A knowledge question and answer recording method and system fusing traceability verification

Abstract

The application discloses a knowledge question and answer record method and system fusing traceability verification, and belongs to the technical field of information retrieval and natural language processing. The method comprises the following steps: constructing a question and answer record basic block, collecting question and answer original information, extracting a knowledge source node, and establishing a preliminary mapping relationship; generating a forward traceability path chain, defining an evidence anchor point sequence, constructing a directed path, and calculating path transition confidence; performing bidirectional verification, generating bidirectional sequences through forward traversal and reverse traversal, performing consistency comparison of the evidence anchor point sequence, calculating traceability consistency scores, and determining a credible level; completing final storage and index updating of the question and answer record, differentiating processing records according to the credible level, generating final record entries with verification stamps, establishing double-path indexes, and updating a global traceability quality statistical table. The application realizes bidirectional closed-loop verification and quantitative evaluation of the traceability logic of the question and answer record, and effectively improves the authority, reliability and verifiability of the knowledge question and answer record.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and knowledge management technology, specifically to a knowledge question-and-answer recording method and system that integrates source tracing and verification. Background Technology

[0002] With the development of artificial intelligence technology, knowledge-based question-answering systems have been widely applied in various fields such as enterprise internal knowledge management, medical clinical decision support, and online legal services. Existing knowledge-based question-answering systems generally adopt a one-way source annotation mechanism when recording the question-and-answer process; that is, after generating an answer, they simply record the knowledge base entries or document fragments referenced by the answer.

[0003] This one-way tracing method only establishes a link from the answer to the knowledge source, and cannot verify whether this link constitutes a logical loop. The system may reference a knowledge source node to support an answer, but it cannot confirm whether the knowledge source node truly contains the logic that leads to the answer, nor can it detect whether there are contradictory circular references between different parts of the answer and multiple knowledge source nodes.

[0004] Furthermore, existing technologies lack quantitative assessment methods for the logical consistency of tracing paths. During subsequent audits or backtracking, it is difficult to confirm whether the answer is truly and credibly derived from the source. It is also impossible to effectively detect illusions or contradictions caused by logical gaps or erroneous connections between knowledge source nodes. This seriously affects the authority, reliability, and verifiability of knowledge question-and-answer records, limiting the application of knowledge question-and-answer systems in scenarios with high credibility requirements. Summary of the Invention

[0005] The technical problems to be solved by this invention include at least one of the following: the existing knowledge question-and-answer record tracing mechanism is a one-way open-loop structure, which cannot verify whether the reference relationship from the answer to the knowledge source constitutes a logical closed loop; the existing technology lacks the means to detect the logical consistency between multiple knowledge source nodes, and cannot detect the logical gaps or contradictory references between knowledge source nodes; the existing technology fails to quantitatively evaluate the reliability of the tracing path, and cannot distinguish between high-quality references and low-quality references; the existing technology does not perform differentiated processing based on the credibility of tracing when storing question-and-answer records, resulting in low-quality records entering the formal knowledge base.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solutions.

[0007] A knowledge question-and-answer recording method integrating traceability and verification includes the following steps:

[0008] Step 1: Construct the basic block of question and answer records, collect the original information of questions and answers and bind timestamps, extract all knowledge source nodes referenced by the answer item and assign them a globally unique traceability identifier, divide the answer item into independent semantic fragments and establish a mapping pair list with the corresponding knowledge source nodes, and encapsulate them into structured data objects;

[0009] Step 2: Generate a forward tracing path chain. Identify the evidence anchors that play a key logical supporting role in the answer items and determine the sequence of evidence anchors according to semantic structure. Connect each evidence anchor with the corresponding knowledge source node in sequence, starting from the question item, to form a directed path. Assign a quantified path transfer confidence to each directed edge.

[0010] Step 3: Perform bidirectional verification. Perform forward and reverse traversal respectively to generate forward node identifier sequences and reverse node identifier sequences. Extract the evidence anchor point sequences from the two sequences, perform mirror symmetry comparison and record the source conflict points. Calculate the source consistency score based on the matching results, path transition confidence and sequence consistency. Determine the credibility level of the record based on the score and generate a verification report.

[0011] Step 4: Complete the final storage and index update of the question and answer records. Differentiate the question and answer records according to their trust level, generate final question and answer record entries with verification stamps, write them to storage, and establish a dual-path index based on content and source. Update the global traceability quality statistics table.

[0012] In one embodiment of the present invention, step 1, which involves constructing the basic block of question-and-answer records, includes:

[0013] Step 1-1: Collect the original question and answer information, obtain the query statement submitted by the user as the question item, record the generated text reply as the answer item, and generate a unified timestamp to bind to the question item and the answer item;

[0014] Step 1-2: Extract knowledge source nodes. During the process of generating answer items, record all referenced knowledge source nodes simultaneously and assign a globally unique source identifier to each knowledge source node.

[0015] Steps 1-3: Establish preliminary mapping relationships, segment the answer item text into independent semantic fragments, associate each answer fragment with the corresponding knowledge source node, and generate a list of mapping pairs;

[0016] Steps 1-4: Encapsulate the basic block of the question and answer record, encapsulating the question item, answer item, knowledge source node, mapping pair list and timestamp into a structured data object and temporarily storing it in the cache area.

[0017] In one embodiment of the present invention, step 2, generating a forward tracing path chain, includes:

[0018] Step 2-1: Define the evidence anchor sequence, identify the evidence anchors that play a key logical supporting role in the answer items, and determine the evidence anchor sequence according to the semantic structure of the answer items;

[0019] Step 2-2: Construct a forward tracing path chain. Starting with the question item as the logical starting point, connect the corresponding knowledge source nodes in sequence according to the order of the evidence anchor sequence to form a directed forward tracing path chain.

[0020] Steps 2-3: Calculate the path transition confidence and assign a quantified path transition confidence to each directed edge in the forward tracing path chain;

[0021] Steps 2-4: Solidify the forward tracing path chain, and integrate the forward tracing path chain sequence with the confidence of each path transition into a structured forward tracing path chain object.

[0022] In one embodiment of the present invention, step 3, performing bidirectional verification, includes:

[0023] Step 3-1: Perform bidirectional path traversal and sequence generation, performing forward and reverse traversal respectively to generate forward node identifier sequences and reverse node identifier sequences;

[0024] Step 3-2: Perform consistency comparison of evidence anchor sequence, extract evidence anchor sequence from the forward node identifier sequence and the reverse node identifier sequence, align and compare item by item and record the source conflict points;

[0025] Step 3-3: Calculate the source consistency score, based on the number of successfully matched evidence anchors, path transition confidence, and sequential consistency to calculate the comprehensive score;

[0026] Steps 3-4: Determine the credibility level of the records and generate a verification report. Based on the traceability consistency score, divide the question and answer records into different credibility levels and generate a verification report containing details of the conflict points.

[0027] In one embodiment of the present invention, step 4, which completes the final storage and index update of the question-and-answer records, includes:

[0028] Step 4-1: Filter and correct records based on their trust level, and adopt different processing strategies for question and answer records with different trust levels;

[0029] Step 4-2: Generate the final question and answer record entry with a verification stamp, embed the verification report in the basic question and answer record block to form the final question and answer record entry with a verification stamp;

[0030] Step 4-3: Store and build a dual-path index, write the final question and answer record entries to the persistent storage medium, and build a dual-path index based on content and source;

[0031] Step 4-4: Update the global source traceability quality statistics table, and record the number of citations, average source traceability consistency score, and most recent citation time for each knowledge source node.

[0032] In one embodiment of the present invention, in steps 1-2, the source identifier includes at least the library number of the source knowledge base, the document number of the specific document, and the location information of the knowledge source node in the document.

[0033] In one embodiment of the present invention, in steps 2-3, the confidence level of the path transfer from the evidence anchor point to the knowledge source node is jointly determined by the semantic matching degree and word frequency of the evidence anchor point in the content of the corresponding knowledge source node, as well as the logical importance of the evidence anchor point in the answer item.

[0034] In one embodiment of the present invention, step 3-2, the content of the evidence anchor sequence consistency comparison includes: checking whether there is a unique and semantically corresponding reverse evidence anchor in the reverse evidence anchor sequence for each positive evidence anchor; checking whether the order of appearance of the corresponding evidence anchor in their respective sequences is consistent; and checking whether the knowledge source nodes associated with the corresponding evidence anchor are consistent.

[0035] In one embodiment of the present invention, in step 4-1, high-confidence records are fully retained, medium-confidence records trigger an auxiliary correction process, and low-confidence records are transferred to the pending review area and the administrator is notified for manual review.

[0036] This invention also provides a knowledge question-and-answer record system that integrates traceability and verification, comprising:

[0037] The question and answer record basic block construction unit is used to collect the original information of questions and answers and bind timestamps, extract all knowledge source nodes referenced by the answer item and assign them a globally unique traceability identifier, divide the answer item into independent semantic fragments and establish a mapping pair list with the corresponding knowledge source nodes, encapsulate and generate the question and answer record basic block and output it to the forward traceability path chain generation unit;

[0038] The forward tracing path chain generation unit is used to receive the question and answer record basic block, identify the evidence anchor points that play a key logical supporting role in the answer items and determine the evidence anchor point sequence according to the semantic structure, connect each evidence anchor point with the corresponding knowledge source node in sequence with the question item as the logical starting point to form a directed path, assign a quantified path transfer confidence to each directed edge, generate a forward tracing path chain object and output it to the bidirectional verification unit.

