Enterprise knowledge intelligent agent generation result verification method based on reflection checking mechanism

Abstract

The present application relates to the technical field of enterprise knowledge agent, and particularly relates to a method for verifying the generation result of enterprise knowledge agent based on a reflection verification mechanism, which obtains a response text to be verified and a knowledge retrieval record, decomposes the text into paragraphs, and reversely constructs a query for retrieval comparison to identify abnormal paragraphs, analyzes the context dependence of the abnormal paragraphs to determine the missing knowledge type, combines adjacent paragraphs to construct a bridging retrieval condition to obtain a supplementary knowledge segment and generate a verification comparison, replaces the abnormal paragraphs based on the verification comparison, extracts the abnormal paragraph distribution features and the bridging retrieval success mode to adjust the reverse retrieval strategy, thereby effectively verifying and correcting the generation result of the knowledge agent, and improving the accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of enterprise knowledge intelligence technology, and in particular to a method for verifying the generation results of enterprise knowledge intelligence based on a reflection and verification mechanism. Background Technology

[0002] In the field of verifying the results generated by knowledge agents, existing conventional practices mainly rely on static comparison mechanisms under the Retrieval Augmented Generation (RAG) framework. The usual approach is to pre-map user queries to an enterprise knowledge base, obtain several candidate knowledge fragments as the basis for generation, and then, after the response is generated, determine whether the generated content matches the retrieval results by simply calculating the lexical or semantic similarity between the response text and the knowledge fragments. Some solutions also introduce manually annotated verification rules to match and verify key entities or values ​​in the response. However, these methods are essentially post-hoc verifications, focusing on the compliance of local facts rather than examining the completeness of the generation logic chain.

[0003] Current conventional methods cannot effectively identify semantic gaps caused by leaps in reasoning or omissions in knowledge. When an agent introduces implicit inferences not included in the original retrieval results during the response process, static comparison may misjudge them as errors because it cannot find directly corresponding knowledge fragments, or conversely, ignore seemingly reasonable but actually unsupported inference paragraphs. This limitation makes it difficult for the verification process to distinguish between "reasonable reasoning" and "unfounded fabrication," especially in knowledge-intensive enterprise scenarios, where a conclusion may depend on a chain of multiple knowledge fragments, and conventional methods struggle to capture such dependencies.

[0004] Furthermore, existing conventional verification methods lack actionable remediation capabilities. Even when inconsistencies or anomalies are detected in the response, conventional methods typically only provide a binary conclusion of "pass" or "fail," or mark suspicious text, but cannot automatically infer the type and scope of missing knowledge, nor do they possess the ability to proactively retrieve and generate supplementary information. This reduces the verification process to an independent detection step, failing to form a closed loop with the generation process. Users still need to rely on manual intervention to correct errors, diminishing the practical value of automation for enterprise knowledge intelligence agents. Summary of the Invention

[0005] This invention provides a method for verifying the generation results of enterprise knowledge intelligent agents based on a reflection and verification mechanism, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for verifying the generation results of an enterprise knowledge intelligent agent based on a reflection verification mechanism, comprising:

[0007] Obtain the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments;

[0008] The response text to be verified is decomposed into text paragraphs that carry independent semantic units. Key terms and argumentation logic are extracted for each text paragraph. A verification query is constructed in reverse and an independent search is performed in the enterprise knowledge base. The reverse search results are aligned and compared with the knowledge search records to identify abnormal paragraphs.

[0009] For the identified anomalous paragraphs, the missing knowledge type and topic scope are determined by analyzing the contextual dependencies of the anomalous paragraphs in the response text to be verified;

[0010] Based on the knowledge type and topic range, bridging retrieval conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging retrieval is performed in the enterprise knowledge base to obtain supplementary knowledge fragments that can establish semantic coherence, and a verification comparison containing the original paragraph and the supplementary knowledge fragments is generated.

[0011] Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted.

[0012] The response text to be verified is decomposed into text paragraphs carrying independent semantic units. Key terms and logical arguments are extracted from each text paragraph. A verification query is constructed in reverse and an independent search is performed in the enterprise knowledge base. The reverse search results are aligned and compared with the knowledge retrieval records to identify anomalous paragraphs, including:

[0013] The semantic unit boundary is identified in the response text to be verified, where the subject-predicate-object structure is complete and all pronouns refer to objects within the current semantic unit. The semantic unit boundary is then segmented to obtain multiple independent text paragraphs.

[0014] For each independent text paragraph, the predicate structure is parsed to extract the core terms representing the action, and the modification dependency relationship is parsed to extract the logical terms representing the constraint conditions. Based on the dependency hierarchy of the core terms and the logical terms in the modification dependency relationship, a hierarchical query expression is constructed, and the hierarchical query expression sets the terms that are dependent on the preceding terms as the retrieval constraints of the subsequent terms.

[0015] The hierarchical query expression is submitted to the enterprise knowledge base to perform a retrieval and obtain a reverse retrieval knowledge set. The original knowledge set referenced when generating the independent text paragraph is extracted from the knowledge retrieval records.

[0016] Calculate the semantic coverage of the argument content in the independent text paragraph in the reverse retrieval knowledge set, perform intersection verification between the original knowledge set and the reverse retrieval knowledge set, and mark the independent text paragraph as an abnormal paragraph when the semantic coverage is lower than the coverage threshold.

[0017] Based on the dependency hierarchy of the core terms and logical terms in the modifying dependency relationship, a hierarchical query expression is constructed. This hierarchical query expression sets the terms of preceding dependencies as retrieval constraints for subsequent terms, including:

[0018] The semantic role labeling between verbs and nouns in the modification dependency relationship is analyzed to identify the action semantics carried by the core terms and the argument semantics carried by the logical terms. Based on the modification direction of the action semantics by the argument semantics, the order of each term in the dependency chain is determined.

[0019] The term located at the beginning of the dependency chain is used as the retrieval anchor to locate the knowledge fragment containing the term in the enterprise knowledge base, and the storage location of the knowledge fragment is recorded as the retrieval boundary.

[0020] A search instruction is generated for subsequent terms in the dependency chain. A range parameter pointing to the search boundary is embedded in the search instruction so that the search instruction searches for the subsequent terms only within the knowledge fragment defined by the search boundary, and obtains the intersection knowledge fragment that contains both preceding and following terms.

[0021] The storage location of the intersection knowledge fragment is updated to the new retrieval boundary. The next term in the dependency chain is used as the new subsequent term. The process of generating retrieval instructions and obtaining intersection knowledge fragments is repeated until all terms are processed. Then, the retrieval instructions are assembled into the hierarchical query expression according to the position of the dependency chain.

[0022] For the identified anomalous paragraphs, the contextual dependencies of the anomalous paragraphs within the response text to be verified are analyzed to determine the type and scope of missing knowledge, including:

[0023] Extract the state description of the subject entity from the preceding paragraph of the abnormal paragraph to construct the preceding state vector, extract the state description of the subject entity from the following paragraph to construct the following state vector, construct the state transition matrix based on the following state vector and the preceding state vector and perform singular value decomposition, and extract the left singular vector and right singular vector corresponding to the principal singular value.

[0024] The left singular vector is projected onto the semantic space of the preceding paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the preceding anchor point knowledge. The right singular vector is projected onto the semantic space of the following paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the following anchor point knowledge.

[0025] Retrieve original knowledge fragments containing the predecessor anchor knowledge and the successor anchor knowledge from the enterprise knowledge base, extract the association path connecting the two original knowledge fragments, count the distribution of relationship types of each node in the association path, and take the relationship type with the highest proportion as the missing knowledge type.

[0026] Calculate the rank of the state transition matrix, extract the corresponding number of principal singular vectors, map the principal singular vectors to the concept hierarchy structure of the enterprise knowledge base, and extract the common parent node of the mapped concepts as the topic scope.

[0027] Based on the aforementioned knowledge type and topic scope, bridging search conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging search is then performed in the enterprise knowledge base to obtain supplementary knowledge fragments that establish semantic coherence. A verification comparison containing the original paragraph and the supplementary knowledge fragments is generated, including:

[0028] Extract the core proposition of the preceding paragraph of the abnormal paragraph as the pre-proposition, extract the core proposition of the following paragraph as the post-proposition, parse the predicate structure of the pre-proposition and the post-proposition, and identify the missing argument roles between the two propositions;

[0029] Based on the knowledge type, determine the relational pattern required to fill the argument role, retrieve knowledge fragments containing the relational pattern from the knowledge subset defined by the topic scope, and convert the retrieved knowledge fragments into propositional expressions to construct an intermediate proposition set;

[0030] Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with the preceding proposition and the premise support relationship with the subsequent proposition. The knowledge fragments of the intermediate propositions that pass the verification are used as candidate bridging knowledge.

