence as a definitive conclusion, and not omitting low-confidence prompts.

[0115] Step 316: Declarative-level trusted verification.

[0116] Specifically, it performs declaration-level credibility verification and outputs the results, checks whether key claims are supported by evidence, calculates confidence levels, and outputs the answer, evidence chain, conflict warning, and supplementary explanations. In other words, it breaks down the initial draft of the answer into a set of claims and performs verification on each claim for evidence implication, path integrity, time validity, conflict status, and source reliability.

[0117] Step 318: Feedback Update.

[0118] Specifically, the system records feedback and updates the graph. Based on user adoption, corrections, evidence clicks, and manual review results, it incrementally updates entity relationships, source weights, evidence ranking, and generation strategies. This means outputting answers based on claim-level confidence, outputting cautiously, deleting and rewriting, or manually reviewing them. The feedback is then mapped back to evidence units, graph edges, and generation strategies to update the graph quality score and evidence ranking weights.

[0119] It should be noted that after the feedback is received in step 318, the process will return to step 310 to continue updating and executing.

[0120] This embodiment provides an example illustrating the input and output data of an AI model: Inputs: User question Q, multi-turn context Ctx, domain ontology O, knowledge graph G, evidence index V Output: Answer A, Chain of Evidence E, Confidence Level C, Risk Warning R T = ParseQuestion(Q, Ctx) # Parse intent, entities, relations, time, and answer type P = GraphSearch(G, T, O, max_hop, rel_set) # Search for candidate graph paths D = EvidenceRetrieve(V, Q, T, top_k) # Retrieve original text fragments or structured records E = AlignAndRank(P, D) # Align and sort paths and text evidence A0 = GenerateWithEvidence(T, E, constraints) # Generate a draft with constraints based on evidence C, R = VerifyClaims(A0, E, G) # Declaration-level trust verification and conflict identification A = ReviseOrEscalate(A0, C, R) # Revise, output cautiously, or transfer to manual review In summary, the knowledge graph-based intelligent question answering method provided in this application has at least the following beneficial effects: 1. More clearly defined application scenarios. This invention is designed for intelligent question-answering tasks in the field of artificial intelligence. It specifically models the question intent, entity relationships, answer types, and evidence chains, avoiding the problem of generalized chat-style generation schemes being difficult to interpret.

[0121] 2. Reduce the risk of AI illusions. This method requires that the generated results align with the knowledge graph path and the original text evidence, and de-emphasizes, deletes, or transfers key statements without supporting evidence to human review.

[0122] 3. Improve answer interpretability. The method not only outputs the answer, but also the entity relationship path, source location, original text fragment, time information, and confidence level, enabling users to trace the source of the answer.

[0123] 4. Enhance multi-hop reasoning capabilities. Knowledge graphs can explicitly represent entities, attributes, events, and relationship paths, overcoming the difficulty of simple vector retrieval in handling structured relationships.

[0124] 5. Improve evidence consistency control. The method checks entity consistency, relation direction, attribute conditions, time conditions, and source conflicts in the recall, alignment, and generation stages to reduce erroneous splicing.

[0125] 6. Improve low-confidence output control. When evidence is insufficient, conflicts cannot be resolved, or key conditions are missing, the method should output a request for supplementary information or a prompt for manual review, rather than providing a definitive answer.

[0126] 7. Enhance continuous evolution capabilities. User error correction, evidence clicks, and manual review results can influence the graph quality score, evidence weight, and generation strategy, enabling the method to continuously optimize as knowledge evolves.

[0127] Based on this, this application takes intelligent question answering tasks as the clear application boundary, knowledge graph paths as the factual skeleton, original text evidence as the basis, answer types and constraint rules as the generation boundary, and performs declaration-level credible verification after generation, thereby forming a stronger accuracy, interpretability, traceability and controllable output capability compared with ordinary RAG solutions.

[0128] It should be noted that this application can have various equivalent or alternative implementation methods. The knowledge graph can be stored using attribute graphs, RDF graphs, relational database simulated graph structures, or in-memory graph structures; the evidence index can use dense vectors, sparse vectors, keyword inverted indexes, structured indexes, or hybrid retrieval; the AI ​​model can use local models, private cloud models, model gateways, or multi-model collaboration; and the declaration-level verification can be implemented using natural language entailment models, fact consistency models, rule inferencers, graph constraint validators, or multi-model voting methods.

[0129] In scenarios with frequent knowledge updates, the method can adopt an incremental processing approach, triggering fragment segmentation, entity relationship extraction, evidence binding, and graph updates upon the arrival of new documents, web pages, question-and-answer records, or structured data. In scenarios with high response speed requirements, the evidence chain of high-frequency questions can be cached as a templated evidence package, and quickly matched and dynamically supplemented when users ask questions.