[0039] The bidirectional verification unit is used to receive the forward tracing path chain object, perform forward traversal and reverse traversal respectively to generate forward node identifier sequence and reverse node identifier sequence, extract the evidence anchor point sequence in the two sequences, perform mirror symmetry comparison and record the tracing conflict points, calculate the tracing consistency score based on the matching result, path transition confidence and sequence consistency, determine the credibility level of the record according to the score and generate a verification report, and output the question and answer record basic block, tracing consistency score, credibility level and verification report to the storage and indexing unit.

[0040] The storage and indexing unit is used to receive the question and answer record base block, traceability consistency score, trust level and verification report, perform differentiated processing on the question and answer records according to the trust level, generate the final question and answer record entries with verification stamps, store and establish a dual-path index based on content and source, and update the global traceability quality statistics table.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] This invention achieves bidirectional closed-loop verification of the question-and-answer record tracing logic by constructing a forward tracing path chain and a reverse tracing path chain, overcoming the defect of existing technologies where unidirectional tracing cannot verify the logical closed loop.

[0043] This invention, through consistency comparison of evidence anchor sequence, can accurately detect logical gaps, disordered sequences, and contradictory references between multiple knowledge source nodes, significantly improving the detection accuracy of tracing logic.

[0044] This invention introduces a source tracing consistency score as a quantitative indicator, which comprehensively considers three dimensions: the breadth of evidence anchor matching, the confidence of path transition, and the consistency of logical order, thereby achieving an objective quantitative assessment of the reliability of the source tracing path.

[0045] This invention classifies question-and-answer records into trust levels based on traceability consistency scores and adopts differentiated storage strategies for different levels, effectively preventing low-trust records from entering the formal knowledge base and improving the overall authority and verifiability of knowledge question-and-answer records.

[0046] This invention establishes a dual-path index and a global source traceability quality statistics table, which supports both content-based fast retrieval and source-based reverse tracing queries, while also providing data support for the dynamic optimization of the knowledge base. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of the knowledge question and answer recording method that integrates traceability and verification according to the present invention.

[0048] Figure 2 This is a schematic diagram of the forward traceability path chain structure in the Q&A embodiment of the enterprise rules and regulations of the present invention.

[0049] Figure 3 This is a schematic diagram of the module architecture of the knowledge question-and-answer record system that integrates traceability and verification according to the present invention.

[0050] Figure 4 This is a schematic diagram of the backend interface for the unified management and quality monitoring of knowledge source nodes in the knowledge question and answer record system that integrates traceability and verification according to the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0053] like Figure 1 As shown, the first embodiment of the present invention provides a knowledge question-and-answer record method that integrates source tracing and verification. The method includes the following steps:

[0054] Step 1: Construct the basic blocks of the question-and-answer record. Step 1 includes the following steps:

[0055] Step 1-1: Collect raw question and answer information. Obtain the user-submitted query statement through a pre-defined interface and record it as a question item. Simultaneously, retrieve and generate a text response corresponding to the question item from the source knowledge base and record this response as an answer item. To ensure time-series traceability, generate a unified timestamp for this question and answer interaction. This timestamp records the moment the question and answer occurred and is bound to both the question item and the answer item.

[0056] Steps 1-2: Extracting Knowledge Source Nodes. During the generation of answer items, all knowledge sources referenced by the answer items are recorded simultaneously. Each referenced knowledge source, whether it's a paragraph in a document, a record in a database, or an entity and its attributes in a knowledge graph, is defined as an independent knowledge source node. Each knowledge source node is assigned a globally unique traceability identifier, which includes at least the library number of the source knowledge base, the document number of the specific document, and the location information of the knowledge source node within the document (such as byte offset or paragraph number).

[0057] Steps 1-3: Establish preliminary mapping relationships. Establish preliminary mapping relationships between the answer items and the knowledge source nodes extracted in Step 1-2.

[0058] Specifically, the answer text is segmented into semantically independent fragments, typically sentences or key phrases. For each answer fragment, it is associated with one or more knowledge source nodes that support it, generating a list of mapping pairs. Each entry in the list contains an answer fragment identifier and at least one source identifier, indicating which specific knowledge source nodes the answer fragment originates from.

[0059] Steps 1-4: Encapsulate the question-and-answer record base block. The question items, answer items, all knowledge source nodes and their identifiers obtained in the previous steps, and the mapping pair list generated in Steps 1-3, along with the timestamps, are encapsulated in a structured data object, defined as the question-and-answer record base block. This question-and-answer record base block serves as the raw input for all subsequent processing flows and is temporarily stored in a cache.

[0060] Step 2: Generate a forward tracing path chain.

[0061] This step aims to transform the discrete mapping relationship established in step 1 into a structured path with a clear direction and logical order, so as to accurately describe the reasoning trajectory from the user's question to the original knowledge source. Step 2 further includes the following steps:

[0062] Step 2-1: Define the evidence anchor sequence. Based on the question-and-answer record base block, analyze the answer items to identify key logically supporting words or phrases, and define these words or phrases as evidence anchors. Evidence anchors are semantic bridges connecting answer items and knowledge source nodes. Based on the semantic structure of the answer items and the content of the knowledge source nodes, determine an evidence anchor sequence that reflects the logical evolution from introducing the topic to drawing a conclusion within the answer item.

[0063] Step 2-2: Execute forward tracing path chain construction.

[0064] Starting with the question item, the process begins by locating the first evidence anchor in the answer item's evidence anchor sequence. Next, based on the mapping list established in steps 1-3, the knowledge source node corresponding to this evidence anchor is found. Then, within the content of the knowledge source node, the specific text position semantically matching the evidence anchor is located. This process logically forms a directed edge from the "question item" to "evidence anchor A1 in the answer item" and then to "knowledge source node K1." Following the order of the evidence anchor sequence, each evidence anchor is processed sequentially, connecting it to its corresponding knowledge source node. Finally, all the connected paths are organized into a directed path starting from the question item, passing through each evidence anchor sequentially, and ultimately reaching each knowledge source node. This path is defined as a forward tracing path chain.

[0065] Step 2-3: Calculate the path transition confidence. During the construction of the forward tracing path chain, each directed edge in the chain is assigned a quantified path transition confidence. This path transition confidence characterizes the reliability of reasoning or association from an upstream node to a downstream node.

[0066] Specifically, for a connection from an evidence anchor to a knowledge source node, the path transfer confidence is jointly determined by the semantic matching degree and word frequency of the evidence anchor in the content of the corresponding knowledge source node, as well as the logical importance of the evidence anchor in the answer item. The higher the semantic matching degree, the more prominent the appearance, and the more crucial the logic, the higher the path transfer confidence. Preferably, for a connection from a question item to the first evidence anchor, the path transfer confidence can be calculated based on the overall semantic relevance between the question item and the answer item.

[0067] Step 2-4: Consolidate the forward tracing path chain. Integrate the forward tracing path chain sequence constructed in Step 2-2 with the path transition confidence scores calculated in Step 2-3 to generate a structured forward tracing path chain object. This forward tracing path chain object not only records the order of all nodes in the forward tracing path chain, but also records each directed edge between adjacent nodes in the forward tracing path chain and its corresponding path transition confidence score.

[0068] Step 3: Perform bidirectional verification. Step 3 further includes the following steps:

[0069] Step 3-1: Perform bidirectional path traversal and sequence generation. While recording the forward tracing path chain, perform a reverse derivation to generate the mirror path chain of the forward tracing path chain, i.e., the reverse tracing path chain.

[0070] First, perform a forward traversal: starting from the first evidence anchor in the answer item, follow the direction of the edges in the forward tracing path chain generated in step 2, and sequentially visit each evidence anchor and knowledge source node on the path, recording the sequence of node identifiers traversed, denoted as F. Seq =[A1,K1,A2,K2,…].

[0071] Next, perform a reverse traversal: starting from each knowledge source node recorded in the forward node identifier sequence. Starting from this point, try to find the anchor points of evidence in the answer that support the connection between the logical deduction and the corresponding evidence in the answer. And ultimately, it can respond to the logical chain of the question item.

[0072] Specifically, for knowledge source nodes Search for anchor points of evidence within its content. Identify semantically related keywords or logical predicates, and use this information as clues to try to locate logically connected evidence anchors within the answer choices. If it is possible to establish... arrive The semantic backtracking link then continues to attempt from We then reverse-engineer whether it logically responds to the fundamental query intent of the question item. After traversing all knowledge source nodes, we record the node identifier sequence of the knowledge source nodes that have successfully established semantic backlinks, according to their logical order corresponding to the forward node identifier sequence, denoted as R. Seq =[K 1' A 1' ,K 2' A 2' It should be noted that, due to the possibility of failure or path bifurcation in the reverse derivation, R... Seq The knowledge source nodes and their order in F may be related to F Seq Not exactly the same.

[0073] Step 3-2: Perform consistency comparison of the evidence anchor sequence. Obtain the positive node identifier sequence F. Seq and the reverse node identifier sequence R Seq Then, the two are compared. The focus of the comparison is the mirror symmetry of the evidence anchor sequence. From F Seq Extract all evidence anchor points and form a positive evidence anchor point sequence F. AnchorSeq =[A1,A2,A3,…]. Accordingly, from R Seq Extract all the evidence anchors located in the answer items through reverse derivation, and form a reverse evidence anchor sequence R. AnchorSeq =[A 1' A 2' A 3' ,…].