[0031] Calculate the logical tightness of each candidate bridging knowledge in the argument chain, select the candidate bridging knowledge with the highest logical tightness as a supplementary knowledge fragment, organize the abnormal paragraph and the supplementary knowledge fragment into a verification unit according to the argument order, mark the reasoning path of the antecedent proposition, the intermediate proposition and the consequent proposition in the verification unit, and generate a verification check containing logical dependencies.

[0032] Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with its preceding proposition and the premise support relationship with its subsequent proposition. The knowledge fragments of the verified intermediate propositions are used as candidate bridging knowledge, including:

[0033] The preceding propositions are formalized into preceding predicate expressions. Derivable atomic propositions are extracted from these preceding predicate expressions to form a forward reachable set. The following propositions are formalized into following predicate expressions. Required atomic propositions are extracted from these following predicate expressions to form a backward necessary set. For each intermediate proposition in the intermediate proposition set, it is formalized into an intermediate predicate expression.

[0034] The dependent variables of the intermediate predicate expression are parsed, and it is determined whether the dependent variables are fully bound by the forward reachable set. If they are fully bound, the atomic propositions of the forward reachable set are substituted into the dependent variables to perform variable substitution, and the truth value of the expression after substitution is calculated.

[0035] The free variables derived from the intermediate predicate expression are analyzed, and it is determined whether the free variables and the backward required set form an injective mapping. If an injective mapping is formed, the predicate expression, the intermediate predicate expression, and the postdicate expression are constructed into a derivation sequence in the order of derivation.

[0036] Logical resolution is performed on the derivation sequence to verify whether the resolution result is an empty clause. When the resolution result is an empty clause, the knowledge fragment corresponding to the intermediate proposition is taken as a candidate bridging knowledge.

[0037] Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted, including:

[0038] The propositional structure of the supplementary knowledge fragments in the verification comparison is analyzed, the core assertions are extracted, the core assertions are aligned with the abnormal paragraphs, the propositional components that are disproven by the core assertions are identified and removed, the core assertions are inserted into the corresponding logical positions according to the reasoning path marked by the verification comparison, the undisproven propositional components are kept, the verified response text is generated and output.

[0039] The source of the preceding propositions that each anomalous paragraph depends on is marked. Anomalous paragraphs that depend on the same source are classified into a homologous anomalous group. The argument elements that are commonly missing in the homologous anomalous group are extracted as distribution features.

[0040] The system parses the relational patterns and argument roles contained in the search conditions of successful bridging search cases, constructs a dependency rule base between the two, identifies dependency rules that appear repeatedly in multiple cases and marks them as successful patterns, selects corresponding dependency rules from the dependency rule base for the argument elements, and embeds the dependency rules into the dependency hierarchy of the hierarchical query expression.

[0041] A second aspect of this invention provides a verification system for the generation result of an enterprise knowledge intelligent agent based on a reflection verification mechanism, comprising:

[0042] The record acquisition unit is used to acquire the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments;

[0043] An anomaly detection unit is used to decompose the response text to be verified into text paragraphs that carry independent semantic units, extract key terms and argumentation logic for each text paragraph, reverse construct a verification query and perform an independent retrieval in the enterprise knowledge base, and align and compare the reverse retrieval results with the knowledge retrieval records to identify abnormal paragraphs.

[0044] The missing knowledge analysis unit is used to determine the type and scope of missing knowledge by analyzing the contextual dependencies of the identified anomalous paragraphs in the response text to be verified.

[0045] The bridging retrieval unit is used to construct bridging retrieval conditions based on the knowledge type and topic range, combined with the semantics of adjacent paragraphs of the abnormal paragraph, to perform bridging retrieval in the enterprise knowledge base, obtain supplementary knowledge fragments that can establish semantic coherence, and generate a verification comparison containing the original paragraph and the supplementary knowledge fragments.

[0046] The replacement output unit is used to replace the content of the abnormal paragraph based on the verification comparison, generate and output the verified response text, extract the distribution characteristics of the abnormal paragraph and the success pattern of the bridging retrieval, and adjust the query construction strategy of the reverse retrieval.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] This method, based on a reflective verification mechanism and reverse retrieval alignment, significantly improves the accuracy and reliability of the knowledge agent's generated results. By decomposing the response text to be verified into independent semantic paragraphs and extracting key terms and logical arguments to construct queries in reverse, it can systematically identify anomalous paragraphs inconsistent with knowledge retrieval records, effectively discovering factual errors, logical jumps, or fabricated content, fundamentally compensating for the cognitive blind spots of traditional single-search methods.

[0053] For anomalous paragraphs, by analyzing their contextual dependencies within the response text to be verified, the specific type and topic scope of missing knowledge are precisely defined, avoiding the generalization error of blindly completing the missing information. Based on this, bridging search conditions are constructed by combining the semantics of adjacent paragraphs to selectively acquire supplementary knowledge fragments that can establish semantic coherence, eliminating knowledge gaps between paragraphs. This ensures that the final verification response text maintains overall logical integrity and consistency with the facts, significantly reducing the risk of knowledge illusion.

[0054] By extracting the distribution characteristics of abnormal paragraphs and the successful patterns of bridging retrieval, the query construction strategy of reverse retrieval is dynamically adjusted to achieve continuous self-optimization of the verification mechanism. This adaptive learning capability not only improves the efficiency of single verification but also enables the system to automatically enhance its ability to identify and repair various knowledge gaps during iterations, providing long-term assurance for the highly reliable output of enterprise knowledge intelligence agents. It is suitable for accurate generation and verification in environments with massive and complex knowledge bases. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the method for verifying the generation results of an enterprise knowledge intelligent agent based on a reflection verification mechanism, according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating the method for constructing a verification query and performing a reverse retrieval in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1This is a flowchart illustrating the method for verifying the generation results of an enterprise knowledge intelligent agent based on a reflection verification mechanism, according to an embodiment of the present invention. The present invention provides a method for verifying the generation results of an enterprise knowledge intelligent agent based on a reflection verification mechanism, comprising:

[0060] Obtain the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments;

[0061] The response text to be verified is decomposed into text paragraphs that carry independent semantic units. Key terms and argumentation logic are extracted for each text paragraph. A verification query is constructed in reverse and an independent search is performed in the enterprise knowledge base. The reverse search results are aligned and compared with the knowledge search records to identify abnormal paragraphs.

[0062] For the identified anomalous paragraphs, the missing knowledge type and topic scope are determined by analyzing the contextual dependencies of the anomalous paragraphs in the response text to be verified;

[0063] Based on the knowledge type and topic range, bridging retrieval conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging retrieval is performed in the enterprise knowledge base to obtain supplementary knowledge fragments that can establish semantic coherence, and a verification comparison containing the original paragraph and the supplementary knowledge fragments is generated.

[0064] Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted.

[0065] Figure 2 This is a flowchart illustrating the method for constructing a verification query and performing a reverse retrieval in an embodiment of the present invention. The response text to be verified is decomposed into text paragraphs carrying independent semantic units. Key terms and argumentative logic are extracted from each text paragraph. A verification query is constructed in reverse and an independent retrieval is performed in the enterprise knowledge base. The reverse retrieval results are aligned and compared with the knowledge retrieval records to identify abnormal paragraphs, including:

[0066] The semantic unit boundary is identified in the response text to be verified, where the subject-predicate-object structure is complete and all pronouns refer to objects within the current semantic unit. The semantic unit boundary is then segmented to obtain multiple independent text paragraphs.

[0067] For each independent text paragraph, the predicate structure is parsed to extract the core terms representing the action, and the modification dependency relationship is parsed to extract the logical terms representing the constraint conditions. Based on the dependency hierarchy of the core terms and the logical terms in the modification dependency relationship, a hierarchical query expression is constructed, and the hierarchical query expression sets the terms that are dependent on the preceding terms as the retrieval constraints of the subsequent terms.

[0068] The hierarchical query expression is submitted to the enterprise knowledge base to perform a retrieval and obtain a reverse retrieval knowledge set. The original knowledge set referenced when generating the independent text paragraph is extracted from the knowledge retrieval records.

[0069] Calculate the semantic coverage of the argument content in the independent text paragraph in the reverse retrieval knowledge set, perform intersection verification between the original knowledge set and the reverse retrieval knowledge set, and mark the independent text paragraph as an abnormal paragraph when the semantic coverage is lower than the coverage threshold.

