Trusted knowledge question answering method based on declaration-evidence bipartite graph constraint propagation

Abstract

The application discloses a kind of trusted knowledge question and answer method based on declaration-evidence bipartite graph constraint propagation. Including: obtaining user query and access permission, retrieving in document set to obtain candidate evidence unit set by filtering, and generating candidate answer draft;The draft is semantically segmented to obtain a set of atomic declaration nodes;According to the association between declaration node and evidence unit, construct declaration-evidence bipartite graph structure;Based on the pre-configured coverage, conflict, timeliness and permission constraints, perform iterative constraint propagation on the bipartite graph structure, obtain the convergent declaration trusted verification state;Global consistency optimization is solved, to determine the final output state and target evidence reference unit;According to this, rewrite draft and output trusted question and answer result.The application can realize global consistency reasoning under the conflict of multiple sources of evidence, can reduce false reference, and improve the reliability of question and answer result in high compliance scenario.

Description

Technical Field

[0001] This invention relates to artificial intelligence and natural language processing technologies, and in particular to a reliable knowledge question answering method based on declaration-evidence bipartite graph constraint propagation. Background Technology

[0002] By introducing external knowledge bases as context, knowledge-based question-answering systems can effectively reduce the illusion of fact when large language models generate content. In scenarios with high requirements for factual consistency, such as corporate norms, legal contracts, and policy interpretations, ensuring that the output conclusions have reliable retrieval basis and can be rigorously logically deduced is the core technological foundation for ensuring the usability and reliability of knowledge-based question-answering systems.

[0003] Existing knowledge-answering systems typically employ a pipelined retrieval-enhanced generation architecture. Specifically, the system first retrieves relevant knowledge fragments based on the user's query, directly generates the response text using a language model, and then performs independent evidence verification by calling a natural language inference model sentence by sentence in a later stage. In environments where the knowledge base is continuously updated or contains multi-source heterogeneous documents, a coarse-grained whole-document replacement strategy is usually relied upon for historical version control, and when verifying long texts, a single-evidence-supported-single-conclusion extraction logic is used by default. Under this architecture, when facing complex fact-checking with multi-dimensional constraints, the system is prone to logical breaks caused by fragmented information; furthermore, it lacks a global, exclusive judgment mechanism when dealing with cross-document version iterations and conflicting knowledge sources.

[0004] In summary, existing knowledge verification mechanisms struggle to simultaneously ensure accurate identification of fine-grained knowledge evolution and global logical consistency in the face of multi-source conflicts when dealing with massive amounts of long texts and frequently revised rules. Therefore, there is an urgent need to research a data processing method that can improve the robustness of fact verification and the consistency of reasoning in complex, heterogeneous knowledge base scenarios. Summary of the Invention

[0005] In view of this, this application provides a reliable knowledge question answering method based on declaration-evidence bipartite graph constraint propagation, in order to solve the above-mentioned problems of the prior art.

[0006] A reliable knowledge question answering method based on declaration-evidence bipartite graph constraint propagation includes:

[0007] Retrieve user queries, corresponding access permissions, and a pre-stored collection of documents;

[0008] Based on user queries and access permissions, the document set is retrieved and filtered to obtain a set of candidate evidence units.

[0009] Text is generated based on user queries and a set of candidate evidence units to obtain draft candidate answers.

[0010] Semantic segmentation of the candidate answer drafts yields a set of atomic declaration nodes;

[0011] Based on the relationship between the set of atomic declaration nodes and the set of candidate evidence units, a declaration-evidence bipartite graph structure is constructed;

[0012] Iterative constraint propagation is performed on the declaration-evidence bipartite graph structure based on pre-configured verification constraint rules to obtain a converged declaration-credible verification state. The verification constraint rules include coverage, conflict, timeliness, and permission constraints.

[0013] Based on the declaration-evidence bipartite graph structure, a global consistency optimization solution is performed on the declaration credibility verification state to obtain the final output state and the target evidence reference unit. Based on this (final output state and target evidence reference unit), the candidate answer draft is rewritten to output the credible question answer result.

[0014] Based on the above technical solutions, the present invention can achieve globally consistent reasoning under multi-source evidence conflict, reduce erroneous citations, and improve the reliability of question-and-answer results in high compliance scenarios. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 This is a flowchart of a trusted knowledge question-answering method based on declaration-evidence bipartite graph constraint propagation provided in this application.

[0017] Figure 2 This application provides a flowchart for retrieving and filtering a document set based on user queries and access permissions to obtain a set of candidate evidence units.

[0018] Figure 3 This application provides a flowchart for retrieving and filtering permissions for a document collection using a multi-level dynamic gating mechanism.

[0019] Figure 4 This application provides a flowchart for constructing a declaration-evidence bipartite graph structure based on the relationship between the set of atomic declaration nodes and the set of candidate evidence units.

[0020] Figure 5 This is the offline security audit flowchart provided in this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] Example 1

[0023] This embodiment provides a reliable knowledge question-answering method based on declaration-evidence bipartite graph constraint propagation, referring to... Figure 1 The method may include the following steps:

[0024] Step 101: Obtain the user query, the corresponding access permissions, and the pre-stored document set.

[0025] The system's front-end interface receives user queries input in natural language and simultaneously obtains the permission identifier of the query initiator to determine access permissions. The underlying knowledge base consists of a pre-stored collection of documents. In practice, this collection typically includes internal management regulations, human resource guidelines, business contracts, bidding documents, system operating procedures, and anonymized customer files, primarily in unstructured and semi-structured PDF and Word document formats.

[0026] During the extraction of underlying documents and the construction of candidate data, the pre-stored document set is sliced. To ensure the granularity of retrieval and the integrity of subsequent graph reasoning, the documents are divided into several candidate evidence units of fixed dimensions. For example, text paragraphs are forcibly segmented into sub-blocks containing 120 to 320 Chinese characters, while retaining metadata such as the original document's title hierarchy, physical page numbers, chapter numbers, clause numbers, effective timestamps, expiration timestamps, and version iteration markers. Furthermore, when the document set contains structured tables, a linear concatenation logic of table name-row header-column header-cell value can be used to flatten them into a one-dimensional sequence that can be processed by the text encoder, preventing the loss of the spatial structure of the table information.

[0027] To provide tamper-proof capabilities in subsequent auditing processes, a fixed-length unique hash value is generated based on the content of the sliced ​​document blocks and key metadata combinations, such as document identifiers and concatenated strings of page number character ranges, using a cryptographic hash algorithm (e.g., SHA-256, a secure hash algorithm). This hash value serves as the basic anti-counterfeiting signature for the candidate evidence unit.

[0028] Step 102: Based on user queries and access permissions, the document set is retrieved and filtered to obtain a set of candidate evidence units.

[0029] After acquiring the underlying knowledge base data, the document set undergoes high-dimensional feature reduction and vectorization calculations. User queries are compared with the segmented document data to extract candidate contexts highly relevant to the query intent. Simultaneously, leveraging acquired access permissions, a strict initial admission filter is applied to the document set, eliminating document units not within the current query subject's permission list, thus contributing to horizontal security isolation of sensitive knowledge. After the retrieval algorithm performs multi-dimensional scoring and rearrangement of document features, a subset of documents with high confidence levels is extracted as a candidate evidence unit set to reduce the redundant load of subsequent complex graph calculations.

[0030] Step 103: Based on the user query and the set of candidate evidence units, text is generated to obtain a draft of the candidate answer.

[0031] After extracting relevant background evidence, the user query and the selected set of candidate evidence units are concatenated and used as a contextual prompt input into the generative model, such as a pre-trained large language model. The generative model is based on the pre-trained large language model and, after multi-task fine-tuning using the joint loss function as described in Example 8, is deployed to the production environment. The generative model progressively generates consecutive text tokens through the decoding layer, outputting draft candidate answers.

[0032] It should be noted that the draft text generated at this stage represents only a draft of candidate answers derived from combining the large model parameter space with extracted evidence. Its content may contain multiple factual complex sentences or create misleading descriptions. Therefore, this draft text is an intermediate variable and is not directly output to the user. Instead, it serves as the input basis for the next step of fine-grained decomposition and cross-validation.

[0033] Step 104: Semantically segment the candidate answer drafts to obtain a set of atomic declaration nodes.