[0130] For scanned documents, images, charts, audio / video transcriptions, or complex tables, the method can first generate knowledge fragments through OCR, table parsing, speech transcription, or image descriptions, and then integrate them into the same graph evidence chain alignment process. This alternative approach only changes the way the evidence source is parsed, without altering the core technical process of "graph constraint—evidence retrieval—evidence chain alignment—controlled generation—credible verification—feedback update".

[0131] Each functional unit can be replaced with an equivalent one according to actual technical conditions: graph storage can be simulated by a relational database with indexes, vector retrieval can be supplemented by keyword retrieval and sparse retrieval, AI models can be implemented by models from different vendors or local models, and evidence verification can be achieved by a combination of rule models, implication models, or manual review. None of the above replacements affect the essential technical features of this invention.

[0132] Supplementary alternatives include: evidence chain alignment can be achieved using natural language entailment models, cross-coding and sorting models, rule templates, graph neural network path encoders, or multi-model voting; evidence gating can be achieved using hard thresholds, soft weights, learned sorting, or manual review queues; and declaration-level credibility can be achieved using minimum declaration confidence, core declaration weighted confidence, or risk level segmented output. These alternative implementations do not change the core idea of ​​the three-level binding of "evidence unit - graph path - answer declaration" and the generation of evidence constraints.

[0133] Corresponding to the above method embodiments, this application also provides embodiments of an intelligent question-answering device based on knowledge graphs. Figure 4 This diagram illustrates the structure of a generative model-based recommendation model training device according to an embodiment of this application. Figure 4 As shown, the device includes: The identification module 402 is configured to acquire and identify user questions and context information, and generate a structured question representation, wherein the structured question representation includes question intent, entities, relationships, attributes, limiting conditions, time and answer type; The knowledge graph path retrieval module 404 is configured to perform knowledge graph path retrieval based on the structured question representation to obtain candidate knowledge graph data, wherein the candidate knowledge graph data includes entity nodes, edges, relations, attributes, time, and answer source; The semantic retrieval module 406 is configured to perform semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question, wherein the answer statement slot represents the slot information of the answer type statement required to answer the user question; The evidence package construction module 408 is configured to construct the binding relationship between the candidate graph data, the answer declaration slot and the candidate original text evidence, and generate a candidate evidence package; Answer generation module 410 is configured to call a large language model, input the structured question representation and the candidate evidence package into the large language model, and output a draft answer corresponding to the user question; The answer credibility verification module 412 is configured to split the initial answer based on the answer declaration slot to obtain at least one answer unit, and to perform credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question.

[0134] Optionally, the device further includes: The multi-source data processing module is configured to first acquire multi-source data, standardize the multi-source data, and abstract it into at least one knowledge fragment with attribute information, including source identifier, location identifier, time information, and quality score. A knowledge graph is constructed based on at least one knowledge fragment with attribute information.

[0135] Optionally, the semantic retrieval module 406 is further configured to: Based on the structured problem representation and the candidate graph data, a similarity calculation is performed, and the graph data with a similarity that meets a preset threshold is used as the recalled knowledge fragment; Determine the answer statement slot and candidate original text evidence corresponding to the user question in the recalled knowledge fragment.

[0136] Optionally, the evidence package construction module 408 is further configured to: Based on the slot declared in the answer, extract the map data to be aligned from the candidate map data; Based on the answer statement slot, extract the original text evidence to be aligned from the candidate original text evidence; Calculate the alignment score between the target graph data and the target original text evidence. The alignment score represents the consistency relationship between the attribute content of the knowledge fragment corresponding to each answer statement slot and the attribute content of the extracted original text evidence. Based on the alignment score, the candidate graph data, the answer statement slot, and the candidate original text evidence, a candidate evidence package is constructed.

[0137] Optionally, the evidence package construction module 408 is further configured to: If the alignment score is determined to be lower than a preset alignment score threshold, the candidate graph data corresponding to the alignment score and the candidate original text evidence are determined as excluded answer content. Record the answer declaration slot, the content of the excluded answer, and the reason for excluding the content of the excluded answer.

[0138] Optionally, the answer generation module 410 is further configured to: The large language model is invoked, and the structured question representation, candidate evidence package, constraints, and output answer template are input into the large language model. The structured question representation includes the identified answer entities, the candidate evidence package includes candidate answer types, graph paths, and original text evidence that completes the alignment score calculation, the constraints include exclusion constraints for original text evidence that does not meet the alignment score conditions, and the output answer template includes risk level categories. Obtain a draft answer output by the large language model, wherein each answer unit in the draft answer is bound to at least one evidence number and risk level.