[0074] Subsequently, the two evidence anchor sequence sequences were aligned and compared item by item.

[0075] The comparison includes: First, examining each anchor point of positive evidence. Is it in the reverse evidence anchor sequence R? AnchorSeq There exists a unique and semantically corresponding reverse evidence anchor point. Second, check and Third, check whether the order of appearance in their respective sequences is consistent; Related knowledge source nodes and Related knowledge source nodes Consistency. If the following conditions are met: existence, uniqueness, consistent order, and consistency of the associated knowledge source nodes, then the evidence anchor pair is determined to pass the mirror comparison. If a positive evidence anchor is missing in the reverse node identifier sequence, corresponds to multiple nodes, is out of order, or is associated with different source knowledge source nodes, then that position is recorded as a source tracing conflict point.

[0076] Step 3-3: Calculate the traceability consistency score.

[0077] This invention introduces comprehensive quantitative indicators to evaluate the reliability of the entire forward traceability path chain. The traceability consistency score... Calculated in the following way:

[0078]

[0079] in: This represents the total number of evidence anchors in the forward tracing path chain. This indicates the number of evidence anchor pairs that were successfully matched in the mirror comparison in step 3-2. Indicates the first The path transition confidence of the corresponding forward tracing path chain (from the knowledge source node to the evidence anchor) for each successfully matched evidence anchor is calculated in step 2-3. It is a sequence consistency coefficient between 0 and 1. It is 1 when the order of the positive evidence anchor sequence and the negative evidence anchor sequence is completely consistent. If there is a sequence inconsistency, it is deducted according to the degree of inconsistency. These are preset weighting coefficients, and they satisfy... . Used to measure the breadth of evidence anchor matching. Used to measure the quality of the matching path (i.e., path transition confidence). Used to measure the consistency of logical order. Preferably, it can be set to... To balance the quantity and quality of matches.

[0080] In the above methods for assessing the reliability of the forward tracing path chain, not only is the number of evidence anchors that can be verified in reverse ( It also considered the credibility of these verified positive tracing paths themselves. ), and whether the order of the entire logical flow is reversible ( This results in a traceability consistency score that comprehensively reflects the quality of the traceability logic. Its value is between 0 and 1.

[0081] Step 3-4: Determine the trust level of the records and generate a verification report. This is based on the traceability consistency score calculated in Step 3-3. In conjunction with the details of the source tracing conflict points discovered in step 3-2, the credibility of the current question and answer record is graded and determined.

[0082] Multiple score threshold ranges are preset, for example: when When a record is judged as "highly reliable," it indicates that the logical closed loop of the forward tracing path is rigorous and the two-way verification is highly consistent; when When a record is classified as "medium confidence," it indicates that the forward tracing path is basically established, but some links have weak connections or non-critical inconsistencies; when If a record is deemed "low credibility," it indicates that there are significant flaws or contradictions in the tracing logic.

[0083] Preferably, a structured verification report is generated, which includes the final traceability consistency score. The credibility level of the determination, the detailed location of all source conflict points (the specific evidence anchors and knowledge source nodes involved), and the description of the conflict type.

[0084] Step 4: Complete the final storage and index update of question-and-answer records. Based on the deep validation results of Step 3, this step differentiates the question-and-answer records, stores high-value records along with validation credentials, and updates the global index to support efficient retrieval and knowledge base optimization. Step 4 further includes the following steps:

[0085] Step 4-1: Filter and correct records based on trust level.

[0086] Based on the credibility level determined in steps 3-4, different processing strategies are adopted for the basic blocks of question-and-answer records. Records with "high credibility" are retained in their entirety. For records with "medium credibility," the system can trigger an auxiliary correction process: attempting to find semantically more semantically matching alternative knowledge source nodes in the knowledge base for the weakly related source tracing conflicts discovered in step 3-2. If an alternative knowledge source node is found and can improve the recalculated source consistency score to above the "high credibility" threshold, the original knowledge source node is replaced with the new knowledge source node, the mapping relationship and forward source tracing path chain are updated, and the record level is upgraded. For records with "low credibility," they are transferred to the pending review area or the abnormal log database, not included in the formal knowledge record database, and the administrator is notified for manual review.

[0087] Step 4-2: Generate the final question-and-answer record entry with a verification stamp. For records approved for storage (including original high-confidence records and corrected records), embed the verification report generated in Step 3 into the basic question-and-answer record block to form a complete final question-and-answer record entry with a verification stamp. This verification stamp is a digital credential of the record's credibility and includes a traceability consistency score. Trust level, verification timestamp, and key verification digests (such as the number of source conflict points).

[0088] Step 4-3: Store and build a dual-path index. Write the final question-and-answer record entries with verification stamps to persistent storage media such as a database. Simultaneously, build an efficient dual-path index to support fast retrieval: the first path index is constructed using keywords for the question and answer items, supporting content-based semantic retrieval; the second path index is constructed using identifiers of all knowledge source nodes referenced by the record, supporting source-based reverse tracing queries. When a user queries a record in the future, they can quickly find it through the first path index; when a knowledge source node is updated or needs to be audited, all question-and-answer records referencing that knowledge source node can be quickly located through the second path index.

[0089] Step 4-4: Update the Global Source Tracing Quality Statistics Table. After the records are stored, update a global source tracing quality statistics table. This global source tracing quality statistics table records the number of times each knowledge source node is successfully cited, the average source tracing consistency score of all question-and-answer records in which it participates, and the time of its most recent citation. By analyzing this table, high-quality, highly reliable "authoritative" knowledge source nodes (frequently cited and with high average scores) can be identified, as well as "weak" knowledge source nodes that may have problems and frequently cause source tracing conflicts (with low average scores).

[0090] The present invention will be further illustrated by the following examples.

[0091] Example 1: Q&A on Corporate Rules and Regulations

[0092] Scenario: In the internal knowledge management system of a large enterprise, an employee queries "What should be done if an employee is rated C twice in the annual performance evaluation?" through a question-and-answer interface. The system generates an answer based on the enterprise's rules and regulations knowledge base and performs integrated source verification.

[0093] Step 1: Construct the basic blocks for question and answer records.

[0094] An employee entered the query "What should be done if an employee is rated C twice consecutively in the annual performance evaluation?" into the Q&A interface. The system recorded this query as a question. Based on the company's rules and regulations knowledge base, the system retrieved relevant clauses and generated a text response: "According to Article 18 of the 'Employee Performance Management Measures' and Article 32 of the 'Labor Contract Management Regulations,' an employee who receives a C rating twice consecutively in the annual performance evaluation is considered incompetent for their current position. The employer should provide training or adjust their job position; if they are still incompetent after training or job adjustment, the employer may legally terminate the labor contract." This response was recorded as the answer. The system generated a unified timestamp for this Q&A interaction: "2024-03-15T09:23:17+08:00" and bound this timestamp to both the question and answer.

[0095] During the generation of answer items, the system simultaneously records all knowledge sources referenced by each answer item. Analysis revealed that the answer item referenced three independent knowledge sources:

[0096] The first source of knowledge is Article 18 of the "Employee Performance Management Measures" document, which clearly stipulates the performance rating standards and subsequent handling principles. The second source is Article 32 of the "Labor Contract Management Regulations" document, which stipulates the procedures for handling incompetence. The third source is the entity node "Incompetence" in the enterprise knowledge graph and its associated attributes, including its definition, identification standards, and legal consequences.

[0097] The three knowledge sources mentioned above are all defined as independent knowledge source nodes, each assigned a globally unique tracing identifier:

[0098] Knowledge source node K1: The traceability identifier is HR-POL-2023-001#SEC18#PARA1, where HR-POL-2023-001 is the library number of the source knowledge base (human resources policy library), SEC18 is the specific document number (article 18), and PARA1 is the location information of the knowledge source node in the document (first paragraph).

[0099] Knowledge source node K2: The traceability identifier is HR-POL-2023-002#SEC32#PARA1.

[0100] Knowledge source node K3: The traceability identifier is HR-KG-2023-001#ENT-NSW#ATTR-ALL, where HR-KG-2023-001 is the knowledge graph database number, ENT-NSW is the entity number, and ATTR-ALL is the location information of all attributes of the entity.

[0101] The system segments the answer text into segments with independent semantics:

[0102] Answer fragment S1: "According to Article 18 of the 'Employee Performance Management Measures' and Article 32 of the 'Labor Contract Management Regulations'";

[0103] Answer snippet S2: "Employees who receive a C grade in two consecutive annual performance evaluations are considered to be incompetent for their current position."

[0104] Answer snippet S3: "Employers should provide training or adjust job positions for them";

[0105] Answer snippet S4: "If an employee is still unable to perform their job duties after training or reassignment, the employer may terminate the employment contract in accordance with the law."

[0106] Create a list of mapping pairs:

[0107] Mapping to item 1: Answer fragment S1 corresponds to knowledge source nodes K1 and K2 (legal basis reference);

[0108] Mapping to item 2: Answer fragment S2 corresponds to knowledge source nodes K1 (performance level assessment) and K3 (definition of incompetence).

[0109] Mapping to entry 3: Answer fragment S3 corresponds to knowledge source node K2 (first step of the processing procedure);

[0110] Mapping to entry 4: Answer fragment S4 corresponds to knowledge source node K2 (second step of the processing procedure).

[0111] The system encapsulates the question item, answer item, three knowledge source nodes and their source identifiers, the above mapping pair list, and the timestamp "2024-03-15T09:23:17+08:00" into a basic block of question and answer records, and temporarily stores it in the system cache.