[0034] Forced word segmentation and dependency parsing are performed on the generated proposed long text to break down the ambiguity of complex sentence structures. By setting punctuation marks such as periods, semicolons, and parallel conditional conjunctions as segmentation boundaries, and combining a boundary recognition model (segmentation model) to break down deep semantics, the candidate answer draft is decomposed into multiple indivisible atomic statements. Each segmented atomic statement represents a single, independent, and verifiable factual logical loop, for example, containing only a single subject, predicate, and numerical time constraints. The smallest factual units after normalization and dimensionality reduction constitute the set of atomic statement nodes.

[0035] Step 105: Construct a declaration-evidence bipartite graph structure based on the association between the set of atomic declaration nodes and the set of candidate evidence units.

[0036] A logical verification network is established in the topological space. Each fact unit in the atomic declaration node set is used as the left node set of the bipartite graph, and the underlying document fragments in the candidate evidence unit set are used as the right node set. For any node on the left and any node on the right, a deep inference algorithm is used to determine whether there is substantial semantic implication, contradiction, mutual exclusion, or information irrelevance between them.

[0037] In the specific construction process, the data representation of the declaration-evidence bipartite graph structure is B=(C,E,L,W,M). Here, B is the declaration-evidence bipartite graph, C is the set of nodes representing atomic declarations, E is the set of nodes representing candidate evidence units, L is the set of connected edges established between nodes, W is the weight matrix identifying the strength of each edge, and M is the set of attribute metadata bound to the evidence node E. The construction of this graph structure transforms the discrete sentence verification task into a global topological network state solving task.

[0038] Step 106: Perform iterative constraint propagation on the declaration-evidence bipartite graph structure based on the pre-configured verification constraint rules to obtain a converged declaration trustworthy verification state. The verification constraint rules include coverage, conflict, timeliness, and permission constraints.

[0039] After the graph structure is established, pre-configured conditional rules containing multi-dimensional compliance attributes are injected into the bipartite graph to initiate the evolution and deduction of the graph state. The credibility of atomic declaration nodes is represented as a discrete state machine, and its initial value is set as a highly optimistic verification hypothesis.

[0040] A discrete state machine can contain the following ordered set of states: fully supported states, partially supported states, conflict states, rejected states, and insufficient authority states. The fully supported state is the highest confidence state and is assigned as the initial value to each atomic declaration node; the insufficient authority state and the rejected state are the lowest confidence states. During iterative propagation, node states are only allowed to monotonically transition negatively from high confidence to low confidence; reverse transitions are not permitted.

[0041] A cyclic propagation is performed within the graph network. In each propagation cycle, each atomic declaration node is checked for violations of any pre-configured constraint criterion, such as: lack of unidirectional support from valid evidence nodes, conflicting node content, expired validity of the relied-upon evidence, or access barriers to underlying evidence. Upon detection of an out-of-bounds violation, the activation state of the corresponding edge in the bipartite graph is forcibly changed, triggering a monotonically negative transition in the declaration node's state machine. This iterative propagation oscillates within the graph structure in a cascading manner. The iterative computation stops when no connected edges in the network experience state reversals and all node states remain static, capturing the current static state as the converged declaration credibility verification state.

[0042] Step 107: Based on the declaration-evidence bipartite graph structure, perform global consistency optimization to solve the declaration credibility verification state, and obtain the final output state and the target evidence reference unit.

[0043] After obtaining the static graph cross-section of iterative convergence, to eliminate the entanglement of redundant nodes that may be left over from local rules, global optimization computation is initiated under the boundary constraints of the declaration-evidence bipartite graph structure. By constructing a mathematical programming model describing the overall graph's gains and costs, the algorithm performs search or combinatorial enumeration within multiple feasible solution spaces. While eliminating interference noise, a minimal subset of evidence is selected from the right-hand node set that maximizes the legitimacy gains of the left-hand declaration nodes. After optimization convergence, a definite final output state is locked for each atomic declaration node, and the target evidence citation unit that provides strong corroboration (or contradictory accusation) for it is output simultaneously.

[0044] Step 108: Rewrite the candidate answer draft based on the final output state and the target evidence citation unit, and output the credible question answer result.

[0045] The rendering engine extracts the final output state and locked evidence. Comparing the original proposed text (i.e., candidate answer drafts) stored in step 103, it performs mandatory content filtering and reconstruction operations on the candidate answer drafts based on the state markers. For factual fragments with compliant states, they are retained, and target evidence citation units containing the corresponding source identifiers are injected at the end of the text; for factual fragments with abnormal states, such as insufficient evidence, logical conflicts, or restricted permissions, text erasure is performed, and they are replaced with the corresponding system disclaimer, making the information on the output interface transparent and verifiable. After a full rewrite and assembly, the credible question-and-answer result is delivered to the front-end caller.

[0046] Example 2

[0047] Based on the above embodiment one, the hybrid retrieval scoring items, multi-level dynamic permission gating, and query cache matching process are further described in detail.

[0048] In one possible implementation, the document set is retrieved and filtered based on user queries and access permissions to obtain a candidate evidence unit set, which is then referred to... Figure 2 Specifically, this can be achieved through the following comprehensive scoring mechanism:

[0049] Semantic dense features and sparse keyword features are extracted from each document unit in the user query and document set, respectively;

[0050] Semantic relevance is calculated based on dense semantic features, and keyword matching degree is calculated based on sparse keyword features;

[0051] The timeliness decay of a document unit is calculated based on the query timestamp of the user query and the effective timestamp of the document unit, combined with the pre-configured decay period.

[0052] Extract the coverage of user queries in the headings of the document unit's level to obtain the heading coverage;

[0053] The semantic relevance, keyword matching degree, timeliness decay degree and title coverage are weighted and summed to obtain the comprehensive ranking score of each document unit;

[0054] Based on this (comprehensive ranking score), a set of candidate evidence units is obtained.

[0055] Upon receiving a user query, the retrieval system uses a pre-trained vector encoding model to perform low-dimensional embedding calculations on the query, mapping it to a dense vector with fixed dimensions. Simultaneously, it retrieves a pre-built inverted index to segment the query into words, obtaining sparse statistical features. Correspondingly, for each document unit in the pre-stored document set, its corresponding dense vector and sparse word frequency statistical features have been pre-calculated and stored during the offline preprocessing stage.

[0056] In order to filter out clauses that are highly consistent in both semantics and wording from a massive amount of corporate policies or legal contracts, the comprehensive ranking score of each document unit is output by combining the following four detailed indicators.

[0057] The first metric is semantic relevance. It characterizes the degree of closeness between the dense vector of the user query and the dense vector of the document unit at a deep abstract semantic level by calculating the cosine similarity between them.

[0058] The second metric is keyword matching accuracy. To eliminate the inconsistency in the units of sparsity features across documents of different lengths, the base BM25 score is normalized and scaled based on the maximum and minimum word frequencies of the current user query within a preset candidate set, as calculated below:

[0059] BM25 norm =(BM25-BM25 min ) / (BM25 max -BM25 min +ε);

[0060] Among them, BM25 norm BM25 represents the keyword matching score, and BM25 is the original BM25 score of the current document unit. min The minimum original BM25 score of document units within the current candidate set, BM25 max The maximum value of the original BM25 score of document units within the current candidate set, where ε is a pre-configured division-to-zero constant with a value of 10.-8 After this formula conversion, the keyword matching degree is strictly limited to a closed interval between 0 and 1.

[0061] The third metric is the time-lapse rate. When dealing with time-sensitive corporate guidelines or supplementary agreements, the time difference between the evidence release time and the time of the inquiry is crucial in determining the validity of a citation. The time-lapse rate is generated by measuring the difference between the query timestamp associated with the user's query and the effective timestamp of the corresponding document unit. The calculation process is as follows:

[0062] fresh=max(0,1-(t q -t eff ) / T decay );

[0063] Where fresh represents the time-related decay rate, and t q For the query timestamp included in the user query, t eff T is the effective timestamp corresponding to the document unit. decay The pre-configured decay period is set to 730 days in this embodiment. This decay rate reflects a qualitative mapping relationship where newer document units score higher. When the time difference exceeds the decay period, this item scores 0. eff >t q If the document is not yet effective, it will be directly removed during the initial admission filtering stage and will not be included in the subsequent ranking calculation. If there are some outdated evidence units that do not have an explicit effective time stamp, their time decay rate will be uniformly assigned a neutral default value of 0.5.

