Deep learning and semantic structure information-based large model hallucination detection method

Abstract

The present application provides a large model hallucination detection method based on deep learning and semantic structure information, aiming at the problem that existing methods are difficult to determine and locate the model fictitious information within the range of verifiable knowledge, and the evidence is not traceable, the present application proposes: performing sentence, dependency and event analysis on the generated text, constructing a semantic graph and a claim list, extracting standardized identification features, retrieving candidate objects and combining them with the least conditions to form a set of distinguishing conditions, generating a semantic query plan and recording ambiguity exclusion, retrieving the knowledge base and document library and forming an evidence list, attributing the evidence according to the distinguishing conditions, generating claim level support, conflict and unknown judgment, calculating the external consistency value on the target object, and checking the internal contradiction, combining the double threshold to output the final judgment, positioning information and evidence chain, the present application can improve the accuracy and explainability of hallucination detection, and is suitable for model output audit and fact verification scenes.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and natural language processing technology, and in particular relates to a large-scale hallucination detection method based on deep learning and semantic structure information. Background Technology

[0002] Large-scale language models are prone to producing illusions when generating text in open domains, meaning they output content that is inconsistent with factual knowledge or cannot be verified. In scenarios such as risk control and compliance, content moderation, and knowledge-based question answering, verifiable fact-checking and locating of specific claims generated by the model are necessary. Such checks typically rely on structured knowledge bases and unstructured document libraries, requiring the decomposition of natural language expressions into elements such as subject, predicate, object, time, location, type, and numerical values, establishing semantic connections and temporal order within cross-sentence and cross-paragraph contexts. Furthermore, phenomena such as pronoun references, ellipsis, aliases, and event restatements in the text increase the difficulty of entity / event disambiguation; differences in the timeliness and type of evidence from different sources require checks to be conducted under constraints such as time windows, type consistency, and location range to ensure the results are interpretable and traceable.

[0003] Existing technologies primarily employ two approaches for hallucination detection and fact-checking: First, enhanced retrieval verification involves segmenting the generated text, performing named entity recognition and dependency analysis to extract elements such as subject-relationship-object and time. Keywords or structured fields are then used for knowledge base and document database searches to retrieve candidate evidence. Second, semantic matching and classification-based comparisons determine whether candidate evidence supports, conflicts with, or is unknown in relation to the original sentence, or classify it using rules and matching degrees. Some methods introduce entity links and event extraction to establish cross-sentence citation relationships, filter evidence using time fields and type labels, and summarize the results to claim-level or paragraph-level conclusions using thresholds or weighted strategies for downstream review and alerts.

[0004] Existing solutions generally lack a unique positioning mechanism for claims: candidate objects mostly rely on keyword matching, failing to complete the minimum combination of distinguishable conditions under strong constraints such as time windows and types, making it difficult to clearly point to specific objects; the attribution of evidence does not clearly distinguish between target objects and excluded objects, and the reasons and boundaries for exclusion lack traceable records; support / conflict determination mostly stays at the sentence level matching, lacking a systematic verification of the uniqueness and sequence of claims, making it difficult to output locatable and interpretable verification conclusions and complete chains of evidence in complex narratives.

[0005] Therefore, how to provide a large-scale hallucination detection method that can solve the above-mentioned technical problems is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to provide a large-model illusion detection method based on deep learning and semantic structure information. In real-world application scenarios, this method aims to achieve uniquely identifiable object alignment and temporal retrieval of claims generated by large models within a knowledge base / document library. Under strong constraints such as time and type, it completes evidence attribution and judgment, while forming a traceable evidence chain that includes ambiguity exclusion records. This enables the final conclusion to accurately locate specific claims and fields and meet the usability requirements of scenarios such as review and verification.