[0112] Step 2: Generate a forward tracing path chain.

[0113] The system analyzes the answer options and identifies four evidence anchors, which are arranged in logical order to form an evidence anchor sequence:

[0114] Anchor of evidence A1: "Performance evaluation grade C" (introducing the topic and corresponding to the performance grade determination);

[0115] Anchor of evidence A2: "Incompetence in the job" (logical progression, linking performance results with job competency);

[0116] Anchor of evidence A3: "Training or job transfer" (handling measures, legal obligations of the employing unit);

[0117] Anchor of evidence A4: "Termination of the labor contract in accordance with the law" (conclusion, corresponding to the final handling method).

[0118] Step 2-2: Execute forward tracing path chain construction.

[0119] Starting with the question item as the logical starting point, construct directed edges according to the sequence of evidence anchor points:

[0120] Directed edge E1: Problem item → Evidence anchor point A1. According to the mapping pair 1, A1 corresponds to knowledge source node K1 (the provisions on C-level rating in Article 18 of the "Employee Performance Management Measures").

[0121] Directed edge E2: Evidence anchor point A1 → Evidence anchor point A2. According to mapping entry 2, A2 corresponds to knowledge source nodes K1 and K3 (K1 contains the expression "considered incompetent", and K3 contains the definition "incompetent at work").

[0122] Directed edge E3: Evidence anchor point A2 → Evidence anchor point A3. According to the mapping pair 3, A3 corresponds to knowledge source node K2 (the provision of "providing training or adjusting job position" in Article 32 of the "Regulations on Labor Contract Management").

[0123] Directed edge E4: Evidence anchor point A3 → Evidence anchor point A4. According to the mapping pair item 4, A4 corresponds to the knowledge source node K2 (the same provision in the following item, "the labor contract may be terminated in accordance with the law").

[0124] All directed edges are organized into a forward tracing path chain that starts from the problem item, passes through evidence anchors A1, A2, and A3 in sequence, and finally reaches the knowledge source nodes K1, K2, and K3.

[0125] CF(E i ) represents the path transition confidence of the i-th directed edge, and its value is jointly determined by the semantic matching degree and word frequency of the evidence anchor in the content of the corresponding knowledge source node, as well as the logical importance of the evidence anchor in the answer item.

[0126] CF(E1): The semantic match between the question item "received a grade of C twice consecutively" and the answer fragments S1 and K1 "received a grade of C twice consecutively in annual performance evaluations" is 0.95, with significant word frequency. A1 has high logical importance in introducing the topic in the answer item. CF(E1) = 0.93.

[0127] CF(E2): The logical connection between A1 "Performance Evaluation Grade C" and A2 "Incompetent in the position" is directly established by "Considered incompetent" in K1, with a semantic matching degree of 0.91. The definition in K3 is completely consistent. CF(E2) = 0.90.

[0128] CF(E3): The logical connection between A2 "incompetent in the position" and A3 "training or reassignment" is directly established by K2 "should provide training or reassignment", with a semantic matching degree of 0.94. CF(E3) = 0.92.

[0129] CF(E4): The logical connection between A3 "training or job transfer" and A4 "termination of labor contract in accordance with law" is directly established by the preceding and following sentences of the same clause in K2, with a semantic matching degree of 0.96, and there is a conditional limitation of "still unable to perform the duties after...". CF(E4) = 0.94.

[0130] The system generates a forward tracing path chain object, recording the order of all nodes on the forward tracing path chain: problem item → A1 → A2 → A3 → A4 → K1 / K2 / K3, as well as four directed edges E1 to E4 and their path transition confidence. This forward tracing path chain object also records each directed edge between adjacent nodes on the forward tracing path chain and its corresponding path transition confidence.

[0131] Figure 2 This is a schematic diagram of the forward tracing path chain structure in the enterprise rules and regulations Q&A scenario in this embodiment. Taking the question "How should we deal with employees who are rated C twice in the annual performance evaluation?" as the logical starting point, the diagram presents four key evidence anchors in sequence according to the semantic evolution of the answer: "Performance evaluation C grade", "Incompetent for the job", "Training or reassignment", and "Legal termination of contract". It also shows the knowledge source node with a globally unique tracing identifier associated with each evidence anchor and marks the path transition confidence (CF) value corresponding to each reasoning link, demonstrating the complete traceable reasoning trajectory from the user's question to the final answer.

[0132] Step 3: Perform two-way verification.

[0133] While recording the forward tracing path chain, a reverse derivation is performed to generate a mirror path chain of the forward tracing path chain, i.e., the reverse tracing path chain.

[0134] Perform a forward traversal: Starting from the first evidence anchor in the answer, follow the direction of the edges in the forward tracing path chain generated in step 2, and visit each evidence anchor and knowledge source node on the path in turn, recording the sequence of node identifiers traversed, denoted as F. Seq =[A1,K1,A2,K3,A3,K2,A4,K2].

[0135] Perform a reverse traversal: start from each knowledge source node K recorded in the forward node identifier sequence. i Starting from this point, try to follow the reverse direction of logical deduction to find an anchor point A in the corresponding answer option that can support its connection. j And ultimately, it can respond to the logical chain of the question item.

[0136] Specifically, for knowledge source node K1, the phrase "received a C grade in two consecutive annual performance evaluations" in its content semantically matches the evidence anchor point A1 "performance evaluation C grade," establishing a semantic backlink. By deducing backward from A1, its content "received a C grade twice consecutively" directly addresses the fundamental query intent of the question item. The backlink is successfully established.

[0137] For knowledge source node K2, its content "provide training or adjust job position" semantically matches evidence anchor point A3 "training or job adjustment," and "can legally terminate the labor contract" semantically matches evidence anchor point A4 "legally terminate the labor contract." By deducing backwards from A3 and A4, its content fully responds to the fundamental query intent of the question item. The reverse link was successfully established.

[0138] For knowledge source node K3, the definition of "incompetent at work" in its content semantically matches the evidence anchor node A2's "incompetent in the position"; deducing backward from A2, it serves as a logical bridge from A1 to A3, fully responding to the fundamental query intent of the question item. The reverse link is successfully established.

[0139] Record the reverse node identifier sequence R Seq =[K1,A1,K3,A2,K2,A3,K2,A4].

[0140] Obtain the forward node identifier sequence F Seq and the reverse node identifier sequence R Seq Then, the two are compared. The focus of the comparison is the mirror symmetry of the evidence anchor sequence.

[0141] From F Seq Extract all evidence anchor points and form a positive evidence anchor point sequence F. AnchorSeq =[A1,A2,A3,A4]. Accordingly, from R Seq Extract all the evidence anchors located in the answer items through reverse derivation, and form a reverse evidence anchor sequence R. AnchorSeq =[A1,A2,A3,A4].

[0142] Subsequently, the two evidence anchor sequence sequences were aligned and compared item by item.

[0143] The comparison includes: First, examining each positive evidence anchor point A. i Is it in the reverse evidence anchor sequence R? AnchorSeq There exists a unique and semantically corresponding reverse evidence anchor point A. i Second, check A. i With A i 'Whether the order of appearance in their respective sequences is consistent; third, check A' iThe associated knowledge source node K i With A i 'The associated knowledge source node K' i 'Whether they are consistent. If they are present, unique, in the same order, and associated with the same knowledge source nodes, then the evidence anchor pair is determined to pass the mirror comparison. If a positive evidence anchor is missing in the reverse node identifier sequence, corresponds to multiple nodes, is in the wrong order, or is associated with different source knowledge source nodes, then that position is recorded as a source tracing conflict point.'

[0144] Item-by-item comparison results:

[0145] A1: There exists a unique corresponding reverse evidence anchor point A1, with the same order (position 1) and the same associated knowledge source node (K1). This is verified through mirror comparison.

[0146] A2: There exists a unique corresponding reverse evidence anchor point A2, with the same order (2nd position) and the same associated knowledge source node (K3). This is verified through mirror comparison.

[0147] A3: There exists a unique corresponding reverse evidence anchor point A3, with the same order (3rd position) and the same associated knowledge source node (K2). This is verified through mirror comparison.

[0148] A4: There exists a unique corresponding reverse evidence anchor point A4, with the same order (4th position) and the same associated knowledge source node (K2). This is verified through mirror comparison.

[0149] No conflicting points were found in the source tracing.

[0150] The traceability consistency score S is calculated as follows:

[0151] S=α·(N match / N total )+β·(ΣCF i / N total )+γ·C order .

[0152] For this embodiment: N total =4, N match =4.

[0153] ΣCF i =CF(E1)+CF(E2)+CF(E3)+CF(E4)=0.93+0.90+0.92+0.94=3.69.

[0154] C order =1.0 (the order is exactly the same).

[0155] Assume α=0.4, β=0.4, γ=0.2:

[0156] S=0.4×(4 / 4)+0.4×(3.69 / 4)+0.2×1.0=0.4+0.369+0.2=0.969.

[0157] S=0.969≥0.9, indicating a "high credibility" record, signifying that the forward tracing path chain is logically consistent and rigorous, and the two-way verification is highly consistent. The verification report includes: tracing consistency score 0.969, credibility level "high credibility", and number of tracing conflict points 0.

[0158] Step 4: Complete the final storage and index update of the question and answer records.

[0159] "Highly reliable" records are retained in their entirety without any corrections.