[0064] The fourth metric is title coverage. To strengthen the constraint of document structure hierarchy on query intent, all parent headings (including first-level headings to the current section heading) of the paragraph containing the document unit are retrieved. The number of intersections between the word segmentation set in the user query and the word segments in the parent heading sets is calculated, and the ratio of this intersection to the total number of word segments in the user query is calculated. The maximum coverage rate among multiple parent headings is extracted as the title coverage. The calculation process is as follows:

[0065] title=max(intersect(tokens q tokens h ) / tokens count );

[0066] Where title represents the title coverage, and tokens q The set of terms used by the user for querying, tokens h Given a title set of tokens, intersect(tokens) q tokens h(tokens) are used by users to query the number of intersections between a token and a specific title token. count The total number of terms in the user query, max(...) represents the maximum value after traversing all the level headings to which the evidence unit belongs.

[0067] After calculating the four basic metrics for each document unit, the pre-configured weighting coefficients are retrieved, and the semantic relevance, keyword matching, timeliness decay, and title coverage are weighted and summed to obtain the final comprehensive ranking score. An example of the calculation formula is as follows:

[0068] score = 0.65 * cos + 0.35 * BM25 norm +0.08*fresh + 0.05*title;

[0069] Where score is the overall ranking score, cos is the semantic relevance, and BM25 is the BM25 ... norm represents keyword matching accuracy, fresh represents time-lapse rate, and title represents title coverage.

[0070] It should be noted that this formula is a non-normalized linear ranking score. Since semantically dense features are generated using a normalized forward spatial coding model, the output range of semantic relevance is limited to 0 to 1. Therefore, this formula is a non-normalized linear ranking score, with a total value range between 0 and 1.13. It is only used to rank the candidate queue by relative position and does not have a probability density distribution meaning. Based on the comprehensive ranking score, documents are aligned from high to low. The top 30 documents are then fed into the re-ranking model, which ultimately selects the top 12 documents as the candidate evidence unit set. The re-ranking model can employ a cross-encoder re-ranking model known to those skilled in the art.

[0071] It should be further noted that the specific values ​​of the above weighting coefficients can be determined by those skilled in the art through cross-validation or grid search based on actual business datasets. The values ​​given in this embodiment are only a preferred configuration.

[0072] In some alternative implementations, the generation process of candidate answer drafts can be switched from a single natural language-driven mode to a dual-track fusion-driven mode during this step. Specifically, a query intent analyzer is used to extract the query subject, query attributes, time range, and conditional premises from the user query to generate a structured query Q. If the user's question is transformed into a query about corporate reimbursement limits, the natural language question can be transformed into a structured query Q: the query subject is the finance department, the query attribute is travel standards, the time range is the current year, and the conditional premise is senior manager. The generation model simultaneously receives the user query text, the structured query Q, and the candidate evidence unit set, and jointly generates the draft answer in a dual-track input constraint form, thereby reducing the risk of the draft answer escaping at the entity attribute level.

[0073] In one possible implementation, refer to Figure 3 The document collection is searched and access-filtered using a multi-level dynamic gating mechanism, which may include the following processes:

[0074] When acquiring the set of candidate evidence units, access permissions are used to perform preliminary admission filtering on the document set.

[0075] When constructing the declaration-evidence bipartite graph structure, a secondary comparison and verification is performed using access permissions and the permission identifiers of each evidence unit in the candidate evidence unit set;

[0076] If the access permission is detected to be inconsistent with the permission identifier of the target evidence unit (i.e., a certain evidence unit in the candidate evidence unit set), the association relationship corresponding to the target evidence unit will be forcibly severed in the declaration-evidence bipartite graph structure.

[0077] To avoid the vulnerability of unauthorized data penetration that may occur when multiple users share the same context, this method adopts a multi-level dynamic gating mechanism that spans multiple stages.

[0078] In the initial access filtering stage, such as step 102, when the retrieval module loads documents, it filters the document set according to the identity and permission token of the current query subject. Only entries whose security level or access control list is included in the access permission scope can be loaded into the candidate queue.

[0079] However, because enterprise systems support dynamic permission changes, there is a possibility that a user's access permissions may be asynchronously revoked midway through a query, or that the permission identifier corresponding to the original evidence unit may have changed during incremental updates or when loading existing historical graph data directly from the historical graph cache. To defend against such security risks, a secondary comparison and verification gate is set when constructing the declaration-evidence bipartite graph structure and initializing graph variables.

[0080] The system obtains the current access permissions detected in real time and performs a secondary comparison with the inherent permission identifier within the right-hand evidence unit node in the bipartite graph structure, using an element inclusion set. Specifically, a Boolean indicator function is introduced. When it is determined that the access permissions do not contain the permission identifier of the target evidence unit, the Boolean indicator function outputs 0, causing the activation initialization scores of all supporting edges and conflicting edges associated with that evidence node to be forcibly multiplied by 0. In other words, the association relationship corresponding to the target evidence unit is forcibly severed on the data link, isolating it from the logical transmission network of the bipartite graph. This design constructs a low-level security defense at the graph construction level through redundancy checks.

[0081] To differentiate the causes of missing evidence during subsequent constraint propagation, a permission blocking flag is synchronously written to the state attribute of the corresponding declaration node when the Boolean indicator function is set to zero. During iterative constraint propagation, when an atomic declaration node triggers state degradation due to a lack of valid supporting edges, it is checked whether the node carries a permission blocking flag. If it carries this flag and the cause of the missing supporting edges is entirely due to permission severance operations, the node's declaration credibility verification state is set to insufficient permissions, rather than the general rejection state. This differentiation mechanism enables the front-end rewrite engine to output accurate access-restricted disclaimers to users, rather than vague insufficient evidence prompts.

[0082] An alternative approach may include a query cache matching step before retrieving and filtering the document collection based on user queries and access permissions:

[0083] Extract the feature vector of the user query and perform similarity matching in the pre-stored historical query cache;

[0084] If a high-confidence historical query is matched, that is, a historical query with a similarity exceeding the preset cache threshold, and the access permissions have not changed, then the associated historical declaration-evidence bipartite graph structure snapshot is directly extracted and reused for local constraint incremental update to obtain the updated declaration-evidence bipartite graph structure, and global consistency optimization is performed based on the updated declaration-evidence bipartite graph structure.

[0085] To reduce the high computational cost of repeated queries in graph computation and optimization, this method incorporates a cache interception mechanism based on graph snapshot reuse.

[0086] In practice, upon receiving a user query in natural language, it is first converted into a high-dimensional feature vector and loaded into a pre-stored historical query cache for distance metric matching. The cache stores historical query feature vectors, corresponding query permission logs, and snapshots of the declaration-evidence bipartite graph topology at that time.

[0087] If the similarity between the feature vector of the current user query and the feature vector of a certain historical query in the cache is higher than the preset high-confidence cache threshold of 0.95, then a high-confidence historical query match is determined. At this point, the currently acquired access permissions are further verified against the historical permissions recorded in the cache log. If the contents of both sets are consistent, confirming that the access permissions have not changed, the question-and-answer process directly intercepts the subsequent global retrieval and generation actions, and directly extracts the corresponding historical declaration-evidence bipartite graph structure snapshot.

[0088] Based on this, only for the timeliness changes or minor metadata changes caused by the current query timestamp, local constraint incremental updates and re-convergence calculations are performed on the graph snapshot, thereby compressing the system response latency to a certain extent.

[0089] Example 3

[0090] Based on the above embodiments one and two, the extraction of logical dependencies of atomic declaration node sets, construction of directed acyclic dependency graphs, and state cascading degradation mechanism in iterative propagation are further described in detail.

[0091] In one possible implementation, after semantically segmenting the candidate answer drafts to obtain a set of atomic declaration nodes, the following steps may also be included:

[0092] Step 301: Identify the logical dependencies between atomic declaration nodes in the atomic declaration node set. Logical dependencies include prerequisite dependencies, mutual exclusion dependencies, and logical implication dependencies.

[0093] A structured mapping is performed on each atomic declaration node in the set of atomic declaration nodes. The processing model (segmentation model) extracts the independent semantic components in each node and normalizes them into declaration tuples containing six dimensions. An example of the specific mathematical representation of the declaration tuple is shown below:

[0094] c i =(text,subject,predicate,object,condition,time);

[0095] Among them, c i Let i be the i-th atomic declaration node, where text is the declaration text, subject is the subject entity, predicate is the relational predicate, object is the object, condition is the precondition, and time is the time constraint.

[0096] After extracting the declaration tuples, logical dependencies between nodes are identified using a joint rule trigger and a sequence classification model. The rule trigger performs regular expression matching on the condition tag words contained in the condition dimension to extract basic dependency associations. Simultaneously, the sequence classification model uses a bidirectional long short-term memory network to process the text sequences of two declaration nodes, outputting a prediction probability matrix. Among these, the premise dependency relationship indicates that the downstream node is true only if the upstream node is true; the mutual exclusion relationship indicates that the states of the two nodes cannot both be true simultaneously; and the logical implication relationship indicates that the truth of the upstream node directly implies the truth of the downstream node.