[0007] A large-model hallucination detection method based on deep learning and semantic structure information according to an embodiment of the present invention includes:

[0008] S1. Receive the generated text from the language model, perform semantic structure parsing, and generate a semantic graph and a list of claims;

[0009] S2. Extract the identification feature set based on the semantic graph, retrieve the candidate object set in the knowledge base, determine the distinguishable conditions of each candidate object based on the identification feature set, select the minimum combination of conditions that can be uniquely pointed to in the claim list to form the distinguishable condition set, generate a semantic query plan based on the distinguishable condition set, and record the ambiguity exclusion records of excluded candidates and corresponding conditions.

[0010] S3. Query the knowledge base and document library according to the semantic query plan, generate an evidence list, divide the evidence into target object evidence and excluded object evidence according to the distinguishing condition set, form the evidence judgment result for the claim list, and output the evidence list and evidence judgment result.

[0011] S4. Based on the evidence list and the evidence judgment results, calculate the external consistency value on the target object defined by the distinguishing condition set.

[0012] S5. Perform uniqueness and time sequence checks on the claim list based on the semantic graph, and generate internal contradiction markers;

[0013] S6. Based on the external consistency value and the internal contradiction marker, generate the final judgment and location information under the constraints of the first threshold and the second threshold, and output the evidence chain including the ambiguity exclusion record and the evidence list reference.

[0014] Optionally, S1 is as follows:

[0015] The generated text from the language model is segmented into sentences and the text boundaries are identified to obtain sentence sequences and paragraph orders;

[0016] Perform part-of-speech tagging and dependency analysis on the sentence sequence to determine the positions of predicates and arguments, and generate intra-sentence relations;

[0017] Perform named entity recognition and event extraction, extracting roles, organizations, locations, times, and values ​​to form entity and event candidates;

[0018] Performs referential resolution and cross-sentence linking, merges the same entity and the same event, and normalizes time expressions into time points or time windows;

[0019] Construct a semantic graph with entities and events as nodes and relationships and roles as edges, and add time, location, numerical and type attributes, and write time order constraints;

[0020] Set a confidence threshold to remove nodes and relationships in the semantic graph with a confidence level below the threshold, and retain components with a confidence level not lower than the threshold.

[0021] Extract a list of claims from the semantic graph, organize the subject, predicate, object, time, location, numerical value and type constraints into claim items, remove duplicate claims, and output the semantic graph and the list of claims.

[0022] Optionally, S2 is as follows:

[0023] Using semantic graphs and claims lists as input, we extract identification feature sets and standardize the representations of roles, organizations, locations, times, values, and types to form an alignable identification feature set.

[0024] Based on the identification feature set, a candidate object set is retrieved from the knowledge base. Field matching and time window overlap determination are used for initial screening to remove objects that do not overlap in time or are inconsistent in type, thus obtaining a candidate object set that meets the semantic boundaries of the claim list.

[0025] For each candidate object in the candidate object set, determine the distinguishable conditions, align the role, organization, location, time, value and type with the identification feature set, and set the value range and matching method for each condition. Among them, time and type are mandatory constraints, and if they are not met, they are marked as unusable conditions.

[0026] Perform minimum condition combination filtering on the candidate object set. With the goal of minimizing the number of conditions, add time, type, location, organization, role, and value in a fixed priority order. Once added, it is determined to be unique. The determination of uniqueness is based on the number of objects in the candidate object set that meet all the added conditions. If the condition is not met, continue to increment the combination until it is met.

[0027] The minimum combination of conditions that achieves unique pointing is summarized into a distinguishing condition set. For candidate objects that are not uniquely pointed to, ambiguity exclusion records are generated based on the mismatch relationship with the distinguishing condition set. The ambiguity exclusion record includes the excluded candidate, the name of the mismatch condition, the mismatch value, and the mismatch time window, which is used to clarify the reason for exclusion.

[0028] Based on the distinguishing condition set, a semantic query plan is generated, which is then mapped to executable fields and filtering conditions for knowledge base query and document retrieval. These include time window filtering, location range filtering, type consistency filtering, numerical range filtering, and role matching rules, forming an executable field list and filtering logic.