[0160] For records approved for storage, the verification report generated in step 3 is embedded into the basic question-and-answer record block to form a complete final question-and-answer record entry with a verification stamp. This verification stamp is a digital credential of the record's credibility, containing a traceability consistency score of 0.969, a credibility level of "high credibility", a verification timestamp of "2024-03-15T09:23:18+08:00", and a key verification summary of "number of traceability conflict points: 0".

[0161] The final question-and-answer record entries with verification stamps are written to persistent storage media such as databases. Simultaneously, an efficient dual-path index is built to support fast retrieval: the first path index is constructed using the core keywords of the question and answer items, supporting content-based semantic retrieval; the second path index is constructed using the source identifiers of all knowledge source nodes referenced in the record, supporting source-based reverse tracing queries. When a user queries a record in the future, they can quickly find it through the first path index; when a knowledge source node is updated or needs to be audited, all question-and-answer records referencing that knowledge source node can be quickly located through the second path index.

[0162] After the records are stored, a global source quality statistics table is updated. This global source quality statistics table records the number of times each knowledge source node is successfully cited, the average source consistency score of all question and answer records in which it participates, and the time of its most recent citation.

[0163] By analyzing this table, we can identify high-quality, highly reliable "authoritative" knowledge source nodes (frequently cited and with high average scores), as well as "weak" knowledge source nodes (with low average scores) that may have problems and frequently lead to source tracing conflicts.

[0164] In this embodiment: K1's successful citation count increases by 1, and the average source consistency score is updated to 0.969; K2's successful citation count increases by 1, and the average score is updated to 0.969; K3's successful citation count increases by 1, and the average score is updated to 0.969. K1, K2, and K3 are all marked as "authoritative" knowledge source nodes.

[0165] Example 2: Medical Knowledge Q&A

[0166] Scenario: In a hospital's clinical decision support system, a doctor queries, "What complication should be considered first when a diabetic patient experiences numbness and tingling in their feet?" The system generates an answer based on a medical knowledge base, performs a fusion-based source verification, and determines the answer to have medium credibility. After an auxiliary correction process, the credibility is increased to high credibility.

[0167] Step 1: Construct the basic blocks for question and answer records.

[0168] A doctor enters a query on the Q&A interface: "What complication should be considered first when a diabetic patient experiences numbness and tingling in their feet?" The system records this query as a question. Based on its medical knowledge base, the system searches relevant literature and textbooks to generate a text response: "When a diabetic patient experiences numbness and tingling in their feet, diabetic peripheral neuropathy (DPN) should be considered first. This is one of the most common chronic complications of diabetes, mainly caused by long-term hyperglycemia leading to nerve fiber metabolic disorders and microvascular complications. It is recommended to perform nerve conduction velocity testing and foot sensory assessment." This response is recorded as an answer. The system generates a unified timestamp for this Q&A interaction: "2024-03-15T14:37:52+08:00" and binds this timestamp to both the question and answer.

[0169] During the generation of answer items, the system simultaneously records all knowledge sources referenced by each answer item. Analysis revealed that the answer item referenced three independent knowledge sources:

[0170] The first source of knowledge is the clinical manifestations of diabetic peripheral neuropathy (DPN) in Chapter X, "Chronic Complications of Diabetes," of the Internal Medicine textbook. The second source is a review article abstract on "Diagnostic Criteria for Diabetic Peripheral Neuropathy" from a medical literature database. The third source is the "Foot Numbness" symptom entity and its associated disease nodes in the hospital knowledge graph.

[0171] The three knowledge sources mentioned above are all defined as independent knowledge source nodes, each assigned a globally unique tracing identifier:

[0172] Knowledge source node K1: The traceability identifier is MED-TXT-2022-015#CH9#PARA3.

[0173] Knowledge source node K2: The traceability identifier is MED-LIT-2023-089#ABS#FULL.

[0174] Knowledge source node K3: The traceability identifier is MED-KG-2023-045#ENT-FNS#ATTR-ALL.

[0175] The system segments the answer text into segments with independent semantics:

[0176] Answer fragment S1: "Diabetic peripheral neuropathy (DPN)";

[0177] Answer fragment S2: "Long-term hyperglycemia leads to nerve fiber metabolic disorders and microvascular lesions";

[0178] Answer fragment S3: "It is recommended to perform nerve conduction velocity testing and foot sensation assessment."

[0179] Create a list of mapping pairs:

[0180] Mapping to item 1: Answer fragment S1 corresponds to knowledge source nodes K1 and K2 (disease name and diagnostic criteria);

[0181] Mapping to entry 2: Answer fragment S2 corresponds to knowledge source node K1 (pathogenesis mechanism);

[0182] Mapping to item 3: Answer fragment S3 corresponds to knowledge source node K2 (check recommendations in diagnostic criteria).

[0183] The system encapsulates the question item, answer item, three knowledge source nodes and their source identifiers, the above mapping pair list, and the timestamp 2024-03-15T14:37:52+08:00 into a basic block of question and answer records, and temporarily stores it in the system cache.

[0184] Step 2: Generate a forward tracing path chain.

[0185] The system analyzes the answer options and identifies four evidence anchors, which are arranged in logical order to form an evidence anchor sequence:

[0186] Anchor of evidence A1: "Numbness and tingling in the feet" (symptom description, introducing the topic);

[0187] Anchor of evidence A2: "Diabetic peripheral neuropathy" (preliminary diagnosis, logical progression);

[0188] Anchor of evidence A3: "Long-term hyperglycemia leads to metabolic disorders of nerve fibers" (etiology and mechanism, supporting the diagnosis);

[0189] Evidence Anchor A4: "Neural Conduction Velocity Test" (Suggested auxiliary examination, conclusion extension).

[0190] Starting with the question item as the logical starting point, construct directed edges according to the sequence of evidence anchor points:

[0191] Directed edge E1: Problem item → Evidence anchor point A1. According to the mapping pair 1, A1 corresponds to knowledge source node K3 (symptom entity).

[0192] Directed edge E2: Evidence anchor A1 → Evidence anchor A2. According to the mapping pair 1, A2 corresponds to knowledge source nodes K1 and K2 (the association between symptoms and disease).

[0193] Directed edge E3: Evidence anchor A2 → Evidence anchor A3. According to the mapping pair 2, A3 corresponds to knowledge source node K1 (pathogenesis mechanism).

[0194] Directed edge E4: Evidence anchor A2 → Evidence anchor A4. According to mapping item 3, A4 corresponds to knowledge source node K2 (examination recommendations in the diagnostic criteria).

[0195] All directed edges are organized into a forward tracing path chain that starts from the problem item, passes through the evidence anchors A1, A2, A3, and A4 in sequence, and finally reaches the knowledge source nodes K1, K2, and K3.

[0196] CF(E i ) represents the path transition confidence of the i-th directed edge, and its value is jointly determined by the semantic matching degree and word frequency of the evidence anchor in the content of the corresponding knowledge source node, as well as the logical importance of the evidence anchor in the answer item.

[0197] CF(E1): The semantic match between the question item "numbness and tingling in the feet" and the symptom "numbness in the feet" in answer fragments S1 and K3 is 0.88, indicating significant word frequency. A1 plays a high logical importance in introducing the topic in the answer item. CF(E1) = 0.88.

[0198] CF(E2): The logical connection between A1 "numbness and tingling in the feet" and A2 "diabetic peripheral neuropathy" is established by the explicit mention in K1 that numbness and tingling are typical symptoms of DPN. However, in K3, "numbness in the feet" is also associated with other diseases such as lumbar spine lesions, resulting in weak ambiguity. Semantic match degree 0.75. CF(E2) = 0.75.

[0199] CF(E3): The logical connection between A2 "diabetic peripheral neuropathy" and A3 "long-term hyperglycemia leading to nerve fiber metabolic disorder" is directly established by the etiological mechanism description in K1, with a semantic matching degree of 0.92. CF(E3) = 0.92.

[0200] CF(E4): The logical connection between A2 "diabetic peripheral neuropathy" and A4 "nerve conduction velocity test" is directly established by the diagnostic criteria in K2, with a semantic match degree of 0.85. CF(E4) = 0.85.

[0201] The system generates a forward tracing path chain object, recording the order of all nodes on the forward tracing path chain: problem item → A1 → A2 → A3 → A4 → K1 / K2 / K3, as well as four directed edges E1 to E4 and their path transition confidence. This forward tracing path chain object also records each directed edge between adjacent nodes on the forward tracing path chain and its corresponding path transition confidence.

[0202] Step 3: Perform two-way verification.

[0203] For knowledge source node K3, the content "numbness in the feet" semantically matches the evidence anchor point A1 "numbness and tingling in the feet," establishing a semantic backtracking link. Deducing backward from A1, the content "numbness and tingling in the feet" directly addresses the fundamental query intent of the question item. The backlink was successfully established. However, during the backtracking process, it was discovered that K3 is also associated with diseases such as "lumbar disc herniation" and "lower extremity arteriosclerosis," indicating that the derivation from A1 to A2 is not a unique path and there are path branches.

[0204] For knowledge source node K1, its content shows that the clinical manifestations of DPN match evidence anchor A2, and the etiology matches evidence anchor A3, thus deducing the fundamental query intent that fully responds to the question item. The backlink is successfully established.

[0205] For knowledge source node K2, its diagnostic criteria and examination recommendations match the evidence anchor A4. However, in reverse derivation, K2 also mentions the need to exclude differential diagnoses such as "vascular lesions," indicating that the derivation from A2 to A4 is not unique. A reverse link is established, but the confidence level is limited.