[0097] Step 302: Based on logical dependencies, construct a directed acyclic dependency graph with topological pointing within the set of atomic declaration nodes to provide conditional constraints in subsequent iterative constraint propagation.

[0098] The identified logical dependencies are extracted as topological edges, and atomic declaration nodes are treated as network nodes. A local topological network connecting these declaration nodes is constructed in the left space of the bipartite graph. To provide monotonically convergent properties and avoid state deadlock during iterative updates, a forced edge-breaking and loop-prevention algorithm is applied to the local topological network. When the topology traversal algorithm detects a closed loop composed of premise dependencies or logical implications, the predicted probability scores of each topological edge constituting the loop are retrieved. The edge with the lowest score is automatically removed through numerical comparison, cutting off the loop path and outputting a directed acyclic dependency graph that meets the topological sorting requirements.

[0099] Optionally, iterative constraint propagation is performed on the declaration-evidence bipartite graph structure based on pre-configured verification constraint rules, including the following declaration dependency cascading rules:

[0100] 1) In each round of iterative updates, monitor each atomic declaration node in the directed acyclic dependency graph in real time.

[0101] Within the computational cycle of iterative constraint propagation in the declaration-evidence bipartite graph structure, cascading update logic for the left-hand declaration nodes is introduced. A topological sorting algorithm for the directed acyclic dependency graph is used to generate the node traversal sequence. In a single iteration, the state machine sequentially reads the current declaration credibility verification state of each atomic declaration node along the pointing path of the graph structure.

[0102] 2) When an atomic declaration node is determined to be a premise declaration node and its declaration credibility verification status is updated to false, the declaration credibility verification status of the downstream conclusion declaration node pointed to by the premise declaration node in the directed acyclic dependency graph is forcibly downgraded to a rejection status.

[0103] During the execution of the traversal sequence, causal constraints between nodes are extracted. The set of invalid states is defined to include conflicting states and initial rejection states. When the state machine detects that the state of a certain atomic declaration node undergoes a negative transition and falls into the set of invalid states, and that the node has outward-pointing prerequisite dependency edges in the directed acyclic dependency graph, the contrapositive cascaded degradation calculation mechanism is triggered.

[0104] Track all downstream conclusion statement nodes connected by the pointing edge, without determining whether there are supporting evidence units on the right side, and directly overwrite their statement credibility verification status to a lower confidence rejection status through graph variable update operations. This constraint intercepts the output of complex facts when the basic premises are missing, ensuring that the complex factual logic tree of the output remains coherent from bottom to top.

[0105] Example 4

[0106] Based on the above embodiments, the process of constructing relationships in the declaration-evidence bipartite graph structure and the adaptive dynamic gating strategy for business scenarios are further described in detail.

[0107] In one possible implementation, a declaration-evidence bipartite graph structure is constructed based on the association between the set of atomic declaration nodes and the set of candidate evidence units, referring to... Figure 4 It may include the following steps:

[0108] Step 401: Compare and infer each atomic declaration node in the atomic declaration node set with each candidate evidence unit in the candidate evidence unit set, and output the corresponding support probability, conflict probability and indeterminate probability respectively.

[0109] After completing atomic statement splitting and evidence retrieval, the set of atomic statement nodes and the set of candidate evidence units are traversed to form pairwise node pairs. For each pair of nodes, the natural language inference model is invoked to perform textual implication computation. Specifically, the natural language inference model adopts a multi-layer Transformer encoder structure, and its input sequence is concatenated according to predetermined placeholders. The concatenation format is: starting classification label, evidence unit text, separator label, statement text, and ending separator label. After processing the concatenated sequence, the model extracts the output vector corresponding to the starting classification label and maps it to three-dimensional space. After normalization by the Softmax function, the natural language inference model outputs a probability distribution vector containing three floating-point numbers. The elements in this vector represent the probability distributions of evidence implication statement, evidence and statement contradiction, and the two being unrelated.

[0110] Step 402: Extract the maximum probability value among the three (support probability, conflict probability, and indeterminate probability).

[0111] After obtaining the probability distribution vector, a numerical comparison operation is performed on the three elements in the vector. The component with the largest value is selected as the maximum probability value, and its corresponding classification label is recorded as the initial candidate relation type. This operation is used to establish the dominant semantic relation dimension between the current evidence and the claim, and to filter out the interference of low-probability noise.

[0112] Step 403: When the maximum probability value reaches the preset gating threshold, establish a corresponding type of association between the corresponding atomic declaration node and the candidate evidence unit. The association includes support relationship, conflict relationship or indeterminate relationship.

[0113] After extracting the maximum probability value, this value is compared with a set truncation reference benchmark. If the value is lower than the benchmark, the current evidence unit is determined to have no substantial reference value for the claim node, and the establishment of a topological connection between the two is abandoned. When the maximum probability value is greater than or equal to a preset gating threshold, the bipartite graph construction module writes a directed edge between the atomic claim node on the left and the candidate evidence unit on the right, and writes the maximum probability value as the basic edge weight into the weight matrix of the graph structure. According to the initial candidate relation type recorded in the previous sequence, the directed edge is instantiated as one of the following: a supporting relation, a conflicting relation, or an indeterminate relation. After traversing all node pairs for filtering and edge construction, an initial claim-evidence bipartite graph structure containing a sparse connection matrix is ​​generated.

[0114] An alternative approach is to preset the gating threshold using a dynamic adaptive strategy based on the business scenario:

[0115] Step 4031: Obtain the pre-configured business scenario tags for the current question-and-answer task. The business scenario tags correspond to the preset fault tolerance sensitivity levels.

[0116] To adapt the graph-based gating logic to the compliance tolerance levels of different vertical domains, the system synchronously parses environmental parameters injected by external systems and extracts pre-configured business scenario tags when receiving user queries. The system pre-stores a mapping configuration table that defines the mapping relationship between tag categories such as general Q&A, casual conversation, financial auditing, and legal contract review, and their corresponding fault tolerance sensitivity levels. By looking up the table, the corresponding fault tolerance sensitivity level parameters are attached to the current task.

[0117] Step 4032: In scenarios with high sensitivity in finance, law, or policy, dynamically increase the threshold for determining support relationships and correspondingly decrease the threshold for determining conflict relationships to implement an asymmetric, conservative determination strategy.

[0118] In standard business scenarios, both the threshold for determining support relationships and the threshold for determining conflict relationships use neutral benchmark constants, such as a uniform setting of 0.45. However, when the current pre-configured business scenario tag is identified as belonging to a high-sensitivity scenario such as financial auditing, legal compliance, or policy interpretation, an asymmetric threshold adjustment mechanism is triggered.

[0119] In such highly sensitive scenarios, strict control over erroneous support results is crucial, while simultaneously exposing all potential contradictions. Basic gating parameters are extracted, and two types of thresholds are updated based on an asymmetric adjustment function. On one hand, the threshold for determining support relationships is raised. This threshold is configured within a high-confidence range, for example, limited to a value between 0.70 and 0.75. Preferably, the support relationship threshold is set to 0.72, requiring evidence to provide strong logical support for the statement to establish a valid connection. On the other hand, the threshold for determining conflict relationships is lowered, configured within a low-confidence range of 0.20 to 0.30. Preferably, the conflict relationship threshold is set to 0.25. This asymmetric configuration logic keeps the system in a conservative operating state, tolerating false positives but not false negatives. Even if there is only a slight probability of textual conflict between the evidence and the statement, the conflict edge will be forcibly retained and participate in subsequent iterative constraint propagation, thereby triggering a mandatory risk warning in the final question-and-answer result.

[0120] Example 5

[0121] Based on the above embodiments, another implementation method for bipartite graph coverage constraints is provided, and the mechanism of conjoint evidence group supported by multiple pieces of evidence is further described in detail.

[0122] In one possible implementation, the coverage constraint in the verification constraint rule includes a mechanism for independent support of a single piece of evidence, or a mechanism for joint evidence support of multiple pieces of evidence.