[0029] Perform field alignment and boundary checks on the semantic query plan to ensure that the fields and filtering conditions correspond to the distinguishing condition set, that the time window does not exceed the time range of the claim list, and that the types are consistent. This completes the execution of the semantic query plan and outputs the distinguishing condition set, the semantic query plan, and the ambiguity exclusion record.

[0030] Optionally, S3 specifically refers to:

[0031] The semantic query plan is mapped to a retrieval request for the knowledge base and document library, the retrieval is executed, the returned content is parsed, the subject, predicate, object, time, location, numerical and type fields are extracted, and the evidence content in the same paragraph as the field is located to form candidate entries.

[0032] Based on the semantic query plan, candidate items are filtered by time window and type constraints. Candidate items whose time falls within the time window of the semantic query plan and whose type is consistent are included in the evidence list. The field representation is unified so that the evidence list and the claim list are associated with each other in terms of subject, predicate, object, time, location, value and type.

[0033] The evidence list is classified according to the differentiation criteria set. Items that meet all the criteria in the differentiation criteria set are classified as target object evidence, while items that do not meet any of the mandatory criteria of time or type are classified as excluded object evidence. The remaining mismatched items of location, institution, role, and value are compared with the remaining criteria in the differentiation criteria set and classified as excluded object evidence.

[0034] Perform consistency comparison on the evidence of the target object, match the subject, predicate, object and location of the claim list with the evidence of the target object item by item, and perform an equality judgment on the value and the value of the claim list under the premise that the time of the evidence is included in the time window of the distinguishing condition set and the type is consistent. If all the above matches are satisfied, it is determined to be supported.

[0035] Conflict detection is performed on the evidence of the target object. Under the premise that the evidence time is included in the time window of the distinguishing condition set and the type is consistent, if any of the following situations occur, such as opposite predicate meanings, unequal values, or inconsistent locations, it is determined to be a conflict. When both supporting and conflicting items appear in the evidence of the target object, the conflict shall prevail.

[0036] When the evidence for the target object is empty and the evidence list only contains the evidence for the excluded object, it is determined to be unknown. When the evidence for the target object is empty and the evidence list is empty, it is determined to be unknown. The evidence for the excluded object does not participate in the generation of support and conflict.

[0037] The supporting, conflicting, and unknown claims are summarized in the claim list dimension to form the evidence judgment result, and the evidence list and evidence judgment result are output.

[0038] Optionally, S4 specifically refers to:

[0039] The evidence list, evidence judgment results, and differentiation condition set are used as input. The evidence attribution is limited according to the differentiation condition set. Items that meet all constraints of the differentiation condition set are selected from the evidence list as target object evidence. Items that do not meet any of the time and type constraints are removed to obtain the target object evidence set used for calculation.

[0040] Based on the claim list, the evidence of the target object is grouped, a mapping relationship between claim items and evidence of the target object is established, a set of evidence of the target object is determined for each claim item for calculation, and time, location and numerical fields are reserved for each piece of evidence for subsequent weight calculation;

[0041] Based on the evidence determination results, the support set and conflict set are determined at the claim item level. When the evidence determination result is supportive, the target object evidence of the claim item is recorded in the support set. When the evidence determination result is conflicting, the target object evidence of the claim item is recorded in the conflict set. When the evidence determination result is unknown, it is not included in the calculation.

[0042] For each piece of evidence of the target object in the support set and conflict set, a time weight is calculated. The time distance is determined by the time window that distinguishes the condition set and the time of the claim list. The corresponding weight is given according to the preset time segment threshold. A higher weight is given for a shorter time distance and a lower weight is given for a longer time distance.

[0043] Set support weight coefficient and conflict weight coefficient. The conflict weight coefficient is greater than the support weight coefficient. Before calculation, check the matching completeness of subject, predicate, object and location. Only when all the above fields match, contribution and penalty are included. Incomplete matching of target object evidence is not included in the calculation.

[0044] Calculate an item-level consistency score for each claim item. The time weights of the supporting set are accumulated according to the support weight coefficient, and the time weights of the conflict set are accumulated according to the conflict weight coefficient. The difference between the two is mapped to an item-level consistency score according to a preset standardization rule. The item-level consistency score ranges from zero to one.