[0070] When performing semantic segmentation on the response text to be verified, it is necessary to accurately locate the boundaries of semantic units, rather than simply truncating them according to punctuation marks or fixed lengths. Specifically, dependency parsing is used to analyze the structure of each sentence in the response text to determine whether it has a complete subject-verb-object triad. For complex or coordinate sentences, it is further checked whether the pronouns within them have a clear referent within the current semantic unit. If a clause contains pronouns such as "it," "that," or "its," and the corresponding referent is located in another independent semantic unit, then the clause containing that pronoun cannot be used as a segmentation unit alone and needs to be merged with the semantic unit containing its referent. Only when the subject-verb-object structure is complete and all pronouns can find a referent within the current unit is semantic boundary segmentation performed at that point, thus obtaining multiple semantically consistent independent text paragraphs. This segmentation method ensures that when verifying each independent paragraph subsequently, the paragraph itself has complete semantic expressive power and the verification conclusion will not be distorted due to cross-paragraph dependencies in referential relationships.

[0071] When performing predicate structure parsing on each independent text paragraph, the focus is on extracting core terms that represent changes in action or state. Predicate structure parsing is accomplished by identifying verb phrases and their argument structures within the paragraph, treating verbs and their direct arguments (agent, patient, instrument, etc.) as the core term set. Simultaneously, modification dependencies within the paragraph are parsed, extracting modifiers such as attributives, adverbs, and complements attached to the core terms and categorizing them as logical terms representing constraints. For example, in the paragraph "conduct a compliance audit of financial data at the end of the quarter," "audit" is the core term, while "end of quarter" and "compliance" are logical terms, representing time constraints and type constraints, respectively.

[0072] Based on the hierarchical position of core terms and logical terms in modifying dependencies, a hierarchical query expression is constructed. The construction logic of the hierarchical query expression is: to set terms with higher dependency levels (i.e., closer to the root node of the syntax tree) as retrieval constraints for subsequent terms. Specifically, if logical terms... The hierarchy depth in the dependency tree is less than that of the core term. The depth of the hierarchy, As Preconditions appearing in the query expression form "in" Search under constraints The structured query format of "related knowledge" avoids the retrieval noise caused by spreading all terms into an unordered set of keywords, enabling retrieval requests to more accurately target knowledge fragments in the enterprise knowledge base that are consistent with the paragraph's logical argument.

[0073] The hierarchical query expression is submitted to the enterprise knowledge base for independent retrieval, resulting in a reverse-search knowledge set. Simultaneously, the original knowledge set referenced when generating the independent text paragraph is located from the knowledge retrieval records. These records, typically indexed by paragraph identifiers, record a list of knowledge fragments invoked by the knowledge agent when generating each paragraph, thus allowing for precise extraction of the original knowledge set corresponding to the current independent text paragraph.

[0074] After obtaining the reverse retrieval knowledge set and the original knowledge set, semantic coverage is used to quantify whether the argument of the current independent text paragraph can be supported by the reverse retrieval results. The semantic coverage is calculated by matching the semantic vector of the independent text paragraph with the semantic vectors of each knowledge segment in the reverse retrieval knowledge set, and selecting the paragraph with a similarity exceeding a preset similarity threshold. The proportion of paragraph semantics covered by a knowledge fragment is used as semantic coverage. .like Below the coverage threshold This indicates that the content of the independent text paragraph cannot be adequately supported by knowledge in the enterprise knowledge base, and there is a potential risk of knowledge illusion or logical overstepping.

[0075] Based on semantic coverage judgment, intersection verification must also be performed on the original knowledge set and the reverse-retrieval knowledge set. The purpose of intersection verification is to detect whether the original knowledge fragments cited by the knowledge agent when generating the paragraph overlap with the knowledge fragments independently retrieved through reverse query. If the size of the intersection of the two sets is significantly smaller than the size of the original knowledge set, it indicates that the knowledge agent cited knowledge fragments that cannot be reproduced by reverse retrieval when generating the paragraph. This means that there is a deviation between the original cited knowledge and the actual content of the paragraph, that is, the knowledge agent may have made inferences or extensions about the original knowledge beyond its actual content. The intersection verification result and semantic coverage together constitute the dual basis for anomaly judgment: only when semantic coverage is... Below the coverage threshold Only when this happens is the independent text paragraph marked as an abnormal paragraph, thus entering the subsequent context dependency analysis and bridging retrieval process.

[0076] Coverage threshold The settings need to be adjusted based on the domain characteristics and knowledge density of the enterprise knowledge base. In core business domains with relatively complete knowledge coverage, It can be set relatively high to ensure strict control over knowledge illusions; while in peripheral areas where the knowledge base coverage is relatively sparse, The threshold can be appropriately lowered to avoid a large number of normal paragraphs being mistakenly marked as abnormal due to the incompleteness of the knowledge base itself. This adaptive threshold strategy enables the entire verification process to maintain stable verification accuracy in business scenarios with different knowledge densities, while reducing the overhead of misjudgments caused by improper threshold settings.

[0077] Similarity threshold The settings also affect the accuracy of semantic coverage calculation. When When the setting is too low, knowledge fragments with weak semantic relevance will also be included in the coverage, resulting in an artificially high coverage and weakening the sensitivity to detect abnormal paragraphs; when When the threshold is set too high, knowledge fragments with different semantic expressions but equivalent content may be excluded from the coverage calculation, leading to misjudgment of normal paragraphs. In actual deployment, statistical analysis can be performed on the semantic coverage distribution of known normal paragraphs and known abnormal paragraphs in historical verification records, using the optimal segmentation point of the two distributions as... The reference values ​​are used to achieve a balance between detection sensitivity and false positive rate. The entire process of segmentation, extraction, query construction, retrieval comparison, and anomaly labeling constitutes the core identification path for enterprise knowledge intelligent agent response verification based on a reflective verification mechanism.

[0078] Based on the dependency hierarchy of the core terms and logical terms in the modifying dependency relationship, a hierarchical query expression is constructed. This hierarchical query expression sets the terms of preceding dependencies as retrieval constraints for subsequent terms, including:

[0079] The semantic role labeling between verbs and nouns in the modification dependency relationship is analyzed to identify the action semantics carried by the core terms and the argument semantics carried by the logical terms. Based on the modification direction of the action semantics by the argument semantics, the order of each term in the dependency chain is determined.

[0080] The term located at the beginning of the dependency chain is used as the retrieval anchor to locate the knowledge fragment containing the term in the enterprise knowledge base, and the storage location of the knowledge fragment is recorded as the retrieval boundary.

[0081] A search instruction is generated for subsequent terms in the dependency chain. A range parameter pointing to the search boundary is embedded in the search instruction so that the search instruction searches for the subsequent terms only within the knowledge fragment defined by the search boundary, and obtains the intersection knowledge fragment that contains both preceding and following terms.

[0082] The storage location of the intersection knowledge fragment is updated to the new retrieval boundary. The next term in the dependency chain is used as the new subsequent term. The process of generating retrieval instructions and obtaining intersection knowledge fragments is repeated until all terms are processed. Then, the retrieval instructions are assembled into the hierarchical query expression according to the position of the dependency chain.

[0083] After decomposing the response text into paragraphs and extracting core and logical terms, the modification dependencies between these terms need to be transformed into hierarchical query expressions to ensure the accuracy and semantic integrity of the reverse retrieval. The construction of hierarchical query expressions is based on modification dependencies, using nested retrieval constraints to ensure that the retrieval scope of each subsequent term is limited by the retrieval results of the preceding terms, thus avoiding semantic drift caused by isolated term matching.

[0084] The first step in constructing a hierarchical query expression is to parse the semantic role annotations between verbs and nouns in the dependency relationships. For each text paragraph, the dependency tree structure is extracted, identifying the verb nodes that act as predicates and the nominal argument nodes distributed around those predicates. Core terms carry the action semantics of the paragraph, that is, the meaning of the predicate describing an operation, state change, or causal relationship; logical terms carry the argument semantics, including the constraint meanings in the direction of modification such as agent, patient, instrument, condition, and scope. By analyzing the direction of modification of action semantics by argument semantics, the order of each term in the dependency chain can be determined. For example, if a logical term acts as a conditional adverbial modifying the action expressed by a core term, then this logical term is located before the core term in the dependency chain and should be given priority as a retrieval constraint to anchor the knowledge scope. If a logical term acts as a result adverbial following a core term, then it is located after the core term in the dependency chain and needs further filtering based on the retrieval results of the core term. The order of the dependency chain determines the nesting level of each term in the hierarchical query expression; the closer a term is to the beginning of the dependency chain, the stronger its retrieval constraint effect and the wider its coverage.