[0206] Record the reverse node identifier sequence R Seq =[K3,A1,K1,A2,K2,A4] (Evidence anchor A3 was deemed weakly correlated in the reverse derivation because K1 focused more on clinical manifestations than etiological descriptions, and was therefore not included in R.) Seq ).

[0207] Obtain the forward node identifier sequence F Seq and the reverse node identifier sequence R Seq Then, the two are compared. The focus of the comparison is the mirror symmetry of the evidence anchor sequence.

[0208] From F Seq Extract all evidence anchor points and form a positive evidence anchor point sequence F. AnchorSeq =[A1,A2,A3,A4]. Accordingly, from RSeq Extract all the evidence anchors located in the answer items through reverse derivation, and form a reverse evidence anchor sequence R. AnchorSeq =[A1,A2,A4].

[0209] Subsequently, the two evidence anchor sequence sequences were aligned and compared item by item.

[0210] Item-by-item comparison results:

[0211] A1: There exists a unique corresponding reverse evidence anchor point A1, with the same order (position 1) and the same associated knowledge source node (K3). This is verified through mirror comparison.

[0212] A2: There exists a unique corresponding reverse evidence anchor point A2, with the same order (2nd position) and the same associated knowledge source node (K1). This is verified through mirror comparison.

[0213] A3: In R AnchorSeq Missing information. Recorded as source conflict point C1: "Evidence anchor point A3 (pathogenesis mechanism) could not be effectively backtracked and verified in reverse derivation. Reason: The etiology description and clinical manifestations in knowledge source node K1 are intertwined, making it difficult to extract independently during reverse localization."

[0214] A4: There exists a unique corresponding reverse evidence anchor point A4, with the same order (3rd position) and the same associated knowledge source node (K2). This is verified through mirror comparison.

[0215] Consistency score for traceability.

[0216] For this embodiment: Ntotal =4, N match =3. ΣCF i = =0.88+0.75+0.85=2.48 (Only the paths corresponding to successfully matched anchor points are calculated).

[0217] Co rder =0.85 (The absence of A3 impairs the order consistency, but the rest are in the same order). Let α=0.4, β=0.4, γ=0.2: S=0.4×(3 / 4)+0.4×(2.48 / 4)+0.2×0.85=0.3+0.248+0.17=0.718.

[0218] S=0.718, 0.7≤S<0.9, classifying it as a "medium confidence" record, indicating that the forward tracing path chain is basically valid, but some links have weak correlations or non-critical inconsistencies. The verification report includes: tracing consistency score of 0.718, confidence level "medium confidence", and details of tracing conflict point C1.

[0219] Step 4: Complete the final storage and index update of the question and answer records.

[0220] For records of "medium credibility", the system can trigger an auxiliary correction process: try to find semantically more semantically matching alternative knowledge source nodes in the knowledge base for the weakly related source conflict points found in step 3-2.

[0221] The system searched the medical literature database and found a literature K4 (source identifier: MED-LIT-2023-156#SEC2#PARA1) that specifically discusses the pathogenesis of diabetic peripheral neuropathy. In it, there is an independent and complete paragraph describing "long-term hyperglycemia leads to nerve fiber metabolic disorder and microvascular lesions". The semantic match with evidence anchor A3 is significantly higher than the intertwined description in K1.

[0222] Replace the original source of the associated source of A3 in the original mapping pair (originally only K1, now K4 is added as the main support) with the alternative knowledge source node K4, and update the mapping relationship and the forward tracing path chain.

[0223] Recalculation: The confidence of E3 pathway transfer is updated to CF(E3)'=0.91 (the etiological description in K4 is independent and complete).

[0224] Re-execute the reverse traversal: the pathogenesis mechanism in K4 matches exactly with A3, the reverse derivation successfully establishes a semantic backtracking link, and the evidence anchor A3 is incorporated into R. Seq .

[0225] Recalculate: N match =4, ΣCF i =0.88+0.75+0.91+0.85=3.39, C order =1.0.

[0226] S'=0.4×1+0.4×(3.39 / 4)+0.2×1=0.4+0.339+0.2=0.939.

[0227] S'=0.939≥0.9, the record level is upgraded to "high credibility".

[0228] For records approved for storage (corrected records), the verification report generated in step 3 is embedded into the basic question-and-answer record block to form a complete final question-and-answer record entry with a verification stamp. This verification stamp is a digital credential of the record's credibility, containing a traceability consistency score of 0.939, a credibility level of "High Credibility (Corrected)", a verification timestamp, and a key verification summary: "One original traceability conflict point, corrected, replacing the knowledge source node: K4".

[0229] The final question-and-answer record entries with verification stamps are written to persistent storage media such as databases. Simultaneously, an efficient dual-path index is built to support fast retrieval: the first path index is constructed using the core keywords of the question and answer entries, supporting content-based semantic retrieval; the second path index is constructed using source identifiers that record all referenced knowledge source nodes, supporting source-based reverse tracing queries.

[0230] After the records are stored, a global source quality statistics table is updated. This global source quality statistics table records the number of times each knowledge source node is successfully cited, the average source consistency score of all question and answer records in which it participates, and the time of its most recent citation.

[0231] In this embodiment: K1's average score is updated (citation count increases, but it was recorded before the correction); the number of successful citations for the alternative knowledge source node K4 is recorded as 1, with an average source consistency score of 0.939, and it is marked as an "authoritative" knowledge source node. K3's average score is updated because its average score decreases slightly due to a path fork in the reverse derivation, and the system marks it as a knowledge source node that needs attention.

[0232] Example 3: Legal Consultation Q&A

[0233] Scenario: On an online legal service platform, a user queries, "Is it legal for a landlord to refuse to return a security deposit after the rental contract expires, citing the need for renovations?" The system generates an answer based on its legal knowledge base, but after performing a fusion-based verification, it finds the answer to be of low credibility and moves it to the pending review area.

[0234] Step 1: Construct the basic blocks for question and answer records.

[0235] A user entered the query "Is it legal for a landlord to refuse to return a security deposit after the lease expires, citing the need for renovations?" on the Q&A interface. The system recorded this query as a question. Based on its legal knowledge base, the system retrieved relevant regulations and generated a text response: "According to Article 703 of the Civil Code and Article 20 of the Measures for the Administration of Commodity Housing Leasing, if the lessee returns the leased property after the lease expires and there is no breach of contract, the lessor shall return the security deposit. The landlord's refusal to return the security deposit on the grounds of needing renovations lacks legal basis, and the lessee has the right to demand its return. If the landlord refuses to return it, the lessee may file a lawsuit in the People's Court or file a complaint with the housing and urban-rural development authority." This response was recorded as an answer. The system generated a unified timestamp for this Q&A interaction: 2024-03-15T16:45:33+08:00, and bound this timestamp to both the question and answer.

[0236] During the generation of answer items, the system simultaneously records all knowledge sources referenced by each answer item. Analysis revealed that the answer item referenced three independent knowledge sources:

[0237] The first source of knowledge is Article 703 of the Civil Code, which defines a lease contract and outlines its basic obligations. The second source is Article 20 of the Measures for the Administration of Commodity Housing Leasing, which stipulates the return of security deposits. The third source is a highly-rated answer on a legal Q&A community regarding "landlords refusing to return security deposits" (not an official legal source).

[0238] The three knowledge sources mentioned above are all defined as independent knowledge source nodes, each assigned a globally unique tracing identifier:

[0239] Knowledge source node K1: The traceability identifier is LAW-COD-2020-001#ART703#FULL.

[0240] Knowledge source node K2: The traceability identifier is LAW-REG-2010-006#ART20#FULL.

[0241] Knowledge source node K3: The traceability identifier is LAW-COM-2023-112#ANS#FULL.

[0242] Steps 1-3: Establish preliminary mapping relationships.

[0243] The system segments the answer text into segments with independent semantics:

[0244] Answer fragment S1: "According to Article 703 of the Civil Code and Article 20 of the Measures for the Administration of Commodity Housing Leasing";

[0245] Answer fragment S2: "If the lessee returns the leased property after the lease contract expires and there is no breach of contract, the lessor shall return the deposit."

[0246] Answer snippet S3: "The landlord's refusal to return the deposit on the grounds that the house needs renovation lacks legal basis."

[0247] Answer fragment S4: "You can file a lawsuit with the People's Court or file a complaint with the housing and urban-rural development authorities."

[0248] Create a list of mapping pairs:

[0249] Mapping to item 1: Answer fragment S1 corresponds to knowledge source nodes K1 and K2 (legal basis reference);

[0250] Mapping to item 2: Answer fragment S2 corresponds to knowledge source node K2 (deposit refund conditions);

[0251] Mapping to item 3: Answer fragment S3 corresponds to knowledge source node K3 (legal viewpoints in community answers);

[0252] Mapping to item 4: Answer fragment S4 corresponds to knowledge source node K3 (rights protection suggestions in community answers).

[0253] The system encapsulates the question item, answer item, three knowledge source nodes and their source identifiers, the above mapping pair list, and the timestamp 2024-03-15T16:45:33+08:00 into a basic block of question and answer records, and temporarily stores it in the system cache.

[0254] Step 2: Generate a forward tracing path chain.