[0123] In the iterative propagation of the declaration-evidence bipartite graph structure, evidence coverage verification is performed on each atomic declaration node. Conventional logical verification uses disjunctive logic, requiring a single candidate evidence unit to cover all facts in the declaration. To accommodate scenarios where long-sentence declarations contain fragments across documents, a dual-track coverage constraint system is configured. When the current atomic declaration node contains only a single-dimensional basic fact, a single-evidence independent support mechanism is triggered; when the current atomic declaration node contains composite facts spanning multiple business entities or distributed across different physical documents, a conjunctive evidence group mechanism is triggered, jointly supporting multiple pieces of evidence.

[0124] One possible implementation of a conjoint evidence group mechanism supported by multiple pieces of evidence may include the following steps:

[0125] Step 501: Split the atomic declaration node into a semantic component vector consisting of multidimensional independent facts.

[0126] Upon entering the conjunctive evidence group mechanism, the processing model performs syntactic dependency analysis and entity recognition on the content of the current atomic declaration node. Based on the space, time, subject, and preconditions of the event, the originally continuous text sequence is deconstructed into discrete semantic slices. These semantic slices are encapsulated into fixed-length numerical and character features, forming semantic component vectors.

[0127] V sem =[sem1,sem 2, ...,sem K ];

[0128] Among them, V sem The semantic component vector of the target atom declaration node, from sem1 to sem K These represent the first to the Kth independently derived factual components, where K is the total number of factual components.

[0129] Step 502: For each candidate evidence unit with a supporting type association, calculate its component coverage result for each fact component in the semantic component vector.

[0130] After extracting the semantic component vector, all edges connected to the current atomic declaration node are traversed in the declaration-evidence bipartite graph structure to filter out a subset of candidate evidence units with supporting relationships. Using local inference operators, local implication calculations are performed sequentially between each fact component in the semantic component vector and each candidate evidence unit in the subset. If the calculated semantic alignment score reaches the baseline alignment threshold, the component coverage result is marked as 1; otherwise, it is marked as 0. Based on this, a component coverage indicator matrix is ​​constructed, whose elements consist of the aforementioned 0s or 1s, to characterize the local support state of the predetermined candidate evidence units for the predetermined fact components.

[0131] Step 503: When all fact components of an atomic declaration node are covered by at least one candidate evidence unit, the coverage constraint is satisfied.

[0132] After obtaining the component coverage indication matrix, branch logic decision-making is performed. In the first decision-making scenario, multi-evidence accumulation verification is performed for each fact component, and the decision calculation logic is as follows:

[0133] S k =∑ j (x j *C kj );

[0134] Among them, S k For the accumulated support value of the k-th fact component, ∑j This represents the summation calculation over all candidate evidence units j with supporting relationships, x j Let C be the activation variable for the support edge corresponding to the j-th candidate evidence unit. kj This represents the component coverage result of the j-th candidate evidence unit for the k-th fact component.

[0135] When the cumulative support value S of all factual components corresponding to variable k is detected k When ≥1, the current graph state is determined to satisfy the conjunctive constraint. Under this condition, the current atomic declaration node retains its support state within the high confidence interval, allowing multiple candidate evidence units with complementary attributes to jointly provide legitimacy endorsement for it.

[0136] Step 504: If some factual components are not covered, it is determined to be a partial support state, and a list of specific factual components of the missing evidence in the atomic declaration node is output.

[0137] Under the second judgment condition, continue to monitor the cumulative support value S. k The output result. If at least one fact component is detected, its corresponding support accumulation value S k If the cumulative support value is 0, the atomic declaration node is determined to have failed to achieve a closed-loop factual logic. The declaration credibility verification state of this atomic declaration node is then forcibly transferred to a local support state at an intermediate level. Simultaneously, the original character slices of the factual components with a cumulative support value of 0 are extracted, combined, and output as a list of specific factual components lacking evidence. This list is then written to a memory array to expose to the calling client the locally unfalsifiable information caused by external document fragmentation.

[0138] By using the aforementioned joint evidence group mechanism, multiple complementary pieces of evidence can be used to jointly verify a composite statement in scenarios where a single piece of evidence cannot fully cover it, thus avoiding overall false rejection due to the lack of partial evidence.

[0139] Example 6

[0140] Based on the above embodiments, another implementation method for bipartite graph time constraints is provided. A further detailed description of the numerically sensitive clause-level substitution dominance mechanism is given.

[0141] In one possible implementation, the timeliness constraint in the validation constraint rule includes a whole-document replacement domination mechanism for the same source version, or a clause-level replacement domination mechanism for the same source version.

[0142] During the iteration process of corporate rules and regulations or normative documents, multiple historical versions of the same source document typically coexist in the knowledge base. The calculation module identifies old and new source documents with a historical relationship by comparing document identifiers, release timestamps, and version number sequences. In the constraint deduction of the declaration-evidence bipartite graph structure, a dual-track version elimination filtering rule is configured. For short documents that are completely abolished or completely revised, the entire document replacement and domination mechanism is triggered. Through a global state overwrite operation, the edge activation variables corresponding to all evidence units derived from the old document are forcibly set to zero. For long corporate rules and regulations, the new document usually only makes partial revisions to some chapters or monetary standards. In this case, the clause-level replacement and domination mechanism of the same source version is triggered to retain the evidentiary effect of the unmodified parts of the old document and prevent irrelevant clauses from being mistakenly eliminated.

[0143] In one possible implementation, the same-source version's clause-level substitution dominance mechanism may include the following steps:

[0144] Step 601: Based on the document identifier and version mark of each candidate evidence unit in the candidate evidence unit set, identify the old and new versions of documents with version evolution relationship; calculate the basic text difference degree of the corresponding clauses between the old and new versions of documents contained in the candidate evidence unit set.

[0145] After triggering the clause-level replacement dominance mechanism, hierarchical alignment and local comparison are performed on the old and new versions of the same source document. Using the clause numbers or chapter anchors obtained from document parsing, corresponding clause text pairs are extracted from the old and new versions of the same source document. After extraction, the longest common subsequence algorithm is introduced to calculate the proportion of common characters between the two string sequences, as follows:

[0146] diff = 1 - (2 * L) lcs ) / (L old +L new );

[0147] Among them, diff is the basic text difference, L lcs L is the length of the longest common subsequence of the old and new clause texts. old The length of the old terms and conditions text, L new This refers to the length of the new terms and conditions text.

[0148] In some optional implementations, for unstructured prose documents without explicitly labeled clause numbers, a combination of structural path and semantic scope is used to determine the overlap of corresponding clauses. The upper-level title path of the evidence unit is extracted, and structural overlap is determined by calculating the longest common prefix of the old and new node paths. When document restructuring causes path invalidation, the main entities in the text are extracted, and Jaccard similarity is calculated. If the Jaccard similarity is ≥0.5 and the node's subject keywords are consistent, a corresponding clause relationship is determined between the old and new nodes, thus triggering the aforementioned basic text difference calculation process.

[0149] Step 602: Extract the numerical entities in the corresponding clauses. When it is determined that the numerical entities of the old and new clauses are inconsistent, the basic text difference degree is forcibly increased to obtain the final clause difference degree.

[0150] In legal contracts or financial audits, substantive revisions to clauses often involve changes to discrete numerical values ​​such as time, amount, and proportion, while the surrounding grammatical structure and most of the text remain fixed. The basic textual difference, dominated by the longest common subsequence length, exhibits a passive characteristic towards pure numerical replacements. To overcome this algorithmic flaw, a named entity recognition model is invoked to extract discrete numbers, currency units, and time parameters from both the old and new clause texts in parallel, forming a set of numerical entities. The numerical entity sets of the old and new clauses are then cross-referenced, and a numerical indicator function is used to forcibly adjust the basic textual difference.

[0151] diff adj =max(diff,I num );

[0152] Among them, diff adj For the final clause difference, diff is the basic text difference, I num This is a function to indicate inconsistencies in numerical entities; I indicates inconsistencies when the extracted numerical entities are different. num The value is 1 otherwise the value is 0. max(...) means taking the maximum value of the two items inside the parentheses.

[0153] For example, suppose in one application scenario, the basic text difference (diff) between the old and new versions of the clauses is 0.14 based on a basic text comparison, and the preset substantive modification threshold is 0.15. Through entity extraction, it is found that the old clause contains the value 600, while the corresponding position in the new clause contains the value 800. The calculation module determines that the numerical entities are inconsistent, and at this point, the indicator function I... num The value is assigned to 1. Substituting the above parameters into the adjustment formula, the final clause difference (diff) is obtained. adj=max(0.14,1)=1. This mechanism bypasses the tolerance range of text similarity calculation, and makes a significant numerical increase for minor character changes in amount or standard.