[0045] The external consistency value is obtained by averaging the item-level consistency scores of all claim items with equal weights, and then output on the target object defined by the distinguishing condition set.

[0046] Optional, S5 specifically includes:

[0047] Using semantic graphs and a list of claims as input, the subject, predicate, object, time, location, value, and type of each claim item are extracted, and the time is unified into a time point or time window to form a set of claim items for verification.

[0048] The claim set is grouped by subject, predicate, object and type. A uniqueness check is performed within each group. If any two claims have overlapping time windows and the locations or values ​​are inconsistent, they are judged to be in uniqueness conflict.

[0049] Based on the time and relationships in the semantic graph, the preceding and subsequent claims are determined for each group, and a set of sequential constraints is established. The determination of the preceding and subsequent claims is based on the start and end relationships of the claim time and the semantic graph relationships.

[0050] Perform a time order check on the set of sequential constraints. If the preceding claim is a time point, its time must not be later than the time of the subsequent claim. If the preceding claim is a time window, its end time must not be later than the start time of the subsequent claim. If the conditions are not met, it is determined to be a time order conflict.

[0051] The uniqueness conflict and the chronological order conflict are merged at the claim item level to generate an internal contradiction marker, and the internal contradiction marker is output on the claim list.

[0052] Optional, S6 specifically includes:

[0053] Using external consistency values ​​and internal contradiction markers as inputs, load the first threshold and the second threshold, check the relationship between the two thresholds, and confirm that the second threshold is less than the first threshold.

[0054] The final judgment is generated based on the external consistency value and the internal contradiction mark. When the external consistency value is not lower than the first threshold and the internal contradiction mark is empty, it is judged as support. When the external consistency value is lower than the second threshold or the internal contradiction mark is not empty, it is judged as conflict. All other cases are judged as unknown.

[0055] Location information is generated based on the final judgment and internal contradiction markers. For support, the claim item and time window are located; for conflict, the claim item where the inconsistency occurred and the time, location or numerical field are located; for unknown, the claim item with missing evidence is located.

[0056] Based on the set of distinguishing conditions, the evidence of the target object corresponding to the final judgment is screened from the evidence list. The evidence chain is constructed by combining the ambiguity elimination record, and the final judgment, location information and evidence list references are established to correspond.

[0057] Output the final judgment, location information, and the chain of evidence including the ambiguity exclusion record and the evidence list cited by the evidence chain.

[0058] The beneficial effects of this invention are:

[0059] This proposal presents an improved method for uniquely identifying candidate objects and generating semantic retrieval plans. The technical means include extracting standardized identification feature sets from semantic graphs and claim lists; using time and type as mandatory constraints; and performing minimum condition combination filtering according to fixed priorities of time, type, location, institution, role, and numerical value to form a distinguishing condition set. This set is then mapped to executable fields and filtering logic to complete the consistency verification of time window boundaries and type, while simultaneously generating ambiguity exclusion records to mark mismatched candidate objects. Unlike solutions that rely on keyword or single-field hits, this improved method achieves unique identification of specific objects within the knowledge base and document library, reducing object confusion and interference from cross-object evidence, ensuring consistency between retrieval boundaries and claim constraints, and providing clear and traceable evidence for subsequent evidence attribution.

[0060] This proposal puts forward a novel method for attributing and measuring the consistency of claim-level evidence. The technical means include explicitly classifying evidence into target-object evidence and excluded-object evidence based on a set of distinguishing conditions; performing support, conflict, and unknown determinations on target-object evidence, prioritizing conflict; introducing time weighting and weighting rules with different coefficients, only awarding contribution or penalty when the subject, predicate, object, and location completely match; obtaining an item-level consistency score and summing it into an external consistency value. This method establishes a one-to-one correspondence between evidence and claim, quantifies the impact of time distance on verification conclusions, suppresses the interference of different-object evidence on judgment, and outputs a stable claim-level consistency measure under time windows and type constraints, providing calculable and verifiable support for the final conclusion.