[0255] The system analyzes the answer options and identifies five evidence anchors, which are arranged in logical order to form an evidence anchor sequence:

[0256] Anchor of evidence A1: "Expiration of the lease contract" (precondition, introducing the topic);

[0257] Anchor of evidence A2: "Return of leased property and no breach of contract" (condition confirmed, logical progression);

[0258] Anchor of evidence A3: "The deposit should be returned" (legal obligation, preliminary conclusion);

[0259] Anchor of evidence A4: "Lack of legal basis" (negative judgment, logical shift);

[0260] Evidence Anchor A5: "Litigation or Complaint" (Rights Protection Channels, Conclusion Extension).

[0261] Starting with the question item as the logical starting point, construct directed edges according to the sequence of evidence anchor points:

[0262] Directed edge E1: Problem item → Evidence anchor point A1. According to the mapping pair 1, A1 corresponds to the knowledge source node K1 (basic definition of lease contract).

[0263] Directed edge E2: Evidence anchor A1 → Evidence anchor A2. According to mapping entry 2, A2 corresponds to knowledge source node K2 (deposit refund condition).

[0264] Directed edge E3: Evidence anchor A2 → Evidence anchor A3. According to mapping entry 2, A3 corresponds to knowledge source node K2 (deposit return obligation).

[0265] Directed edge E4: Evidence anchor point A3 → Evidence anchor point A4. According to the mapping pair 3, A4 corresponds to the knowledge source node K3 (legal opinions in the community answers).

[0266] Directed edge E5: Evidence anchor point A4 → Evidence anchor point A5. According to the mapping pair 4, A5 corresponds to the knowledge source node K3 (rights protection suggestions in the community answer).

[0267] All directed edges are organized into a forward tracing path chain that starts from the problem item, passes through the evidence anchors A1, A2, A3, A4, and A5 in sequence, and finally reaches the knowledge source nodes K1, K2, and K3.

[0268] CF(E1): The semantic match between "rental contract expiration" in the question item and the general definition of a rental contract in Article 703 of K1 is 0.85. However, Article 703 is a general definition and does not directly involve the deposit. CF(E1) = 0.85.

[0269] CF(E2): The logical connection between A1 "lease contract expires" and A2 "return of leased property without breach of contract" is clearly defined by Article 20 of K2, with a semantic matching degree of 0.90. CF(E2) = 0.90.

[0270] CF(E3): The logical connection between A2 "return the leased property without breach of contract" and A3 "the deposit should be returned" is clearly defined by Article 20 of K2, with a semantic matching degree of 0.92. CF(E3) = 0.92.

[0271] CF(E4): The logical connection between A3 "the deposit should be returned" and A4 "lack of legal basis" was established by the K3 community answer. However, K3 is a community answer, not an official regulation, and "lack of legal basis" is an opinion statement, not a citation of the law. Semantic match degree 0.55. CF(E4)=0.55.

[0272] CF(E5): The logical connection between A4 "Lack of legal basis" and A5 "Litigation or complaint" was established by the K3 community answer, but the community answer lacks authority, and the suggestive statement deviates from the fundamental query intent of "Is it legal?". Semantic match degree 0.50. CF(E5)=0.50.

[0273] The system generates a forward tracing path chain object, recording the order of all nodes on the forward tracing path chain: problem item → A1 → A2 → A3 → A4 → A5 → K1 / K2 / K3, as well as five directed edges E1 to E5 and their path transition confidence. This forward tracing path chain object also records each directed edge between adjacent nodes on the forward tracing path chain and its corresponding path transition confidence.

[0274] Step 3: Perform two-way verification.

[0275] Perform a forward traversal: Starting from the first evidence anchor in the answer, follow the direction of the edges in the forward tracing path chain generated in step 2, and visit each evidence anchor and knowledge source node on the path in turn, recording the sequence of node identifiers traversed, denoted as F. Seq =[A1,K1,A2,K2,A3,K2,A4,K3,A5,K3].

[0276] Perform a reverse traversal: start from each knowledge source node K recorded in the forward node identifier sequence. i Starting from this point, try to follow the reverse direction of logical deduction to find an anchor point A in the corresponding answer option that can support its connection. j And ultimately, it can respond to the logical chain of the question item.

[0277] Specifically, for knowledge source node K1, Article 703 is a general definition of a lease contract. However, during reverse derivation, it cannot directly pinpoint the specific situation of "lease contract expiration" at evidence anchor point A1; it can only support the concept of "lease contract." Semantic backtracking links are established, but the matching granularity is coarse.

[0278] For knowledge source node K2, Article 20 clearly stipulates the return of the deposit upon expiration, which precisely matches evidence anchors A2 and A3, allowing for reverse deduction of the fundamental query intent that fully responds to the question item. The reverse link has been successfully established.

[0279] For knowledge source node K3, the community answers mixed legal opinions, case citations, and rights protection suggestions. In reverse deduction: evidence anchor point A4 "lack of legal basis" is an opinion-based conclusion in K3, without specific legal provisions to support it; evidence anchor point A5 "litigation or complaint" is a suggestive statement in K3, which deviates from the fundamental query intent of the question item "whether it is legal" (the user asked about legality, while K3 focuses on rights protection actions). A reverse link is established, but the direction of logical deduction is inconsistent with the fundamental query intent of the question item.

[0280] Record the reverse node identifier sequence R Seq =[K1,A1,K2,A2,K3,A5] (Evidence anchor A3 is merged in the reverse node identifier sequence because it belongs to K2 support along with A2; evidence anchor A4 could not be located independently in the reverse derivation because "lack of legal basis" in K3 is a general view and cannot correspond precisely to A4).

[0281] Step 3-2: Perform a consistency comparison of the evidence anchor sequence.

[0282] Obtain the forward node identifier sequence F Seq and the reverse node identifier sequence R Seq Then, the two are compared. The focus of the comparison is the mirror symmetry of the evidence anchor sequence.

[0283] From F Seq Extract all evidence anchor points and form a positive evidence anchor point sequence F. AnchorSeq =[A1,A2,A3,A4,A5]. Correspondingly, from R... Seq Extract all the evidence anchors located in the answer items through reverse derivation, and form a reverse evidence anchor sequence R. AnchorSeq=[A1,A2,A5].

[0284] Subsequently, the two evidence anchor sequence sequences were aligned and compared item by item. Item-by-item comparison results:

[0285] A1: There exists a unique corresponding reverse evidence anchor point A1, with the same order (position 1) and the same associated knowledge source node (K1). This is verified through mirror comparison.

[0286] A2: There exists a unique corresponding reverse evidence anchor point A2, with the same order (2nd position) and the same associated knowledge source node (K2). This is verified through mirror comparison.

[0287] A3: In R AnchorSeq The record is missing. It is recorded as the source conflict point C1: "The evidence anchor point A3 (the deposit should be returned) does not appear independently in the reverse node identifier sequence and is merged by A2, resulting in the destruction of mirror symmetry."

[0288] A4: In R AnchorSeq The following is missing. The record is for the source conflict point C2: "Evidence anchor point A4 (lack of legal basis) could not be effectively backtracked and verified in the reverse deduction. The reason is that the knowledge source node K3 is a community answer, in which 'lack of legal basis' is an opinion statement without textual support, and it is impossible to establish a precise semantic backtracking link with A4."

[0289] A5: There is a corresponding pair, but the order is incorrect (R) AnchorSeq A5 is in the 3rd position, F AnchorSeq The record is located at position 5, and although the associated knowledge source node is also K3, the reverse derivation direction deviates from the fundamental query intent of the question item "whether it is legal". The record is the source conflict point C3: "The order of evidence anchor point A5 (litigation or complaint) is inconsistent, and the reverse derivation direction deviates from the fundamental query intent of the question item 'whether it is legal'".

[0290] Calculate the traceability consistency score. For this embodiment: N total =5, N match =2 (Only A1 and A2 successfully matched). ΣCF i =CF(E1)+CF(E2)=0.85+0.90=1.75.

[0291] C order =0.40 (A3 and A4 are missing, A5 is out of order and has a biased intent, and the consistency of the order is severely compromised).

[0292] Assume α=0.4, β=0.4, γ=0.2:

[0293] S=0.4×(2 / 5)+0.4×(1.75 / 5)+0.2×0.40=0.16+0.14+0.08=0.38.

[0294] S=0.38<0.7, indicating a "low credibility" record, suggesting significant flaws or contradictions in the tracing logic. The verification report includes: a tracing consistency score of 0.38, a credibility level of "low credibility", details of tracing conflict points C1, C2, and C3, and a description of the conflict type.

[0295] Step 4: Complete the final storage and index update of the question and answer records.

[0296] Records with "low credibility" will not trigger the auxiliary correction process (the score is too low, and the correction cost is high and the risk is great). They will be directly transferred to the pending review area and the administrator will be notified to conduct manual review.

[0297] After review, the administrator found that evidence anchors A4 and A5 in the answer section overly relied on community answer K3, indicating a reliance on non-authoritative sources. The correct approach should be: evidence anchor A4 should cite relevant clauses in the Civil Code regarding "no property may be detained without legal basis" (such as relevant provisions in the Property Law section); evidence anchor A5 should cite formal regulations such as the Civil Procedure Law and the Ministry of Housing and Urban-Rural Development's Complaint Handling Measures. Because the knowledge base lacks complete inclusion of these clauses, the system was forced to rely on community answers, resulting in a source tracing conflict.

[0298] This record does not generate a final question-and-answer record entry with a verification stamp and is not included in the formal knowledge record repository.

[0299] This record is not written to the formal database and no dual-path index is created.