[0085] After determining the dependency chain positions of each term, the term at the beginning of the dependency chain is used as the retrieval anchor. An initial search is performed in the enterprise knowledge base to locate all knowledge fragments containing the anchor term. The selection of the retrieval anchor follows the principle of maximizing semantic coverage; the starting term is usually the one with the broadest semantics and the widest distribution in the knowledge base. This ensures that relevant knowledge is not missed during subsequent constraint filtering due to an insufficiently small initial result set. The set of storage locations of the knowledge fragments hit by the initial search is recorded as the first-level retrieval boundary. The retrieval boundary can be physically represented as a set of unique identifiers for knowledge fragments, a subspace range in a vector index, or a partition identifier in document storage. Its core function is to shrink the search space of subsequent searches from the entire enterprise knowledge base to the set of hit fragments, thereby achieving progressive constraint filtering.

[0086] When generating a search instruction for a subsequent term in the second layer of a dependency chain, a range parameter pointing to the first-layer search boundary is embedded in the instruction. This range parameter forces the search engine to search for the subsequent term only within the set of knowledge fragments defined by the first-layer search boundary, rather than searching independently across the entire database. The result obtained in this way is a knowledge fragment containing the intersection of the starting and subsequent terms. This intersection operation ensures that the search results semantically satisfy the constraints of both terms in the dependency chain, excluding irrelevant fragments containing only a single term but lacking semantic connection. For example, in a corporate knowledge base, if the starting term of the dependency chain is "approval" and the second-layer term is "time limit," the second-layer search instruction will only search for "time limit" within the knowledge fragments containing "approval." Ultimately, it retrieves a knowledge fragment that describes both the approval action and its time limit constraints, rather than all documents that broadly relate to time limits.

[0087] After obtaining the intersection knowledge fragments, the storage location of the intersection result is updated to the new retrieval boundary, and the next term in the dependency chain is taken as the new subsequent term. The process of generating retrieval instructions and obtaining intersection knowledge fragments is repeated. Each iteration further shrinks the search space based on the retrieval boundary of the previous iteration, forming a nested constraint structure. The iteration process continues until all terms in the dependency chain have been processed. If the set of intersection knowledge fragments is empty in a certain iteration, it indicates that there are no knowledge fragments in the enterprise knowledge base that simultaneously satisfy all the preconditions of the current dependency chain. At this time, it is necessary to backtrack to the previous retrieval boundary and try to relax the matching conditions of the current term, such as switching from exact matching to semantic similarity matching, to ensure that the hierarchical query expression still has a certain robustness when facing incomplete knowledge base coverage. When relaxing the matching conditions, the similarity matching is determined based on the cosine similarity of semantic vectors to ensure consistency with the dependency chain constraints at the semantic level.

[0088] After all terms have been processed, the retrieval instructions at each level are assembled into a complete hierarchical query expression according to their position in the dependency chain. During assembly, each retrieval instruction carries its corresponding retrieval boundary parameters, forming a nested structure from broad to narrow, from pre-constraints to post-constraints. Structurally, the hierarchical query expression is similar to a context-defined compound query, where the outer instruction provides semantic anchoring, and the inner instruction performs fine-grained filtering within the anchored scope. This assembly method ensures that the hierarchical query expression not only records the retrieval intent of each term but also preserves the dependency constraints between terms, providing structured query semantics for subsequent reverse retrieval in the enterprise knowledge base.

[0089] After the hierarchical query expression is generated, it is used to perform a reverse retrieval in the enterprise knowledge base, and the retrieval results are aligned and compared with the original knowledge retrieval records. Because the hierarchical query expression makes full use of the modification dependencies between terms, its retrieval accuracy is significantly higher than that of flat queries based on single terms. It can more accurately locate knowledge fragments that are highly consistent with the semantics of the text paragraphs, thus providing more reliable knowledge support for subsequent abnormal paragraph identification and bridging retrieval.

[0090] For the identified anomalous paragraphs, the contextual dependencies of the anomalous paragraphs within the response text to be verified are analyzed to determine the type and scope of missing knowledge, including:

[0091] Extract the state description of the subject entity from the preceding paragraph of the abnormal paragraph to construct the preceding state vector, extract the state description of the subject entity from the following paragraph to construct the following state vector, construct the state transition matrix based on the following state vector and the preceding state vector and perform singular value decomposition, and extract the left singular vector and right singular vector corresponding to the principal singular value.

[0092] The left singular vector is projected onto the semantic space of the preceding paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the preceding anchor point knowledge. The right singular vector is projected onto the semantic space of the following paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the following anchor point knowledge.

[0093] Retrieve original knowledge fragments containing the predecessor anchor knowledge and the successor anchor knowledge from the enterprise knowledge base, extract the association path connecting the two original knowledge fragments, count the distribution of relationship types of each node in the association path, and take the relationship type with the highest proportion as the missing knowledge type.

[0094] Calculate the rank of the state transition matrix, extract the corresponding number of principal singular vectors, map the principal singular vectors to the concept hierarchy structure of the enterprise knowledge base, and extract the common parent node of the mapped concepts as the topic scope.

[0095] After identifying the anomalous paragraph, it's crucial to analyze its semantic connection function within the overall response text to accurately determine the missing knowledge and its subject matter. Anomalous paragraphs are not isolated; they logically connect to preceding and following paragraphs. Preceding paragraphs describe the state of a specific entity at a particular stage, while subsequent paragraphs describe its evolved state. The anomalous paragraph should ideally serve as a narrative of this state transition. The lack of knowledge support in the anomalous paragraph creates a semantic gap between its preceding and following sections. Therefore, by structurally modeling the state descriptions of the preceding and following paragraphs, the missing knowledge outline of the anomalous paragraph can be effectively reconstructed.

[0096] Extract the state description of the subject entity from the preceding paragraph of the anomalous paragraph, and encode the state description into a vector form to obtain the preceding state vector. Extract the state description of the same topic entity from subsequent paragraphs and encode it as a subsequent state vector. The "state description" referred to here refers to semantic segments in a paragraph that explicitly depict the attributes, relationships, or behavioral outcomes of the subject entity, such as the entity's current conditions, constraints, and operational results. During extraction, dependency parsing is used to locate the predicate-argument structure. The direct modifiers and argument roles of the entity are collected, and these components are mapped to a high-dimensional vector space using a pre-trained semantic encoding model, forming... and .

[0097] Based on the predecessor state vector With successor state vector Construct the state transition matrix Its construction method is to and The outer product expansion, i.e. This matrix encodes the directional change information from the predecessor state to the successor state. Perform singular value decomposition to obtain ,in It is a left singular vector matrix. It is a diagonal singular value matrix. This is a right singular vector matrix. Extract the maximum singular value. The corresponding left singular vector With right singular vectors These two vectors represent the most significant directions of state change in the successor and predecessor spaces, respectively, and carry the semantic transition direction information that should be filled in for the abnormal paragraph.

[0098] left singular vector Projecting onto the semantic space of the preceding paragraph involves specifically calculating... The cosine similarity between the semantic vectors of the knowledge segments and the preceding paragraphs is used to identify the knowledge segment with the highest similarity as the preceding anchor knowledge. This anchor knowledge represents the knowledge node most relevant to the state transition in the state described by the preceding paragraph and serves as the starting point for subsequent bridging retrieval. Similarly, the right singular vectors are... Projected onto the semantic space of subsequent paragraphs, the knowledge fragment with the highest similarity is identified as the subsequent anchor knowledge. This anchor knowledge represents the knowledge node most relevant to the state transition in the state described by the subsequent paragraph and serves as the target anchor point for bridging retrieval. The two anchor knowledge points together define the semantic boundary of the missing knowledge in the anomalous paragraph.

[0099] In the knowledge graph structure of the enterprise knowledge base, the system retrieves original knowledge fragments that simultaneously contain entity nodes corresponding to both predecessor and successor anchor knowledge. The knowledge in the knowledge base is organized in a graph structure, where nodes represent knowledge entities or concepts, and edges represent the types of relationships between entities. These relationship types encompass various semantic categories such as causal relationships, dependency relationships, transformation relationships, and conditional relationships. A graph path search algorithm extracts all associated paths connecting the nodes corresponding to predecessor and successor anchor knowledge, and the distribution of relationship types carried by each edge in these paths is statistically analyzed. The relationship type with the highest frequency across all associated paths is identified as the missing knowledge type in the anomalous segment. For example, if the "transformation relationship" edge has the highest proportion in an associated path, the missing knowledge type describes the entity's state transformation process; if the "conditional relationship" edge has the highest proportion, the missing knowledge type describes the triggering condition. This method of determination based on path relationship distribution objectively reflects the semantic leap between preceding and succeeding states at the structural level.