[0154] It should be noted that the specific value of the aforementioned preset substantial modification threshold can be determined by those skilled in the art through cross-validation or grid search based on actual business datasets. The value given in this embodiment is only a preferred configuration.

[0155] Step 603: If the final clause difference reaches the preset substantive modification threshold, it is determined that the corresponding clause is covered by the new version, and the association relationship of the corresponding candidate evidence unit in the old version document is forcibly severed in the declaration-evidence bipartite diagram structure; if the final clause difference does not reach the preset substantive modification threshold, the association relationship of the corresponding candidate evidence unit in the old version document is retained.

[0156] After obtaining the final clause difference degree, it is numerically compared with a preset substantial modification threshold. If the comparison result shows that the final clause difference degree is greater than or equal to the preset substantial modification threshold, it is determined that the content of that clause in the old document has been substantially covered by the new version, and its evidentiary effect has expired. In the declaration-evidence bipartite graph structure, the candidate evidence unit node derived from the old clause is located, and the activation variables of all edges connecting this node to support and conflict relationships are multiplied by the Boolean value 0. This operation forcibly severs the association relationship of the invalid node in the graph topology, preventing it from transmitting false confidence support to the left-hand declaration node during iterative constraint propagation.

[0157] In practice, if the final difference in terms does not reach the preset substantive modification threshold, the clause is considered to be in an unmodified, static state, and a cross-version content conflict detection mechanism is further activated. The unmodified evidence text corresponding to the old version of the same source document is injected into the precondition sequence of the natural language inference model, and the associated evidence text corresponding to the new version document is injected into the hypothesis sequence. The model outputs the probability of contradiction for the two version sequences. If this probability is ≥0.60, it is marked as a real business data conflict not caused by version change. At this point, no severance operation is performed; instead, the conflict relationship is preserved in the declaration-evidence bipartite graph structure to trigger subsequent graph node state machine rejection degradation or conflict alarm responses.

[0158] Example 7

[0159] Based on the above embodiments, this embodiment further details the two-level acceleration algorithm in the global consistency optimization solution stage, the terminal polymorphic text rewriting mechanism, and the front-end interactive presentation method.

[0160] In one possible implementation, a global consistency optimization solution is performed on the declaration-evidence bipartite graph structure to solve the declaration credibility verification state. Specifically, the state variables of each atomic declaration node and the edge activation variables of each association are used as constraint variables to construct a 0-1 mixed integer linear programming model for solution, thereby obtaining the final output state and the target evidence reference unit. The comprehensive optimization objective of the 0-1 mixed integer linear programming model includes: maximizing the weight score of the activated support relationship, while minimizing the penalty weight value of the activated conflict relationship, the baseline penalty term of the rejection state, and the missing penalty term of the local support state.

[0161] In some embodiments, the overall optimization objective of the 0-1 mixed integer linear programming model includes: maximizing the weight scores and evidence coverage of activated support relations, while minimizing the penalty weights of activated conflict relations, the baseline penalty term for the refusal state, and the penalty term for the missing local support state.

[0162] After convergence of iterative constraint propagation, the atomic declaration nodes are assigned a definite declaration credibility verification state. Using this initial state as the starting point for optimization, a global optimization computation layer is initiated under the topological constraints of the declaration-evidence bipartite graph structure. In this computation layer, the state variables of each atomic declaration node and the edge activation variables of each association are all used as constraint variables to be solved. Among these, the state of nodes forcibly downgraded by permission constraints or declaration dependency cascade rules is a hard constraint and cannot be changed; the states of other nodes can be adjusted within the legal state space as edge activation combinations change. By constructing a mathematical programming model describing the overall graph's gains and losses, the algorithm performs search or combinatorial enumeration within the feasible solution space, and its comprehensive optimization objective is calculated as follows:

[0163] Obj=∑(P s *X s )-β*∑(P c *X c )-γ*N abs -δ*N part ;

[0164] Where Obj is the comprehensive optimization objective value, ∑(P s *X s The sum of the weighted scores of all activated support relationships is represented by P. Since activating each support relationship edge signifies the selection of a candidate evidence unit into the support system for the claim, the increase in the value of this sum reflects the improvement in evidence coverage. s X represents the weight score of the support relationship. s This represents the activation variable that supports the relation edge, corresponding to x as described in Example 5. j β is the preset conflict penalty coefficient, ∑(P c *X cP represents the sum of the penalty weights for all activated conflict relations. c X represents the penalty weight value for conflict relationships. c N represents the activation variable of the conflict edge, γ is the baseline penalty coefficient for the rejection state, and N is the activation variable of the conflict edge. abs N represents the number of atomic declaration nodes in the refusing state, δ represents the missing penalty coefficient for local support states, and N represents the number of atomic declaration nodes in the refusing state. part The values ​​of the coefficients β, γ, and δ are calibrated to balance the magnitudes of the terms, representing the number of atomic declaration nodes in a locally supported state. This comprehensive optimization objective drives the solver to seek evidence activation combinations that allow more nodes to escape unanswered or locally supported states.

[0165] When solving this model, the algorithmic instructions aim to maximize the overall optimization objective value. When configuring specific coefficient parameters, the local support penalty is set lower than the rejection penalty to output locally verifiable data. In this embodiment, β is set to 2.0, γ to 0.3, and δ to 0.15. The specific values ​​of the above penalty coefficients can be determined by those skilled in the art based on actual business datasets through cross-validation or grid search; the values ​​given in this embodiment are merely a preferred configuration.

[0166] According to one aspect of this application, the 0-1 mixed-integer linear programming model includes the following constraints:

[0167] Constraint 1 (State Mutual Exclusion): The sum of the fully supported state variables, locally supported state variables, conflicting state variables, and unanswered state variables of each atomic declaration node is equal to 1;

[0168] Constraint 2 (Support Activation Coupling): If an atomic declaration node is in a fully supported state, then there exists at least one support relationship edge associated with it with an activation variable of 1;

[0169] Constraint 3 (Conflict Activation Mutual Exclusion): For the same atomic declaration node, if the activation variable of any conflicting edge is 1, then the node cannot be in a fully supported state.

[0170] An alternative approach is to use a two-stage solution strategy for the 0-1 mixed-integer linear programming model:

[0171] In the first stage, the declared credible verification state reached by the iterative constraint propagation is directly used as the initial feasible solution for greedy state initialization;

[0172] The second stage involves performing a finite number of random searches for edge activation combinations within the local neighborhood of the initial feasible solution within a preset near real-time solution timeframe. This enumerates the activation combinations that maximize the overall optimization objective and solves for the final output state and the target evidence reference unit.

[0173] The local neighborhood is defined as follows: based on the initial feasible solution, each time a declared node is selected, the activation state of all its connected edges is flipped by a single bit to form a candidate combination; or the activation states of the edges of two adjacent nodes are flipped at the same time to form a second-order neighborhood.

[0174] To address the computational latency limitations imposed by online data flow, a two-stage computation acceleration scheme is adopted. The maximum number of atomic declaration nodes in a single response is constrained to 15, and the number of candidate evidence is constrained to 12, thereby compressing the total number of constraint variables in the graph space to within 180 dimensions. During the first stage of execution, each node is traversed, and the two connected edges with the highest weights are directly extracted and activated in descending order of support probability, outputting a greedy computational solution.

[0175] After obtaining the aforementioned greedy solution, the second-level optimization is driven. A time monitor is activated, and within a preset near-real-time solution timeframe of 50 milliseconds, declared nodes in the graph are randomly located, and the Boolean activation combinations of their connected edges are enumerated. Within a finite time window, the comprehensive optimization objective value after each combination update is calculated, and combination schemes that cause this value to show an upward trend are extracted and saved. If a matrix constraint conflict is encountered during the enumeration search, rendering the solution infeasible, the calculation is interrupted, and the static graph cross-section at the end of the iterative constraint propagation is extracted as a fallback solution. This establishes the final output state of each node and locks the currently activated connected edge as the target evidence reference unit.

[0176] In non-real-time offline compliance audit computation scenarios, as an alternative, the local neighborhood random search step in the two-level solution strategy is disabled, and instead, the branch and bound algorithm is invoked to perform constraint space solution. The search depth limit parameter is configured to twice the total variable size, and when the physical computation time reaches the blocking threshold, it is forcibly truncated, returning the optimal feasible solution under the current optimization path. This allows the optimal edge activation combination of the entire graph to be obtained while relaxing the time limit requirements.