[0300] This record does not update the authoritative node statistics in the global traceability quality statistics table. However, the system records the event in the anomaly log database, marks K3 as a "weak" knowledge source node (with an extremely low average score), and triggers the knowledge base supplementation and improvement process, prompting the administrator to supplement the relevant clauses of the Property Law section of the Civil Code and the complaint handling regulations.

[0301] The above three embodiments demonstrate the workflow of the present invention in different application scenarios, covering three typical cases: high credibility, medium credibility improved by correction, and low credibility rejected.

[0302] like Figure 3 As shown, in another embodiment of the present invention, a knowledge question-and-answer record system integrating traceability and verification is provided, comprising:

[0303] The question and answer record basic block construction unit is used to collect the original information of questions and answers and bind timestamps, extract all knowledge source nodes referenced by the answer item and assign them a globally unique traceability identifier, divide the answer item into independent semantic fragments and establish a mapping pair list with the corresponding knowledge source nodes, encapsulate and generate the question and answer record basic block and output it to the forward traceability path chain generation unit;

[0304] The forward tracing path chain generation unit is used to receive the question and answer record basic block, identify the evidence anchor points that play a key logical supporting role in the answer items and determine the evidence anchor point sequence according to the semantic structure, connect each evidence anchor point with the corresponding knowledge source node in sequence with the question item as the logical starting point to form a directed path, assign a quantified path transfer confidence to each directed edge, generate a forward tracing path chain object and output it to the bidirectional verification unit.

[0305] The bidirectional verification unit is used to receive the forward tracing path chain object, perform forward traversal and reverse traversal respectively to generate forward node identifier sequence and reverse node identifier sequence, extract the evidence anchor point sequence in the two sequences, perform mirror symmetry comparison and record the tracing conflict points, calculate the tracing consistency score based on the matching result, path transition confidence and sequence consistency, determine the credibility level of the record according to the score and generate a verification report, and output the question and answer record basic block, tracing consistency score, credibility level and verification report to the storage and indexing unit.

[0306] The storage and indexing unit is used to receive the question and answer record base block, traceability consistency score, trust level and verification report, perform differentiated processing on the question and answer records according to the trust level, generate the final question and answer record entries with verification stamps, store and establish a dual-path index based on content and source, and update the global traceability quality statistics table.

[0307] Figure 4 This is a backend interface of the knowledge question-and-answer record system that integrates traceability and verification according to the present invention, which displays the unified management and quality monitoring results of the knowledge source nodes involved in the above three embodiments.

[0308] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0309] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A knowledge question-and-answer recording method integrating traceability and verification, characterized in that, Includes the following steps: Step 1: Construct the basic block of question and answer records, collect the original information of questions and answers and bind timestamps, extract all knowledge source nodes referenced by the answer item and assign them a globally unique traceability identifier, divide the answer item into independent semantic fragments and establish a mapping pair list with the corresponding knowledge source nodes, and encapsulate them into structured data objects; Step 2: Generate a forward tracing path chain. Identify the evidence anchors that play a key logical supporting role in the answer items and determine the sequence of evidence anchors according to semantic structure. Connect each evidence anchor with the corresponding knowledge source node in sequence, starting from the question item, to form a directed path. Assign a quantified path transfer confidence to each directed edge. Step 3: Perform bidirectional verification. Perform forward and reverse traversal respectively to generate forward node identifier sequences and reverse node identifier sequences. Extract the evidence anchor point sequences from the two sequences, perform mirror symmetry comparison and record the source conflict points. Calculate the source consistency score based on the matching results, path transition confidence and sequence consistency. Determine the credibility level of the record based on the score and generate a verification report. Step 4: Complete the final storage and index update of the question and answer records. Differentiate the question and answer records according to their trust level, generate final question and answer record entries with verification stamps, write them to storage, and establish a dual-path index based on content and source. Update the global traceability quality statistics table.

2. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 1, characterized in that, The steps in step 1 of constructing the basic block of question-and-answer records include: Step 1-1: Collect the original question and answer information, obtain the query statement submitted by the user as the question item, record the generated text reply as the answer item, and generate a unified timestamp to bind to the question item and the answer item; Step 1-2: Extract knowledge source nodes. During the process of generating answer items, record all referenced knowledge source nodes simultaneously and assign a globally unique source identifier to each knowledge source node. Steps 1-3: Establish preliminary mapping relationships, segment the answer item text into independent semantic fragments, associate each answer fragment with the corresponding knowledge source node, and generate a list of mapping pairs; Steps 1-4: Encapsulate the basic block of the question and answer record, encapsulating the question item, answer item, knowledge source node, mapping pair list and timestamp into a structured data object and temporarily storing it in the cache area.

3. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 2, characterized in that, Step 2, which generates the forward tracing path chain, includes: Step 2-1: Define the evidence anchor sequence, identify the evidence anchors that play a key logical supporting role in the answer items, and determine the evidence anchor sequence according to the semantic structure of the answer items; Step 2-2: Construct a forward tracing path chain. Starting with the question item as the logical starting point, connect the corresponding knowledge source nodes in sequence according to the order of the evidence anchor sequence to form a directed forward tracing path chain. Steps 2-3: Calculate the path transition confidence and assign a quantified path transition confidence to each directed edge in the forward tracing path chain; Steps 2-4: Solidify the forward tracing path chain, and integrate the forward tracing path chain sequence with the confidence of each path transition into a structured forward tracing path chain object.

4. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 3, characterized in that, Step 3, which involves performing bidirectional verification, includes: Step 3-1: Perform bidirectional path traversal and sequence generation, performing forward and reverse traversal respectively to generate forward node identifier sequences and reverse node identifier sequences; Step 3-2: Perform consistency comparison of evidence anchor sequence, extract evidence anchor sequence from the forward node identifier sequence and the reverse node identifier sequence, align and compare item by item and record the source conflict points; Step 3-3: Calculate the source consistency score, based on the number of successfully matched evidence anchors, path transition confidence, and sequential consistency to calculate the comprehensive score; Steps 3-4: Determine the credibility level of the records and generate a verification report. Based on the traceability consistency score, divide the question and answer records into different credibility levels and generate a verification report containing details of the conflict points.

5. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 4, characterized in that, Step 4, which completes the final storage and index update of the question-and-answer records, includes: Step 4-1: Filter and correct records based on their trust level, and adopt different processing strategies for question and answer records with different trust levels; Step 4-2: Generate the final question and answer record entry with a verification stamp, embed the verification report in the basic question and answer record block to form the final question and answer record entry with a verification stamp; Step 4-3: Store and build a dual-path index, write the final question and answer record entries to the persistent storage medium, and build a dual-path index based on content and source; Step 4-4: Update the global source traceability quality statistics table, and record the number of citations, average source traceability consistency score, and most recent citation time for each knowledge source node.

6. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 5, characterized in that, In steps 1-2, the source identifier includes at least the library number of the source knowledge base, the document number of the specific document, and the location information of the knowledge source node in the document.

7. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 6, characterized in that, In steps 2-3, the confidence level of the path transfer from the evidence anchor to the knowledge source node is jointly determined by the semantic matching degree and word frequency of the evidence anchor in the content of the corresponding knowledge source node, as well as the logical importance of the evidence anchor in the answer item.

8. The knowledge question-and-answer recording method for fusion traceability and verification according to claim 7, characterized in that, In step 3-2, the content of the evidence anchor sequence consistency comparison includes: checking whether there is a unique and semantically corresponding reverse evidence anchor in the reverse evidence anchor sequence for each positive evidence anchor; checking whether the order of appearance of the corresponding evidence anchor in their respective sequences is consistent; and checking whether the knowledge source nodes associated with the corresponding evidence anchor are consistent.

9. The knowledge question-and-answer recording method for integrated traceability verification according to claim 8, characterized in that, In step 4-1, high-credibility records are fully retained, medium-credibility records trigger an auxiliary correction process, and low-credibility records are transferred to the pending review area and the administrator is notified for manual review.

10. A knowledge question-and-answer recording system integrating traceability and verification, characterized in that, include: The question and answer record basic block construction unit is used to collect the original information of questions and answers and bind timestamps, extract all knowledge source nodes referenced by the answer item and assign them a globally unique traceability identifier, divide the answer item into independent semantic fragments and establish a mapping pair list with the corresponding knowledge source nodes, encapsulate and generate the question and answer record basic block and output it to the forward traceability path chain generation unit; The forward tracing path chain generation unit is used to receive the question and answer record basic block, identify the evidence anchor points that play a key logical supporting role in the answer items and determine the evidence anchor point sequence according to the semantic structure, connect each evidence anchor point with the corresponding knowledge source node in sequence with the question item as the logical starting point to form a directed path, assign a quantified path transfer confidence to each directed edge, generate a forward tracing path chain object and output it to the bidirectional verification unit. The bidirectional verification unit is used to receive the forward tracing path chain object, perform forward traversal and reverse traversal respectively to generate forward node identifier sequence and reverse node identifier sequence, extract the evidence anchor point sequence in the two sequences, perform mirror symmetry comparison and record the tracing conflict points, calculate the tracing consistency score based on the matching result, path transition confidence and sequence consistency, determine the credibility level of the record according to the score and generate a verification report, and output the question and answer record basic block, tracing consistency score, credibility level and verification report to the storage and indexing unit. The storage and indexing unit is used to receive the question and answer record base block, traceability consistency score, trust level and verification report, perform differentiated processing on the question and answer records according to the trust level, generate the final question and answer record entries with verification stamps, store and establish a dual-path index based on content and source, and update the global traceability quality statistics table.

Images