[0100] While determining the type of missing knowledge, it is also necessary to determine the topic scope to which the missing knowledge belongs, so as to conduct targeted scope-limited searches in the enterprise knowledge base. Calculate the state transition matrix. rank The rank reflects the dimensional complexity of the semantic change from the predecessor state to the successor state. Extract the corresponding quantity. One principal singular vector, i.e., taking the singular value matrix. Center front The set of left singular vectors corresponding to each non-zero singular value This Each singular vector is mapped to a concept hierarchy in the enterprise knowledge base. The mapping method is to calculate the similarity between each singular vector and the semantic vector of each concept node in the concept hierarchy, and take the concept node with the highest similarity as the mapped concept of the singular vector.

[0101] The concept hierarchy of the enterprise knowledge base is organized using a tree or directed acyclic graph format. Each concept node has a clear parent node, which represents a more macro-level knowledge topic category. Extraction The common parent node of a mapping concept in the concept hierarchy represents the knowledge topic of the missing knowledge. When When mapping concepts are distributed across different branches of the concept hierarchy, the common parent node is likely located at a higher level, indicating that the missing knowledge covers a broad range of topics. Conversely, when mapping concepts are concentrated in the same branch, the common parent node is located at a lower level, indicating that the missing knowledge covers a more focused range of topics. This hierarchical tracing mechanism ensures that the topic range is neither too broad, leading to excessive noise in subsequent searches, nor too narrow, resulting in the omission of relevant knowledge.

[0102] Through the above process, a clear missing knowledge type and topic scope are ultimately determined for each anomalous paragraph. These two factors together constitute the input constraints for subsequent bridging retrieval. The knowledge type limits the semantic relational nature of the retrieval target, while the topic scope limits the location range of the retrieval target in the knowledge base's conceptual space. The superposition of these two constraints can significantly improve the accuracy of bridging retrieval and lay a structured foundation for the subsequent generation of semantically coherent supplementary knowledge fragments.

[0103] Based on the aforementioned knowledge type and topic scope, bridging search conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging search is then performed in the enterprise knowledge base to obtain supplementary knowledge fragments that establish semantic coherence. A verification comparison containing the original paragraph and the supplementary knowledge fragments is generated, including:

[0104] Extract the core proposition of the preceding paragraph of the abnormal paragraph as the pre-proposition, extract the core proposition of the following paragraph as the post-proposition, parse the predicate structure of the pre-proposition and the post-proposition, and identify the missing argument roles between the two propositions;

[0105] Based on the knowledge type, determine the relational pattern required to fill the argument role, retrieve knowledge fragments containing the relational pattern from the knowledge subset defined by the topic scope, and convert the retrieved knowledge fragments into propositional expressions to construct an intermediate proposition set;

[0106] Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with the preceding proposition and the premise support relationship with the subsequent proposition. The knowledge fragments of the intermediate propositions that pass the verification are used as candidate bridging knowledge.

[0107] Calculate the logical tightness of each candidate bridging knowledge in the argument chain, select the candidate bridging knowledge with the highest logical tightness as a supplementary knowledge fragment, organize the abnormal paragraph and the supplementary knowledge fragment into a verification unit according to the argument order, mark the reasoning path of the antecedent proposition, the intermediate proposition and the consequent proposition in the verification unit, and generate a verification check containing logical dependencies.

[0108] After identifying anomalous paragraphs and determining their knowledge type and topic scope, a bridging retrieval mechanism is needed to supplement these paragraphs with semantically coherent knowledge. The core idea of ​​bridging retrieval is that the preceding and following text paragraphs already contain specific argumentative propositions, but the anomalous paragraph itself fails to provide sufficient knowledge support, resulting in an argumentative gap between the preceding and following propositions. Therefore, bridging retrieval uses the core propositions of the preceding and following paragraphs as anchors to identify the missing argumentative roles between them, and then specifically retrieves knowledge fragments from the enterprise knowledge base that can fill this argumentative gap.

[0109] When extracting the core propositions of preceding paragraphs, predicate-argument structure analysis is performed on the text of the preceding paragraphs to identify the predicates expressing the main arguments and their corresponding agent, patient, and target argument roles. The complete predicate framework and its argument filling are recorded as preceding propositions. Preceding propositions typically express facts or inferences that a certain line of reasoning has already been established at the current position, such as "a certain technical solution has been adopted" or "a certain type of risk has been identified." Similarly, the same predicate-argument structure analysis is performed on subsequent paragraphs to extract the core propositions of the subsequent paragraphs as following propositions. Following propositions typically represent conclusions or action recommendations that the line of reasoning needs to be established in subsequent positions, such as "a certain improvement measure should be implemented."

[0110] After obtaining the predicate structures of the antecedent and consequent propositions, a comparative analysis of the argument roles of the two propositions is performed. Specifically, it checks whether the agent, patient, or conditional arguments appearing in the consequent proposition have been fully defined or derived in the antecedent proposition. If there are argument roles that the consequent proposition depends on but are not provided by the antecedent proposition, then the missing argument role is recorded as a semantic gap that needs to be filled through bridging knowledge. For example, if the patient argument of the antecedent proposition is "data quality problem" and the agent argument of the consequent proposition is "data governance team," and there is a lack of a causal or authorization relationship between the two—"the data governance team is authorized to handle data quality problems"—then this relationship type is the missing argument role that needs to be identified.

[0111] Based on the identified knowledge types, missing argument roles are mapped to specific relational schemas. Knowledge types may encompass various categories such as causal relationships, hierarchical relationships, process dependencies, and authority / responsibility authorization relationships, with different knowledge types corresponding to different relational schema templates. For example, if the knowledge type is "process dependency," the relational schema is expressed as "the output of the preceding action is the input condition of the subsequent action"; if the knowledge type is "authority / responsibility authorization," the relational schema is expressed as "subject A has operational authority over object B under condition C." Within a knowledge subset defined by the topic scope, structured semantic retrieval is performed using the relational schemas as search criteria to filter out a candidate set of knowledge fragments containing the aforementioned relational schemas.

[0112] Each knowledge fragment in the retrieved candidate knowledge fragment set is converted into a propositional expression, forming an intermediate proposition set. The propositional expression requires each knowledge fragment to be represented in a predicate-argument structure, ensuring that subsequent verification of inference relationships can be performed at a unified logical representation level. During the conversion process, predicate extraction and argument annotation are performed on the natural language expression of the knowledge fragments, making the implicit logical relationships explicit. For example, "System A is responsible for processing type B data" is converted into the propositional form "process (Agent: System A, Patient: Type B data)".

[0113] For each intermediate proposition in the intermediate proposition set, bidirectional reasoning verification is performed. In forward verification, it is determined whether the preceding and intermediate propositions, together, can deduce the subsequent proposition; that is, verifying whether the conclusion derivation relationship "preceding proposition ∧ intermediate proposition → subsequent proposition" holds. In backward verification, it is determined whether the intermediate proposition is one of the necessary prerequisites for the subsequent proposition to hold; that is, verifying the prerequisite support relationship between the intermediate proposition and the subsequent proposition. The verification method combines semantic implication judgment and knowledge graph path reasoning: semantic implication judgment is achieved by encoding the semantic vectors of each proposition and calculating their implication probabilities; knowledge graph path reasoning is achieved by finding the relationship path between the preceding and subsequent proposition entities in the enterprise knowledge graph through the intermediate proposition entity. Only when an intermediate proposition passes both forward and backward verification is the corresponding knowledge fragment included in the candidate bridging knowledge set.

[0114] Calculate the logical tightness of each candidate bridging knowledge in the argument chain. Logical tightness comprehensively measures the following three dimensions: First, the semantic connection between the intermediate proposition and its preceding proposition, i.e., the semantic similarity between the agent or conditional argument of the intermediate proposition and the conclusion argument of the preceding proposition; second, the semantic connection between the intermediate proposition and its following proposition, i.e., the semantic similarity between the conclusion argument of the intermediate proposition and the premise argument of the following proposition; third, the credibility score of the knowledge fragment corresponding to the intermediate proposition in the enterprise knowledge base, which is calculated by comprehensively considering the source authority, update timeliness, and citation frequency of the knowledge fragment. The scores of the above three dimensions are combined using a weighted summation method to obtain the total logical tightness score, and the candidate bridging knowledge with the highest logical tightness is selected as the final supplementary knowledge fragment.

[0115] After identifying the supplementary knowledge fragments, the anomalous paragraphs and the supplementary knowledge fragments are organized into verification units according to the order of argumentation. The internal structure of the verification unit follows the reasoning path of "preceding proposition → intermediate proposition → subsequent proposition," where the preceding proposition corresponds to the core proposition of the preceding paragraph, the intermediate proposition corresponds to the propositional expression derived from the supplementary knowledge fragment, and the subsequent proposition corresponds to the core proposition of the following paragraph. The original content of the anomalous paragraph is retained in the verification unit as the area to be replaced, presented alongside the supplementary knowledge fragments, so that the two can be compared and analyzed in subsequent verification steps.