[0177] An alternative approach involves rewriting the candidate answer draft based on the final output state and the target evidence citation unit, including performing the following differential text rewriting mapping based on the parallel final output states:

[0178] If the final output state is fully supported, the corresponding statement text in the candidate answer draft is retained and the target evidence citation unit is appended with a digital index mark.

[0179] If the final output state is a partially supported state, then append a list of specific factual components of the missing evidence to the corresponding declaration text in the candidate answer draft.

[0180] If the final output state is a conflict state, the corresponding draft text will be rewritten and modified to indicate that the information source is inconsistent, and the target evidence reference units on both sides of the conflict will be listed respectively.

[0181] If the final output status is "rejected" or "insufficient permissions", the corresponding draft text will be rewritten and replaced with a preset neutral disclaimer text indicating insufficient evidence or restricted access.

[0182] After extracting the optimized graph state matrix, the text processing engine transforms the underlying topological logic into a structured natural language character stream based on the distribution of state labels.

[0183] In the first rewriting scenario, when the node is in a fully supported state, the engine inserts a hypertext anchor at the end of the corresponding segmented text of the candidate answer draft, binding the extracted target evidence index sequence.

[0184] In the second rewrite scenario, when a node is in a locally supported state, the engine reads the array of uncovered fact slices generated in the preceding verification phase, concatenates them, inserts them into the suffix area of ​​the original proposed text (i.e., the candidate answer draft), and applies a visual warning rendering style.

[0185] In the third rewrite scenario, when faced with a graph node marked as conflicting, the engine determines that the corresponding draft text has lost its output validity, performs an overwrite operation, and injects a pre-configured text inconsistency warning field in situ. Simultaneously, the engine extracts source file paragraphs from both the supporting and conflicting ends and outputs them below the warning field in a comparison view format.

[0186] In the fourth rewriting scenario, if a node has a rejection status or insufficient permissions status label, the text processing engine directly erases the original candidate statements, extracts the fields lacking search support or triggering unauthorized interception based on the status code, merges them into the main text stream, and outputs the final reliable question and answer result.

[0187] An alternative approach is to further include a front-end interactive presentation mechanism in the process of outputting credible question-answering results:

[0188] While outputting credible question-answering results, a lightweight visual subgraph of the declaration-evidence bipartite graph structure is generated and rendered, namely the visual subgraph corresponding to the active nodes and edges.

[0189] Interactive tracing links are established between nodes in the lightweight visualization subgraph and target evidence citation units to support user clicks that penetrate to the corresponding highlighted section in the original document.

[0190] After receiving the trusted question-and-answer result data packet, the terminal display device parses the bipartite graph structure snapshot file carried within. It extracts the nodes and topological edges marked as active, projects them into a force-directed network graph model, and outputs it in the sidebar area of ​​the main question-and-answer result view. For each numeric index mark in the main text stream, the front-end controller injects an interactive event listener process. When an activation event triggered by an input device is captured, it extracts the physical page number parameters and character offset coordinate range bound to the target evidence citation unit, sends a jump command to the document viewer, and drives the underlying drawing interface to perform background color rendering, forming an audit penetration link tracing the question-and-answer conclusion back to the original PDF or Word file.

[0191] Example 8

[0192] Based on the same inventive concept as the above embodiments, an offline security mechanism independent of the real-time online question-and-answer process is provided, and the construction of the offline encrypted audit chain and the pre-training process of the base multi-task model are described in detail.

[0193] In one possible implementation, the trusted knowledge question answering method also includes offline security auditing independent of the real-time process, referring to... Figure 5 Specifically, it may include the following steps:

[0194] Step 801: Extract user queries, access permissions, topological snapshots of the declaration-evidence bipartite graph structure, state evolution trajectories of iterative constraint propagation, and final output states during the question-and-answer execution process.

[0195] Specifically, it could also involve extracting user queries, access permissions, topological snapshots of the declaration-evidence bipartite graph structure, node state change records in each round of iterative constraint propagation as state evolution trajectories, and the final output state during the question-and-answer execution process.

[0196] After a single real-time knowledge question-answering task is completed, an asynchronous bypass recording process is triggered. This process detaches from the main computation flow and reads all intermediate state data of the question-answering cycle from the memory bus. To provide complete accountability, the user query text and the corresponding access permission identifier are extracted from the input source, and a topological snapshot of the declaration-evidence bipartite graph structure after the initial edge construction is captured. In addition, using the round as an index, the edge weight activation value change matrix and node state transition logs for each round of iterative constraint propagation computation are recorded to form a state evolution trajectory, and the final output state solidified after global optimization is also extracted.

[0197] Step 802: Serialize and package the extracted data.

[0198] After obtaining the multi-dimensional data objects distributed across different computing nodes, a pre-configured serialization protocol is used to flatten the data structure and convert it into a byte stream. During execution, the set of nodes, the set of edges, and their associated weight matrices in the graph structure are mapped into nested arrays in key-value pair format. The aforementioned topological snapshots, state evolution trajectories, and final output states are then concatenated according to their temporal sequence to generate a one-dimensional continuous byte stream data packet. This eliminates pointer dependencies of programming language entity objects in memory, ensuring the cross-platform parsability of the audit data.

[0199] Step 803: Calculate the secure hash and digital signature of the serialized and packaged data using a pre-configured encryption algorithm.

[0200] After acquiring the continuous byte stream data packets, tamper-proof encapsulation is performed. A metadata header containing the task identifier, processing timestamp, and model version number is injected into the packet header. A pre-configured one-way cryptographic hash function is called to calculate the packet's digest signature. After calculating the digest signature, the pre-allocated asymmetric encryption private key from the offline audit server is extracted, and the digest signature is encrypted to output a digital signature with anti-counterfeiting properties.

[0201] Step 804: Generate and store a replayable audit chain record with a digital signature for offline restoration and replay of the constraint reasoning judgment process of the statement-evidence bipartite graph structure.

[0202] The continuous byte stream data packets are merged and assembled with their corresponding digital signatures to generate an immutable block file. The persistent storage component writes this block file to a read-only storage medium, forming a replayable audit chain record. In the event of compliance reviews or Q&A quality disputes, third-party auditing devices can load this record file and verify the signature's legitimacy using the corresponding public key. Upon successful verification, the parsing program reverse-deserializes the original topology matrix and replays the state evolution trajectory frame by frame in a sandbox environment, transparently reproducing the entire logical deduction chain of the machine's determination of the final conclusion.

[0203] An alternative approach involves pre-training a semantic segmentation model, a natural language inference model for inference comparison, and a text generation model using a joint loss function. This pre-training process may include the following steps:

[0204] Step S1: Obtain a pre-built historical sample database, which includes manually labeled fact segmentation boundaries, real relationship labels, real state labels, and standard sorting labels.

[0205] In the system's underlying deployment phase, to ensure that multiple neural network components have a consistent feature distribution in the vector space, the model training unit initiates a joint pre-training pipeline. A pre-built historical sample database is loaded from the local storage system. This database contains a large amount of domain question-answering corpus and accompanying evidence documents. After removing samples with anomalous formats, manually labeled information attached to the text corpus is extracted. This labeling information includes factual segmentation boundary labels for supervising segmentation actions, true relation labels for supervising inference comparisons, true state labels indicating the expected global convergence, and standard ranking labels for supervising retrieval re-ranking scoring.

[0206] Step S2: Construct a multi-task joint loss function, which includes cross-entropy loss for association classification, boundary recognition loss, final state classification loss, and evidence reordering loss.

[0207] The training control unit integrates the processing model, natural language inference model, and generative model within the same computational graph, and constructs multi-dimensional error penalty terms. The error between the model's forward propagation predictions and the corresponding ground truth labels is calculated separately. The cross-entropy algorithm is used to calculate the cross-entropy loss for association classification and the final state classification loss. The contrastive learning algorithm is used to calculate the evidence reordering loss, and the boundary recognition loss is recorded. The various local loss terms are combined, and the calculation formula is as follows:

[0208] L total =w1*L rel +w2*L seg +w3*L state +w4*L rank ;

[0209] Among them, L total The total loss value is the output of the multi-task joint loss function, w1 is the weight coefficient of the association relationship, and L is the total loss value. rel The cross-entropy loss is used for classifying association relationships, w2 is the weighting coefficient for boundary identification, and L... seg For boundary recognition loss, w3 is the weight coefficient for state classification, and L... state The final state classification loss is given by w4, where w4 is the weighting coefficient for the ranking task, and L is the weighting coefficient for the ranking task. rank Loss due to evidence reordering.