[0061] This proposal presents a comprehensive illusion detection method and technique for large-scale model outputs, encompassing semantic graph construction, claim uniqueness and temporal sequence checks, evidence retrieval and attribution, external consistency calculation, dual-threshold decision-making, and evidence chain construction. By performing uniqueness and temporal sequence checks on the semantic graph to generate internal contradiction markers, and fusing them with external consistency values ​​under dual-threshold constraints, the method locates specific claim entries and time, location, or numerical fields. Simultaneously, it outputs an evidence chain including ambiguity exclusion records and evidence list citations. The overall method completes a closed loop of claim-level object alignment and temporal sequence retrieval in real-world applications, achieving the location, interpretability, and traceability of conclusions. This clearly distinguishes it from existing processes based on sentence matching or coarse-grained filtering, providing a systematic solution to the core technical problem of achieving uniquely identifiable object alignment and temporal sequence retrieval within a knowledge base / document repository and forming a traceable evidence chain. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a flowchart of a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention;

[0064] Figure 2 This is a flowchart of semantic parsing and semantic graph construction for a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention.

[0065] Figure 3 This is a flowchart of the candidate object retrieval and minimum condition combination process of a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention.

[0066] Figure 4 This is a flowchart of the external consistency calculation for a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention.

[0067] Figure 5 This is a flowchart of the dual-threshold fusion and evidence chain output of a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention.

[0068] Figure 6 This is a schematic diagram of evidence attribution and evidence chain for a large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention.

[0069] Figure 7 This is a schematic diagram showing the comparison before and after the standardization and alignment of the recognition features in the large-model hallucination detection method based on deep learning and semantic structural information proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figures 1 to 7 A large-model hallucination detection method based on deep learning and semantic structural information includes:

[0072] S1. Receive the generated text from the language model, perform semantic structure parsing, and generate a semantic graph and a list of claims;

[0073] S2. Extract the identification feature set based on the semantic graph, retrieve the candidate object set in the knowledge base, determine the distinguishable conditions of each candidate object based on the identification feature set, select the minimum combination of conditions that can be uniquely pointed to in the claim list to form the distinguishable condition set, generate a semantic query plan based on the distinguishable condition set, and record the ambiguity exclusion records of excluded candidates and corresponding conditions.

[0074] S3. Query the knowledge base and document library according to the semantic query plan, generate an evidence list, divide the evidence into target object evidence and excluded object evidence according to the distinguishing condition set, form the evidence judgment result for the claim list, and output the evidence list and evidence judgment result.

[0075] S4. Based on the evidence list and the evidence judgment results, calculate the external consistency value on the target object defined by the distinguishing condition set.

[0076] S5. Perform uniqueness and time sequence checks on the claim list based on the semantic graph, and generate internal contradiction markers;

[0077] S6. Based on the external consistency value and the internal contradiction marker, generate the final judgment and location information under the constraints of the first threshold and the second threshold, and output the evidence chain including the ambiguity exclusion record and the evidence list reference.

[0078] In this embodiment, step S1 specifically includes:

[0079] Let the generated text of the language model be denoted as ,by Generate a semantic graph from the input. List of claims The semantic graph From the set of nodes With edge set The structure consists of nodes containing entities and events, edges containing relationships and roles, and additional attributes such as time, location, value, and type. This is the aforementioned list of claims. It consists of several claim entries, each claim entry containing subject, predicate, object, time, location, numerical value, and type constraints;

[0080] Will Sentence segmentation and discourse boundary recognition are performed to obtain sentence sequences. Paragraph order The criteria for judgment are punctuation boundaries, abbreviation lists, and paragraph indentation. Priority is given to ensuring the integrity of parentheses and quotation marks within sentences, establishing the subordinate relationship between sentences and paragraphs. exist The position on it is unique;

[0081] right Perform part-of-speech tagging and dependency parsing sentence by sentence to determine the positions of predicates and arguments in each sentence and generate intra-sentence relations. The system uses a sequence labeling network to output part-of-speech tags, a transition-based parser to output dependency arcs, and gathers argument slots around the predicate to form a predicate-argument-modifier structure description for subsequent event extraction.