[0116] In the verification unit, the reasoning paths between predicates, intermediate propositions, and postdicates are explicitly labeled. The labeling includes the type of reasoning relationship between each pair of adjacent propositions (e.g., causal deduction, conditional implication, inductive generalization), the confidence score of the reasoning relationship, and the source identifier of the knowledge fragment supporting that reasoning relationship. Through this labeling, the verification check not only includes a content comparison between the original paragraph and the supplementary knowledge fragment, but also fully records the structural information of the logical dependencies, enabling subsequent content replacement and verification evaluation to be conducted within a verifiable logical framework. The generated verification check containing logical dependencies provides a complete structured basis for subsequent content replacement of abnormal paragraphs and the generation of post-verification response text.

[0117] Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with its preceding proposition and the premise support relationship with its subsequent proposition. The knowledge fragments of the verified intermediate propositions are used as candidate bridging knowledge, including:

[0118] The preceding proposition is formalized into a preceding predicate expression, and the deducible atomic propositions are extracted from the preceding predicate expression to form a forward reachable set. The following proposition is formalized into a following predicate expression, and the necessary atomic propositions are extracted from the following predicate expression to form a backward necessary set.

[0119] For each intermediate proposition in the set of intermediate propositions, it is formalized into an intermediate predicate expression;

[0120] The dependent variables of the intermediate predicate expression are parsed, and it is determined whether the dependent variables are fully bound by the forward reachable set. If they are fully bound, the atomic propositions of the forward reachable set are substituted into the dependent variables to perform variable substitution, and the truth value of the expression after substitution is calculated.

[0121] The free variables derived from the intermediate predicate expression are analyzed, and it is determined whether the free variables and the backward required set form an injective mapping. If an injective mapping is formed, the predicate expression, the intermediate predicate expression, and the postdicate expression are constructed into a derivation sequence in the order of derivation.

[0122] Logical resolution is performed on the derivation sequence to verify whether the resolution result is an empty clause. When the resolution result is an empty clause, the knowledge fragment corresponding to the intermediate proposition is taken as a candidate bridging knowledge.

[0123] After completing the context dependency analysis of the anomalous paragraphs and determining the type of missing knowledge, rigorous logical validity verification of the candidate bridging knowledge is required. This ensures that the supplementary knowledge fragments ultimately selected for verification are not only semantically related to the anomalous paragraphs but also truly fill the reasoning gaps between the preceding and following propositions at the logical deduction level. To this end, the preceding semantic content of the anomalous paragraphs is abstracted into pre-propositions, the subsequent semantic content into post-propositions, and the semantic content corresponding to the knowledge fragments returned by the bridging retrieval is abstracted into a set of intermediate propositions. Based on this, a formal logical verification process is executed.

[0124] When formalizing a predicate proposition into a predicate expression, it is necessary to identify the entities, attributes, and relationships involved in the predicate proposition and transform them into the standard form of first-order predicate logic. For example, if the predicate proposition describes that the quality inspection status of a product meets a specific standard, the formalized predicate expression contains constants representing product entities, predicates representing quality attributes, and logical assertions representing the inspection conclusion. Starting from this predicate expression, a forward reachable set is constructed by exhaustively enumerating all atomic propositions that can be directly derived from it. Each atomic proposition in the forward reachable set represents the smallest semantic unit that can be established from the predicate proposition through a single step or a finite number of steps of derivation. These atomic propositions collectively characterize the "reachable boundary" of the predicate proposition in the reasoning space.

[0125] Similarly, the post-proposition is formalized into a post-predicate expression, and the atomic propositions necessary for the post-proposition to be true are extracted to construct a backward-necessary set. Each atomic proposition in the backward-necessary set represents a necessary premise for the post-proposition to be true; without any one of these atomic propositions, the post-proposition cannot be logically supported. Constructing the backward-necessary set is essentially performing a reverse analysis on the post-predicate expression, breaking down its conclusion into a set of premises with the smallest granularity.

[0126] For each intermediate proposition in the set of intermediate propositions, it is also formalized into an intermediate predicate expression. The dependent variables contained in this intermediate predicate expression are parsed; that is, the variables that need to be externally bound for the intermediate proposition to be true. It is determined whether these dependent variables can be fully bound by atomic propositions in the forward reachability set; that is, for each dependent variable, there exists at least one atomic proposition in the forward reachability set that can provide a valid assignment. If any dependent variable cannot be bound from the forward reachability set, it means that the preceding proposition cannot provide sufficient derivation support for the intermediate proposition, and the intermediate proposition does not satisfy the conclusion derivation relationship with the preceding proposition, and is directly excluded.

[0127] When the dependent variable is fully bound to the forward reachable set, the corresponding atomic proposition in the forward reachable set is substituted into the dependent variable of the intermediate predicate expression, and a variable substitution operation is performed to obtain a closed predicate expression without any free dependent variables. The truth value of this closed expression is then evaluated to verify whether it is true within the logical context established by the preceding proposition. If the truth value is false, it means that even if the reachable atomic proposition of the preceding proposition can provide an assignment for the dependent variable, the intermediate proposition is still logically incompatible with the preceding proposition and is therefore excluded.

[0128] Intermediate propositions that pass forward verification must then undergo backward verification. This involves analyzing the free variables derived from the intermediate predicate expression after parsing and substitution—the semantic information carriers that can be passed forward after the intermediate propositions are true. It is then determined whether these free variables form an injective mapping with the required set of subsequent propositions. This means that each free variable derived from the intermediate proposition can establish a one-to-one semantic binding relationship with an atomic proposition in the required set of subsequent propositions, and there is no situation where an atomic proposition in the required set cannot be covered by any free variable. The injective mapping requirement ensures that the intermediate propositions not only pass on sufficient semantic content but also that the passed content precisely corresponds to the necessary premises of the subsequent propositions, without any logical ambiguity caused by semantic redundancy.

[0129] After the injective mapping verification passes, the predicate expressions, intermediate predicate expressions, and postdicate expressions are arranged sequentially according to the derivation order to construct a complete derivation sequence. This derivation sequence formally constitutes a proposition chain, where each node is a formalized predicate expression, and the directed connections between nodes represent logical derivation relations. Logical resolution is then performed on this derivation sequence. Logical resolution is a standard method in first-order logic for verifying the consistency of a proposition set. Its core idea is to incorporate the negation forms of each proposition in the derivation sequence along with the original propositions into the resolution process. By repeatedly applying resolution rules to eliminate complementary literal pairs, if an empty clause can ultimately be derived, it proves that the derivation sequence is logically self-consistent and complete; that is, the predicate proposition can indeed logically derive the postdicate proposition through the intermediate proposition.

[0130] During the resolution process, each predicate expression needs to be transformed into clause normal form, i.e., a set of clauses under conjunctive normal form. For any two clauses in the clause set, if there is a complementary literal pair, the resolution operation is performed on them, generating a new clause and adding it to the clause set. This process is repeated until an empty clause is generated or the clause set stops growing. When the resolution result is an empty clause, it is confirmed that the knowledge fragment corresponding to the intermediate proposition satisfies the complete logical bridging condition, and it is marked as candidate bridging knowledge and included in the subsequent verification and comparison construction process. If the resolution process fails to generate an empty clause, it indicates that there is a logical gap in the derivation sequence, and the intermediate proposition cannot completely connect the preceding and following propositions at the formal logic level, and is therefore excluded.

[0131] The candidate bridging knowledge selected through the formal verification process described above possesses strict logical validity guarantees, ensuring that the final verification comparison is not only semantically coherent at the content level but also has a complete reasoning chain. This mechanism is particularly important in enterprise knowledge agent application scenarios. For example, in scenarios involving compliance judgment, process approval, or technical solution recommendation, where the rigor of reasoning is highly demanding, the logical validity of the candidate bridging knowledge directly determines the credibility and usability of the verified response text. Candidate bridging knowledge that has passed logical deduction verification will serve as high-quality supplementary knowledge fragments in subsequent content replacement and response text reconstruction.

[0132] Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted, including:

[0133] The propositional structure of the supplementary knowledge fragments in the verification comparison is analyzed, the core assertions are extracted, the core assertions are aligned with the abnormal paragraphs, the propositional components that are disproven by the core assertions are identified and removed, the core assertions are inserted into the corresponding logical positions according to the reasoning path marked by the verification comparison, the undisproven propositional components are kept, the verified response text is generated and output.

[0134] The source of the preceding propositions that each anomalous paragraph depends on is marked. Anomalous paragraphs that depend on the same source are classified into a homologous anomalous group. The argument elements that are commonly missing in the homologous anomalous group are extracted as distribution features.