[0139] Optionally, the answer trust verification module 412 is further configured as follows: Calculate the confidence level corresponding to at least one answer unit; If the confidence level is greater than a first preset threshold, output at least one answer unit corresponding to the user question and the confidence level; If the confidence level is less than or equal to the first preset threshold and greater than the second preset threshold, at least one answer unit and the risk level corresponding to the answer unit shall be output. If the confidence level is less than or equal to the second preset threshold, output supplementary conditions or manual review prompts corresponding to the user's question.

[0140] Optionally, the device further includes: The event feedback module is configured to record user feedback events, wherein the feedback events include explicit feedback events and implicit feedback events. The explicit feedback events include user likes, copy operations, evidence clicks, correction operations, or manual review feedback events. The implicit feedback events include whether the user continues to ask follow-up questions and whether the same question appears repeatedly. The feedback event is mapped to the knowledge graph, and the graph data in the knowledge graph is updated; wherein, the update operation includes incremental updates to the entity relationships, evidence source weights, evidence ranking, and generation strategies of the graph data.

[0141] The knowledge graph-based intelligent question-answering device provided in this application generates a structured question representation by recognizing user questions and contextual information. It can quickly complete knowledge graph path retrieval to obtain candidate graph data, and then perform semantic retrieval through a recall mechanism to obtain the answer statement slots and candidate original text evidence corresponding to the user question. Subsequently, it aligns and binds the graph path, original text evidence, and answer statement slots to generate a traceable evidence chain. Then, it calls a large language model to output the answer corresponding to the user question, achieving controlled generation of the large model within a traceable evidence package. Content without bound evidence cannot be output as a definitive answer, unlike selecting the optimal candidate answer based on semantic similarity. Based on this, this application uses intelligent question-answering tasks as a clear application boundary, knowledge graph paths as the factual skeleton, original text evidence as the basis, and answer type and constraint rules as the generation boundary. After generation, it performs declaration-level credibility verification, thus achieving stronger accuracy, interpretability, traceability, and controllable output capabilities compared to ordinary RAG solutions.

[0142] The above is an illustrative scheme of a knowledge graph-based intelligent question-answering device according to this embodiment. It should be noted that the technical solution of this knowledge graph-based intelligent question-answering device and the technical solution of the knowledge graph-based intelligent question-answering method described above belong to the same concept. Details not described in detail in the technical solution of the knowledge graph-based intelligent question-answering device can be found in the description of the technical solution of the knowledge graph-based intelligent question-answering method described above.

[0143] Figure 5 A structural block diagram of a computing device 500 according to an embodiment of this application is shown. The components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0144] The computing device 500 also includes an access device 540, which enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 540 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0145] In one embodiment of this application, the aforementioned components of the computing device 500 and Figure 5 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 5 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this application. Those skilled in the art can add or replace other components as needed.

[0146] The computing device 500 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 500 can also be a mobile or stationary server.

[0147] Specifically, the processor 520 implements the steps of the knowledge graph-based intelligent question-answering method when executing the computer instructions.

[0148] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above-described intelligent question answering method based on knowledge graphs belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the above-described intelligent question answering method based on knowledge graphs.

[0149] An embodiment of this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the knowledge graph-based intelligent question-answering method described above.

[0150] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the knowledge graph-based intelligent question answering method described above. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the knowledge graph-based intelligent question answering method described above.

[0151] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0152] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the 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. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0153] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0154] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0155] The preferred embodiments disclosed above are merely illustrative of this application. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this application. These embodiments are selected and specifically described in this application to better explain the principles and practical applications of this application, thereby enabling those skilled in the art to better understand and utilize this application. This application is limited only by the claims and their full scope and equivalents.

Claims

1. A knowledge graph-based intelligent question answering method, characterized in that, include: The system acquires and identifies user questions and contextual information to generate a structured question representation, which includes question intent, entities, relationships, attributes, limiting conditions, time, and answer type. Based on the structured question representation, a knowledge graph path retrieval is performed to obtain candidate graph data, wherein the candidate graph data includes entity nodes, edges, relations, attributes, time, and answer source; Semantic retrieval is performed based on the structured question representation and the candidate graph data to obtain the answer statement slots and candidate original text evidence corresponding to the user question, wherein the answer statement slots represent the slot information of the answer type statement required to answer the user question; Construct the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence to generate a candidate evidence package; The large language model is invoked, and the structured question representation and the candidate evidence package are input into the large language model to output the initial draft of the answer corresponding to the user question; The initial draft answer is split based on the answer declaration slot to obtain at least one answer unit, and the at least one answer unit is verified to obtain the target answer corresponding to the user question.