[0082] Perform named entity recognition and event extraction to construct an entity set. With event candidate set It maps roles, organizations, locations, times, and values ​​to entity types, and identifies event trigger words using a trigger word list and dependency patterns, based on... Locate the agent, patient, location, time, and quantity to the corresponding role slots, output the event candidates and their slot filling results, and retain the position index of the trigger word in the original sentence for subsequent linking;

[0083] The process involves dereference resolution and cross-sentence linking, merging entities and events with the same name to obtain consistent entity and event identifiers across sentences. Referential string matching and the nearest antecedent rule are used, prioritizing merging based on unique identifiers such as person names, organization names, and place names. The merged results are then verified using context-compatible role consistency constraints. Time expressions are standardized, and absolute time is parsed into time points. or time window Relative time is based on document base time. For anchor point derivation, output a unified time point or time window representation;

[0084] Constructing a semantic graph ,Will and Entities and events are added to the node set. Add and collect based on relationships and roles. Furthermore, time, location, numerical, and type attributes are added to nodes and edges. Time order constraints are written based on the order of event trigger words and the order of normalized time, using two types of constraints, namely "prior" and "include", to express the temporal and inclusion relationships between events.

[0085] Set confidence threshold semantic graph The confidence scores of nodes and relationships are evaluated using a linear combination of recognition probability, syntactic consistency score, and temporal normalization consistency score to obtain node confidence and relationship confidence. Nodes with confidence scores below a certain threshold are considered confidence scores. The nodes and relationships from Remove from the list, retaining only those with a confidence level of not less than [a certain value]. The components are used to reduce the sources of noise in subsequent claims;

[0086] From semantic graph Extract the list of claims Centered on an event node, the system reads the associated subject, predicate, object, time, location, numerical, and type attributes, combining them into claim entries. Claims with the same subject, predicate, and object, overlapping time windows, and consistent locations are merged, taking the intersection time window or replacing the time window with a more precise time point. Numerical fields are retained when consistent and separated when inconsistent for subsequent adjudication. Claims with identical field content are deduplicated, and a semantic graph is output. List of claims This provides structured input for subsequent generation of condition sets and adjudication of evidence partitions.

[0087] In this embodiment, step S2 specifically includes:

[0088] semantic graph List of claims As input, the output is a set of recognition features. Candidate object set Distinguishing condition sets With semantic query plan The identification feature set For the subjects, predicates, objects, time, location, values, and types in the claim list, a unified field is used to represent them, which is used to establish an aligned search entry point in the knowledge base;

[0089] The roles, organizations, locations, times, values, and types in the semantic graph are standardized to obtain the recognition feature set. Time is unified into a time window or time point It is then converted into a time window format, with locations uniformly represented as a combination of administrative level and coordinate range, numerical values ​​uniformly represented as closed intervals, types uniformly represented as a controlled vocabulary, and roles and institutions represented using a dual-channel representation of unique identifiers and synonym mappings, forming a set of comparable fields for matching.

[0090] Based on the recognition feature set Retrieve candidate object sets from the knowledge base The process involves exact matching or inclusion matching of the subject, organization, and location fields; consistency matching of the type field; and time window overlap determination for the time field, requiring that the intersection of the time windows be non-empty. Objects with non-overlapping times or inconsistent types are eliminated, and objects that meet the semantic boundaries of the claim list are retained to create candidate objects. A unified field view for subsequent condition alignment;

[0091] For each candidate in the candidate set, determine the distinguishable conditions and match the role, organization, location, time, value, and type with the identification feature set. Align each item one by one, set the value range and matching method for each condition, where time uses the time window inclusion rule, type uses the consistency rule, location uses the hierarchical inclusion rule, value uses the interval inclusion rule, and role and organization use the unique identifier rule. Mark time and type as mandatory constraints, and mark any condition that is not met as unavailable.