[0135] The system parses the relational patterns and argument roles contained in the search conditions of successful bridging search cases, constructs a dependency rule base between the two, identifies dependency rules that appear repeatedly in multiple cases and marks them as successful patterns, selects corresponding dependency rules from the dependency rule base for the argument elements, and embeds the dependency rules into the dependency hierarchy of the hierarchical query expression.

[0136] After completing the bridging retrieval and obtaining the supplementary knowledge fragments, precise content replacement needs to be performed on the abnormal paragraphs, rather than simple text overwriting. The core of the replacement process lies in the deep analysis of the propositional structure of the supplementary knowledge fragments in the verification comparison. Propositional structure analysis is based on semantic role labeling in natural language processing, identifying the subject, predicate, object, and additional arguments expressed by the supplementary knowledge fragments, and extracting the core assertion—that is, the smallest complete unit of fact conveyed by the supplementary knowledge fragment. The core assertion usually exists in the form of a "subject-relation-object" triple, such as a structured factual expression like "a certain process node-dependency-specific approval authority". After extracting the core assertion, it is aligned with the original propositions in the abnormal paragraph one by one. The alignment process determines the correspondence by calculating the degree of overlap between the two at the predicate framework level, thereby determining which original propositional components directly conflict or logically contradict the core assertion.

[0137] Propositional components disproved by the core assertion are represented in the alignment results as statements that contradict the facts established by the supplementary knowledge fragment. For example, if the original paragraph claims that a certain approval process does not require secondary review, while the supplementary knowledge fragment explicitly states that the process must undergo secondary review, then the negative statement in the original paragraph is the disproved propositional component and should be removed. Removal operations must be precise to the propositional boundary to avoid accidentally deleting other statements that are grammatically adjacent to but semantically independent of the disproved component. After removing the disproved components, the core assertion is inserted into the corresponding logical position according to the reasoning path marked in the verification check. The reasoning path markings record the argumentation level of the supplementary knowledge fragment in the original knowledge base and its causal or supporting relationships with surrounding knowledge nodes. Based on this path information, the core assertion is placed in a position that maintains the coherence of the overall argumentation chain of the paragraph. Undisproven propositional components are retained as is, and together with the inserted core assertion, they constitute the verified paragraph content, which is then assembled into a complete verified response text and output.

[0138] After content replacement and output are completed, the distribution feature extraction stage begins. Each anomalous paragraph is labeled with its preceding proposition source, which refers to the source identifier of the upstream knowledge node or original retrieval fragment upon which the anomalous paragraph depends in its argument structure. By clustering the source labels of all anomalous paragraphs, those relying on the same source identifier are grouped into the same homologous anomalous group. The formation of homologous anomalous groups reveals a type of systematic knowledge gap: when multiple anomalous paragraphs rely on the same knowledge source, and this source is missing or biased in the original knowledge retrieval record, these paragraphs will collectively exhibit similar argumentation breaks. A horizontal comparison is performed on the anomalous paragraphs within the homologous anomalous group to extract their commonly missing argumentation elements. The method for extracting argumentation elements is to perform an intersection operation on the core assertions that were supplemented in the verification comparison of each paragraph within the group, extracting the knowledge types and relationship patterns that need to be supplemented in multiple paragraphs, and recording these commonly missing argumentation elements as the distribution features of the current verification batch. Distribution characteristics not only describe the subject categories of missing knowledge, but also include the hierarchical position of missing knowledge in the argument structure, such as "conditional constraint knowledge missing in the causal reasoning layer," which has positioning significance.

[0139] Successful pattern identification is based on a systematic analysis of historical bridging retrieval cases. For each successful bridging retrieval case, the relational schemas and argument roles contained in its retrieval conditions are analyzed. Relational schemas describe the semantic relationship type between subjects and objects in the retrieval conditions, such as "dependency," "causation," and "qualification." Argument roles describe the semantic functions of each semantic component in the retrieval conditions, such as "agent," "patient," "condition," and "result." Relational schemas and argument roles in the same case are combined to construct dependency rules for that case. The form of the dependency rules is "Under relational schema X, the knowledge node of argument role Y can effectively bridge the semantic gap of state type Z." Dependency rules generated from all successful cases are summarized to construct a dependency rule base. In the dependency rule base, dependency rules that recur in multiple different cases are identified, and these high-frequency rules are marked as successful patterns. The recurrence of successful patterns indicates that this type of dependency structure has universal bridging effectiveness in the semantic topology of the enterprise knowledge base and can serve as a reliable reference for subsequent query construction.

[0140] After successful pattern recognition, the extracted distribution features are combined with the dependency rule base to adjust the query construction strategy for reverse retrieval. For each common missing argument element recorded in the distribution features, dependency rules matching the knowledge type and hierarchical position of that argument element are selected from the dependency rule base. The selection criteria are the mapping relationship between the relational pattern described by the dependency rule and the semantic relation type to which the argument element belongs, as well as the correspondence between the argument role and the position of the argument element in the argument structure. The selected dependency rules are embedded in the dependency hierarchy of the hierarchical query expression. The hierarchical query expression is the structured query form used by reverse retrieval, and its dependency hierarchy is specifically used to describe the semantic dependency constraints between the target knowledge node and other knowledge nodes. After embedding the dependency rules into the dependency hierarchy, the query expression will not only match knowledge fragments semantically similar to the key terms during retrieval, but will also prioritize returning knowledge nodes that satisfy specific dependency structures, thereby improving the accuracy and recall rate of reverse retrieval when facing similar argument gaps. The revised query construction mechanism will be automatically applied in subsequent verification batches, enabling the overall verification process to have continuous optimization and adaptive capabilities. As the number of processed cases accumulates, the query quality of reverse retrieval will continue to improve as the dependency rule base becomes richer.

[0141] A second aspect of this invention provides a verification system for the generation results of an enterprise knowledge intelligent agent based on a reflection verification mechanism, comprising:

[0142] The record acquisition unit is used to acquire the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments;

[0143] An anomaly detection unit is used to decompose the response text to be verified into text paragraphs that carry independent semantic units, extract key terms and argumentation logic for each text paragraph, reverse construct a verification query and perform an independent retrieval in the enterprise knowledge base, and align and compare the reverse retrieval results with the knowledge retrieval records to identify abnormal paragraphs.

[0144] The missing knowledge analysis unit is used to determine the type and scope of missing knowledge by analyzing the contextual dependencies of the identified anomalous paragraphs in the response text to be verified.

[0145] The bridging retrieval unit is used to construct bridging retrieval conditions based on the knowledge type and topic range, combined with the semantics of adjacent paragraphs of the abnormal paragraph, to perform bridging retrieval in the enterprise knowledge base, obtain supplementary knowledge fragments that can establish semantic coherence, and generate a verification comparison containing the original paragraph and the supplementary knowledge fragments.

[0146] The replacement output unit is used to replace the content of the abnormal paragraph based on the verification comparison, generate and output the verified response text, extract the distribution characteristics of the abnormal paragraph and the success pattern of the bridging retrieval, and adjust the query construction strategy of the reverse retrieval.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for verifying the generation results of enterprise knowledge intelligent agents based on a reflection verification mechanism, characterized in that: include: Obtain the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments; The response text to be verified is decomposed into text paragraphs that carry independent semantic units. Key terms and argumentation logic are extracted for each text paragraph. A verification query is constructed in reverse and an independent search is performed in the enterprise knowledge base. The reverse search results are aligned and compared with the knowledge search records to identify abnormal paragraphs. For the identified anomalous paragraphs, the missing knowledge type and topic scope are determined by analyzing the contextual dependencies of the anomalous paragraphs in the response text to be verified; Based on the knowledge type and topic range, bridging retrieval conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging retrieval is performed in the enterprise knowledge base to obtain supplementary knowledge fragments that can establish semantic coherence, and a verification comparison containing the original paragraph and the supplementary knowledge fragments is generated. Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted.

2. The method according to claim 1, characterized in that, The response text to be verified is decomposed into text paragraphs carrying independent semantic units. Key terms and logical arguments are extracted from each text paragraph. A verification query is constructed in reverse and an independent search is performed in the enterprise knowledge base. The reverse search results are aligned and compared with the knowledge retrieval records to identify anomalous paragraphs, including: The semantic unit boundary is identified in the response text to be verified, where the subject-predicate-object structure is complete and all pronouns refer to objects within the current semantic unit. The semantic unit boundary is then segmented to obtain multiple independent text paragraphs. For each independent text paragraph, the predicate structure is parsed to extract the core terms representing the action, and the modification dependency relationship is parsed to extract the logical terms representing the constraint conditions. Based on the dependency hierarchy of the core terms and the logical terms in the modification dependency relationship, a hierarchical query expression is constructed, and the hierarchical query expression sets the terms that are dependent on the preceding terms as the retrieval constraints of the subsequent terms. The hierarchical query expression is submitted to the enterprise knowledge base to perform a retrieval and obtain a reverse retrieval knowledge set. The original knowledge set referenced when generating the independent text paragraph is extracted from the knowledge retrieval records. Calculate the semantic coverage of the argument content in the independent text paragraph in the reverse retrieval knowledge set, perform intersection verification between the original knowledge set and the reverse retrieval knowledge set, and mark the independent text paragraph as an abnormal paragraph when the semantic coverage is lower than the coverage threshold.