[0210] In one optional implementation, w1 is set to 1.0, w2 to 0.8, w3 to 0.5, and w4 to 0.3. Those skilled in the art can determine appropriate weight ratios by observing the convergence curves of each subtask on the validation set and employing grid search or gradient-based multi-task weight adaptation methods.

[0211] Step S3: Perform backpropagation iterative calculation based on the historical sample database and the multi-task joint loss function to perform closed-loop optimization of the network parameters of each model, and obtain the pre-trained generative model, segmentation model and inference model for online use.

[0212] Those skilled in the art can use conventional model training methods in the field, such as the backpropagation algorithm combined with the Adam optimizer, to train the model. The hyperparameters during the training process, such as the learning rate, batch size, and number of iterations, can be determined by conventional hyperparameter tuning methods.

[0213] After obtaining the total loss value, the training control unit invokes the pre-configured optimizer engine. The partial derivative gradient of the total loss value with respect to the weight matrices of each network layer is calculated using the chain rule. Based on the gradient direction and the pre-configured learning rate step size, the internal neuron parameters of the processing model, natural language inference model, and generation model are updated synchronously. This batch calculation cycle is executed iteratively on all batches of data in the historical sample database. The iteration ends when the fluctuation of the total loss value on the validation set is detected to be below the halt threshold. The converged network parameters of each model are solidified, and frozen versions of the processing model, natural language inference model, and generation model are output and pushed to the production environment server for execution.

[0214] For example, in one test scenario, this method was applied to a knowledge base containing approximately 500 internal enterprise specification documents (including 12 sets of multi-version iteration documents), and evaluated on 150 test questions covering both single-dimensional and multi-dimensional composite fact queries. The results show that, compared to the baseline method that does not introduce bipartite graph constraint propagation, this method improves the fact consistency metric and reduces the misreference rate, especially in the test question set involving cross-version document conflicts, where the conflict detection capability shows a significant improvement. Those skilled in the art will understand that specific performance metrics may vary depending on the size of the knowledge base, the distribution of document types, and parameter configurations.

[0215] This application introduces a numerically sensitive clause-level difference comparison mechanism. By combining basic text difference with the forced adjustment judgment of discrete digital entities, the coarse overall document repeal is reduced to fine-grained clause-level filtering. This can preserve the historical evidentiary effect of unmodified clauses, while improving the system's sensitivity to fine-tuning of key indicators such as amount and standards.

[0216] This application adopts a multi-dimensional semantic component decomposition and evidence group accumulation verification rule. This mechanism allows fragmented information across documents to jointly support a single complex assertion, avoiding false rejection of overall legality due to missing local evidence fragments. It can also expose specific missing factual components to the front end, improving the transparency of the verification.

[0217] Furthermore, to eliminate logical deadlocks and contradictory statements caused by isolated statement validation, this application constructs a topologically-oriented declaration dependency graph and deeply integrates a bipartite graph constraint propagation algorithm. When a precondition fails, a contrapositive cascade blocking mechanism is automatically triggered. Combined with the underlying mixed-integer programming for optimal state solving in multi-source conflict environments, this achieves an evolution from local text matching to global macro-consistent reasoning, reducing the system's logical conflict output rate.

[0218] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases the steps shown or described may be executed in a different order than that shown here.

[0219] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0220] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A reliable knowledge question-answering method based on declaration-evidence bipartite graph constraint propagation, characterized in that, include: Retrieve user queries, corresponding access permissions, and a pre-stored collection of documents; Based on user queries and access permissions, the document set is retrieved and filtered to obtain a set of candidate evidence units. Text is generated based on user queries and a set of candidate evidence units to obtain draft candidate answers; Semantic segmentation of the candidate answer drafts yields a set of atomic declaration nodes; Based on the relationship between the set of atomic declaration nodes and the set of candidate evidence units, a declaration-evidence bipartite graph structure is constructed; Each atomic declaration node in the atomic declaration node set is compared with each candidate evidence unit in the candidate evidence unit set, and the corresponding support probability, conflict probability and indeterminable probability are output respectively. The maximum probability value among the three is extracted. When the maximum probability value reaches the preset gating threshold, a corresponding type of association relationship is established between the corresponding atomic declaration node and the candidate evidence unit. The association relationship includes support relationship, conflict relationship or indeterminate relationship. Iterative constraint propagation is performed on the declaration-evidence bipartite graph structure based on pre-configured verification constraint rules to obtain a converged declaration-credible verification state. The verification constraint rules include coverage, conflict, timeliness, and permission constraints. In the graph network, a loop propagation is performed. In each round of propagation, it is checked whether each atomic declaration node violates any pre-configured constraint criterion. Once an out-of-bounds trigger is found, the activation state of the corresponding edge in the bipartite graph is forcibly changed, and a monotonically negative transition of the declaration node state machine is triggered. The iterative propagation exhibits a cascading effect and oscillates within the graph structure. When no connected edge in the network experiences a state reversal and all node states remain static, the iterative computation stops, and the current static state is captured as the convergent, credible verification state. Based on the declaration-evidence bipartite graph structure, a global consistency optimization solution is performed on the declaration credibility verification state to obtain the final output state and the target evidence reference unit. Based on this, the candidate answer draft is rewritten and the credible question answer result is output. After obtaining the static graph cross-section of the iterative convergence, the state variables of each atomic declaration node and the edge activation variables of each association relationship are used as constraint variables to construct a 0-1 mixed integer linear programming model for solution. The comprehensive optimization objectives of the 0-1 mixed-integer linear programming model include: maximizing the weight scores of activated support relations, while minimizing the penalty weights of activated conflict relations, the baseline penalty term for the rejection state, and the penalty term for the missing local support states.

2. The method according to claim 1, characterized in that, Based on user queries and access permissions, the document set is retrieved and filtered to obtain a set of candidate evidence units, which is then implemented through the following comprehensive scoring mechanism: Semantic dense features and sparse keyword features are extracted from each document unit in the user query and document set, respectively; Semantic relevance is calculated based on dense semantic features, and keyword matching degree is calculated based on sparse keyword features. Calculate the time-related decay of document units based on the pre-configured decay period; Extract the coverage of user queries in the headings of the document unit's level to obtain the heading coverage; The semantic relevance, keyword matching degree, timeliness decay degree and title coverage are weighted and summed to obtain the comprehensive ranking score of each document unit, and the candidate evidence unit set is obtained based on this score.

3. The method according to claim 1, characterized in that, The document collection is searched and access-filtered using a multi-level dynamic gating mechanism, including: When acquiring the set of candidate evidence units, access permissions are used to perform preliminary admission filtering on the document set. When constructing the declaration-evidence bipartite graph structure, a secondary comparison and verification is performed using access permissions and the permission identifiers of each evidence unit in the candidate evidence unit set; If access permissions are detected to be inconsistent with the permission identifier of the target evidence unit, the association relationship corresponding to the target evidence unit will be forcibly severed in the declaration-evidence bipartite graph structure.

4. The method according to claim 1, characterized in that, After semantically segmenting the candidate answer drafts to obtain the set of atomic declaration nodes, the following is also included: Identify the logical dependencies between atomic declaration nodes in the set of atomic declaration nodes. Logical dependencies include prerequisite dependencies, mutual exclusion dependencies, and logical implication dependencies. Based on logical dependencies, a directed acyclic dependency graph with topological pointing is constructed within the set of atomic declaration nodes to provide conditional constraints in subsequent iterative constraint propagation.

5. The method according to claim 1, characterized in that, The coverage constraints in the verification constraint rules include mechanisms for independent support of a single piece of evidence, or mechanisms for joint evidence groups supported by multiple pieces of evidence.

6. The method according to claim 1, characterized in that, The timeliness constraints in the validation constraint rules include either a whole-document replacement control mechanism for the same source version or a clause-level replacement control mechanism for the same source version.

7. The method according to claim 1, characterized in that, The output of reliable question-and-answer results also includes a front-end interactive presentation mechanism: While outputting credible question-answering results, it generates and renders a lightweight visual subgraph of the statement-evidence bipartite graph structure; Interactive tracing links are established between nodes in the lightweight visualization subgraph and target evidence citation units to support user clicks that penetrate to the corresponding highlighted section in the original document.

8. The method according to claim 1, characterized in that, Before retrieving and filtering documents based on user queries and access permissions, query cache matching is also included: Extract the feature vector of the user query and perform similarity matching in the pre-stored historical query cache; If a high-confidence historical query is matched and the access permissions have not changed, the associated historical declaration-evidence bipartite graph structure snapshot is directly extracted and reused for local constraint incremental updates.

Images