[0092] In the candidate object set The minimum condition combination screening is performed, aiming for uniqueness. Time, type, location, organization, role, and value are added sequentially with a fixed priority. Uniqueness is determined after each addition, based on the number of candidate objects satisfying all added conditions being one. To ensure the computability of the method, the selection of the minimum condition combination is expressed as the constraint optimization of the following formula:

[0093] ;

[0094] ;

[0095] In the formula, Indicates a combination of conditions. Indicates the number of conditions. Represents the set of candidate objects. For indicator functions, This indicates that the candidate object meets the requirements in the fields of time, type, location, organization, role, and numeric value. All added matching constraints in the above formula are satisfied simultaneously with the unique pointer and the mandatory constraint. The solution uses the aforementioned greedy addition strategy with fixed priority. If a unique pointer is not found, continue adding the next condition.

[0096] The minimum number of conditions required to achieve a unique pointer are grouped into a distinguishing condition set. For candidate objects that are not uniquely pointed to, based on and The mismatch relationship generates an ambiguity exclusion record, which records the identifier of the excluded candidate, the name of the mismatch condition, the mismatch value and the mismatch time window, and is used to clarify the reason for exclusion and maintain the boundary of the distinguishing condition set in subsequent steps.

[0097] Based on the distinguishing condition set Generate semantic query plan ,Will The executable fields and filtering conditions are mapped to knowledge base queries and document retrieval, including time window filtering, location range filtering, type consistency filtering, numerical range filtering and role matching rules. Each filtering condition is assigned a field name, operator and value structure to form an executable field list and filtering logic. The predicates in the claim list are converted into query predicate templates to ensure that the retrieval results are consistent with the claim list in terms of predicate semantics.

[0098] semantic query plan Perform field alignment and boundary checks, Each field in The corresponding conditions in the list are checked one by one to ensure that the time window does not exceed the time range of the claim list, that the types are consistent and there is no ambiguous mapping, and that the location level and numerical range match the recognition feature set. Maintain consistency, complete the verification, and output the distinguishing condition set. Semantic query plan The ambiguous exclusion record provides a stable input for subsequent evidence list generation and evidence partitioning decisions.

[0099] In this embodiment, step S3 specifically includes:

[0100] semantic query plan Distinguishing condition sets List of claims As input, generate a list of evidence. The evidence list is consistent with the evidence assessment results. It is a structured sequence of entries, each entry containing subject, predicate, object, time, location, numerical and type fields, and retains the paragraph position of the evidence content in the document for contextual verification;

[0101] semantic query plan The mapping is used for retrieval requests oriented towards knowledge bases and document databases. On the knowledge base side, field retrieval is used to map the subject, predicate, object, and type to exact matches, time to time window constraints, and location to coordinates or administrative division filters. On the document database side, Boolean queries are used to combine predicate templates with the subject and object into query clauses. The returned text is segmented and parsed, and the context of the same paragraph as the field is used as evidence content to extract the subject, predicate, object, time, location, numerical, and type fields to form a set of candidate entries.

[0102] Based on the semantic query plan, candidate items are filtered using time windows and type constraints, and the time field is standardized to a closed interval. Requirements and The intersection of the time windows in the data is not empty, and the type field has the same requirements as... Candidate items that are of the same type and satisfy the above two constraints are included in the evidence list. ,right The fields in the document are represented in a unified manner, making the subject, predicate, object, time, location, value, type, and claim list consistent. Establish corresponding relationships to facilitate subsequent comparisons;

[0103] Based on the distinguishing condition set List of evidence To determine ownership, it will satisfy... Entries meeting all conditions are classified as target object evidence; entries that do not meet any of the mandatory conditions of time or type are classified as excluded object evidence; and other mismatched entries regarding location, institution, role, and value are classified as... The range of values ​​in the criteria is compared with the matching method, and those that do not meet the criteria are included in the excluded object evidence, thus forming the claim list. Each claim entry Establish a target object evidence set as a unit. Evidence set of excluded objects ;

[0104] Evidence of the target object Perform a consistency comparison to assert the entries. The subject, predicate, object, and location are the main keys, and The corresponding fields in the evidence were matched item by item, and the evidence was verified at the time of the evidence. Provided that the time windows are contained within each other and of the same type, an equality check is performed on the numeric fields. The criterion for equality is that the absolute difference does not exceed the numeric tolerance. When all fields meet the matching condition, the evidence is marked as a supporting candidate. Predicate consistency is determined by a combination of predicate template matching and synonym mapping.