2. The method according to claim 1, characterized in that, Before performing knowledge graph path retrieval based on the structured question representation to obtain candidate graph data, the method further includes: First, acquire multi-source data, standardize the multi-source data, and abstract it into at least one knowledge fragment with attribute information, including source identifier, location identifier, time information, and quality score; A knowledge graph is constructed based on at least one knowledge fragment with attribute information.

3. The method according to claim 2, characterized in that, The step of performing semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question includes: Based on the structured problem representation and the candidate graph data, a similarity calculation is performed, and the graph data with a similarity that meets a preset threshold is used as the recalled knowledge fragment; Determine the answer statement slot and candidate original text evidence corresponding to the user question in the recalled knowledge fragment.

4. The method according to claim 1, characterized in that, The process of constructing the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence to generate a candidate evidence package includes: Based on the slot declared in the answer, extract the map data to be aligned from the candidate map data; Based on the answer statement slot, extract the original text evidence to be aligned from the candidate original text evidence; Calculate the alignment score between the target graph data and the target original text evidence. The alignment score represents the consistency relationship between the attribute content of the knowledge fragment corresponding to each answer statement slot and the attribute content of the extracted original text evidence. Based on the alignment score, the candidate graph data, the answer statement slot, and the candidate original text evidence, a candidate evidence package is constructed.

5. The method according to claim 4, characterized in that, The process of constructing a candidate evidence package based on the alignment score binding the candidate graph data, the answer declaration slot, and the candidate original text evidence includes: If the alignment score is determined to be lower than a preset alignment score threshold, the candidate graph data corresponding to the alignment score and the candidate original text evidence are determined as excluded answer content. Record the answer declaration slot, the content of the excluded answer, and the reason for excluding the content of the excluded answer.

6. The method according to claim 1, characterized in that, The process of calling the large language model involves inputting the structured question representation and the candidate evidence package into the large language model, and outputting a draft answer corresponding to the user question, including: The large language model is invoked, and the structured question representation, candidate evidence package, constraints, and output answer template are input into the large language model. The structured question representation includes the identified answer entities, the candidate evidence package includes candidate answer types, graph paths, and original text evidence that completes the alignment score calculation, the constraints include exclusion constraints for original text evidence that does not meet the alignment score conditions, and the output answer template includes risk level categories. Obtain a draft answer output by the large language model, wherein each answer unit in the draft answer is bound to at least one evidence number and risk level.

7. The method according to claim 6, characterized in that, The step of performing credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question includes: Calculate the confidence level corresponding to at least one answer unit; If the confidence level is greater than a first preset threshold, output at least one answer unit corresponding to the user question and the confidence level; If the confidence level is less than or equal to the first preset threshold and greater than the second preset threshold, at least one answer unit and the risk level corresponding to the answer unit shall be output. If the confidence level is less than or equal to the second preset threshold, output supplementary conditions or manual review prompts corresponding to the user's question.

8. The method according to claim 2, characterized in that, The method further includes: Record user feedback events, wherein the feedback events include explicit feedback events and implicit feedback events. The explicit feedback events include user likes, copy operations, evidence clicks, correction operations, or manual review feedback events. The implicit feedback events include whether the user continues to ask follow-up questions and whether the same question appears repeatedly. The feedback event is mapped to the knowledge graph, and the graph data in the knowledge graph is updated; wherein, the update operation includes incremental updates to the entity relationships, evidence source weights, evidence ranking, and generation strategies of the graph data.

9. A knowledge graph-based intelligent question-answering device, characterized in that, include: The acquisition and recognition module is configured to acquire and recognize user questions and context information, and generate a structured question representation, which includes question intent, entities, relationships, attributes, limiting conditions, time, and answer type. The knowledge graph path retrieval module is configured to perform knowledge graph path retrieval based on the structured question representation to obtain candidate knowledge graph data, wherein the candidate knowledge graph data includes entity nodes, edges, relations, attributes, time, and answer source; The semantic retrieval module is configured to perform semantic retrieval based on the structured question representation and the candidate graph data to obtain the answer statement slot and candidate original text evidence corresponding to the user question, wherein the answer statement slot represents the slot information of the answer type statement required to answer the user question; The evidence package construction module is configured to construct the binding relationship between the candidate graph data, the answer declaration slot, and the candidate original text evidence, and generate a candidate evidence package; The answer generation module is configured to call a large language model, input the structured question representation and the candidate evidence package into the large language model, and output a draft answer corresponding to the user question; The answer credibility verification module is configured to split the initial answer based on the answer declaration slot to obtain at least one answer unit, and to perform credibility verification on the at least one answer unit to obtain the target answer corresponding to the user question.

10. A computing device, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the computer instructions, it implements the steps of the method according to any one of claims 1-8.

Images