3. The method according to claim 2, characterized in that, Based on the dependency hierarchy of the core terms and logical terms in the modifying dependency relationship, a hierarchical query expression is constructed. This hierarchical query expression sets the terms of preceding dependencies as retrieval constraints for subsequent terms, including: The semantic role labeling between verbs and nouns in the modification dependency relationship is analyzed to identify the action semantics carried by the core terms and the argument semantics carried by the logical terms. Based on the modification direction of the action semantics by the argument semantics, the order of each term in the dependency chain is determined. The term located at the beginning of the dependency chain is used as the retrieval anchor to locate the knowledge fragment containing the term in the enterprise knowledge base, and the storage location of the knowledge fragment is recorded as the retrieval boundary. A search instruction is generated for subsequent terms in the dependency chain. A range parameter pointing to the search boundary is embedded in the search instruction so that the search instruction searches for the subsequent terms only within the knowledge fragment defined by the search boundary, and obtains the intersection knowledge fragment that contains both preceding and following terms. The storage location of the intersection knowledge fragment is updated to the new retrieval boundary. The next term in the dependency chain is used as the new subsequent term. The process of generating retrieval instructions and obtaining intersection knowledge fragments is repeated until all terms are processed. Then, the retrieval instructions are assembled into the hierarchical query expression according to the position of the dependency chain.

4. The method according to claim 1, characterized in that, For the identified anomalous paragraphs, the contextual dependencies of the anomalous paragraphs within the response text to be verified are analyzed to determine the type and scope of missing knowledge, including: Extract the state description of the subject entity from the preceding paragraph of the abnormal paragraph to construct the preceding state vector, extract the state description of the subject entity from the following paragraph to construct the following state vector, construct the state transition matrix based on the following state vector and the preceding state vector and perform singular value decomposition, and extract the left singular vector and right singular vector corresponding to the principal singular value. The left singular vector is projected onto the semantic space of the preceding paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the preceding anchor point knowledge. The right singular vector is projected onto the semantic space of the following paragraph, and the knowledge segment with the highest semantic similarity in the projection direction is identified as the following anchor point knowledge. Retrieve original knowledge fragments containing the predecessor anchor knowledge and the successor anchor knowledge from the enterprise knowledge base, extract the association path connecting the two original knowledge fragments, count the distribution of relationship types of each node in the association path, and take the relationship type with the highest proportion as the missing knowledge type. Calculate the rank of the state transition matrix, extract the corresponding number of principal singular vectors, map the principal singular vectors to the concept hierarchy structure of the enterprise knowledge base, and extract the common parent node of the mapped concepts as the topic scope.

5. The method according to claim 1, characterized in that, Based on the aforementioned knowledge type and topic scope, bridging search conditions are constructed by combining the semantics of adjacent paragraphs of the abnormal paragraph. A bridging search is then performed in the enterprise knowledge base to obtain supplementary knowledge fragments that establish semantic coherence. A verification comparison containing the original paragraph and the supplementary knowledge fragments is generated, including: Extract the core proposition of the preceding paragraph of the abnormal paragraph as the pre-proposition, extract the core proposition of the following paragraph as the post-proposition, parse the predicate structure of the pre-proposition and the post-proposition, and identify the missing argument roles between the two propositions; Based on the knowledge type, determine the relational pattern required to fill the argument role, retrieve knowledge fragments containing the relational pattern from the knowledge subset defined by the topic scope, and convert the retrieved knowledge fragments into propositional expressions to construct an intermediate proposition set; Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with the preceding proposition and the premise support relationship with the subsequent proposition. The knowledge fragments of the intermediate propositions that pass the verification are used as candidate bridging knowledge. Calculate the logical tightness of each candidate bridging knowledge in the argument chain, select the candidate bridging knowledge with the highest logical tightness as a supplementary knowledge fragment, organize the abnormal paragraph and the supplementary knowledge fragment into a verification unit according to the argument order, mark the reasoning path of the antecedent proposition, the intermediate proposition and the consequent proposition in the verification unit, and generate a verification check containing logical dependencies.

6. The method according to claim 5, characterized in that, Based on the set of intermediate propositions, verify whether each intermediate proposition satisfies the conclusion derivation relationship with its preceding proposition and the premise support relationship with its subsequent proposition. The knowledge fragments of the verified intermediate propositions are used as candidate bridging knowledge, including: The preceding propositions are formalized into preceding predicate expressions. Derivable atomic propositions are extracted from these preceding predicate expressions to form a forward reachable set. The following propositions are formalized into following predicate expressions. Required atomic propositions are extracted from these following predicate expressions to form a backward necessary set. For each intermediate proposition in the intermediate proposition set, it is formalized into an intermediate predicate expression. The dependent variables of the intermediate predicate expression are parsed, and it is determined whether the dependent variables are fully bound by the forward reachable set. If they are fully bound, the atomic propositions of the forward reachable set are substituted into the dependent variables to perform variable substitution, and the truth value of the expression after substitution is calculated. The free variables derived from the intermediate predicate expression are analyzed, and it is determined whether the free variables and the backward required set form an injective mapping. If an injective mapping is formed, the predicate expression, the intermediate predicate expression, and the postdicate expression are constructed into a derivation sequence in the order of derivation. Logical resolution is performed on the derivation sequence to verify whether the resolution result is an empty clause. When the resolution result is an empty clause, the knowledge fragment corresponding to the intermediate proposition is taken as a candidate bridging knowledge.

7. The method according to claim 1, characterized in that, Based on the verification comparison, the abnormal paragraphs are replaced with content, a verified response text is generated and output, the distribution characteristics of the abnormal paragraphs and the success pattern of the bridging retrieval are extracted, and the query construction strategy of the reverse retrieval is adjusted, including: The propositional structure of the supplementary knowledge fragments in the verification comparison is analyzed, the core assertions are extracted, the core assertions are aligned with the abnormal paragraphs, the propositional components that are disproven by the core assertions are identified and removed, the core assertions are inserted into the corresponding logical positions according to the reasoning path marked by the verification comparison, the undisproven propositional components are kept, the verified response text is generated and output. The source of the preceding propositions that each anomalous paragraph depends on is marked. Anomalous paragraphs that depend on the same source are classified into a homologous anomalous group. The argument elements that are commonly missing in the homologous anomalous group are extracted as distribution features. The system parses the relational patterns and argument roles contained in the search conditions of successful bridging search cases, constructs a dependency rule base between the two, identifies dependency rules that appear repeatedly in multiple cases and marks them as successful patterns, selects corresponding dependency rules from the dependency rule base for the argument elements, and embeds the dependency rules into the dependency hierarchy of the hierarchical query expression.

8. A verification system for the generation result of an enterprise knowledge intelligent agent based on a reflection and verification mechanism, used to implement the method as described in any one of claims 1-7, characterized in that, include: The record acquisition unit is used to acquire the response text to be verified generated by the knowledge agent and the corresponding knowledge retrieval record, wherein the knowledge retrieval record contains the retrieved knowledge fragments; An anomaly detection unit is used to decompose the response text to be verified into text paragraphs that carry independent semantic units, extract key terms and argumentation logic for each text paragraph, reverse construct a verification query and perform an independent retrieval in the enterprise knowledge base, and align and compare the reverse retrieval results with the knowledge retrieval records to identify abnormal paragraphs. The missing knowledge analysis unit is used to determine the type and scope of missing knowledge by analyzing the contextual dependencies of the identified anomalous paragraphs in the response text to be verified. The bridging retrieval unit is used to construct bridging retrieval conditions based on the knowledge type and topic range, combined with the semantics of adjacent paragraphs of the abnormal paragraph, to perform bridging retrieval in the enterprise knowledge base, obtain supplementary knowledge fragments that can establish semantic coherence, and generate a verification comparison containing the original paragraph and the supplementary knowledge fragments. The replacement output unit is used to replace the content of the abnormal paragraph based on the verification comparison, generate and output the verified response text, extract the distribution characteristics of the abnormal paragraph and the success pattern of the bridging retrieval, and adjust the query construction strategy of the reverse retrieval.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images