[0105] Evidence of the target object Conflict detection is performed on the evidence time. Given that the time windows are contained and of the same type, if the predicates have opposite meanings and the absolute difference in values ​​is greater than 1, then... If the location does not meet any of the conditions of the level, the evidence is marked as a conflict candidate. If the meaning of the predicate is opposite, it is determined based on the predicate opposition mapping table. If the location is inconsistent, it is determined by the intersection of the coordinate boundaries being empty. When the same claim item has both supporting candidates and conflict candidates, the conflict candidate shall prevail.

[0106] When the evidence set of the target object Empty and evidence list When only evidence of the excluded object is included, mark the claim entry as unknown. When empty, it is also marked as unknown, and the evidence set of excluded objects is also included. It does not participate in the generation of support and conflict, nor does it change the adjudication of the evidence of the target object;

[0107] Supporting, conflicting, and unknown claims are summarized in the claim list to form an evidence assessment result, which is then compared with the evidence list. Establish correspondences to support subsequent measurement; to make the adjudication rules clear and computable, in the claim entries The above uses the following single determination formula to define the evidence determination result. :

[0108] ;

[0109] in, To distinguish condition sets The set of evidence for the target object under certain conditions. To support the determination indicator function, when the subject, predicate, object, and location match item by item and are of the same type, the evidence time is determined. The time window is included, and the absolute difference of the values ​​does not exceed [a certain value]. Time to take Otherwise take , For conflict determination indicator functions, when the types are consistent and the evidence time is... The time window contains, and there are instances where the predicates have opposite meanings or the absolute difference in values ​​is greater than 1. If the location does not meet the requirement of containing any of the following conditions, then take... Otherwise take , This is the numerical tolerance threshold, and it is a positive number.

[0110] Through the above steps, the process is completed from semantic query plan-driven retrieval, to the construction of evidence list under time and type constraints, to evidence partitioning and adjudication with the distinguishing condition set as the boundary, and finally to the generation of evidence judgment results at the claim list dimension. The implementation method ensures that the evidence of the target object and the evidence of the excluded object are strictly separated, and ensures that the generation conditions of support, conflict and unknown are calculable and consistent with the time and type constraints of the semantic query plan.

[0111] In this embodiment, step S4 specifically includes:

[0112] List of evidence Evidence determination results and distinguishing condition sets As input, the list of claims Generate externally consistent values ​​for the index. The aforementioned list of evidence Includes fields for subject, predicate, object, time, location, value, and type; evidence determination result. Label the claims as supported, conflicting, or unknown to differentiate the condition sets. Define the boundaries of the target object to ensure that subsequent calculations depend only on evidence of the target object;

[0113] Distinguish condition sets Applied to the list of evidence For each piece of evidence, the time and type were checked item by item, and the time was determined by... The given time window includes rules for judgment, and the type is judged using consistent rules. Entries that do not meet any constraint are removed, and entries that meet all constraints are marked as target object evidence, forming a target object evidence set for calculation. The original time, location and numerical fields are retained to support subsequent weight calculation and matching verification.

[0114] According to the list of claims Group the evidence of the target object according to the claim items. Using this as the key, evidence of the target object that matches the subject, predicate, object, and location will be aggregated to establish a mapping. The retention period for each piece of evidence The location and numerical fields are used for time weighting and matching completeness checks. If a claim item has no matching evidence, then... It is an empty set;

[0115] Based on the evidence assessment results Determine the supporting and conflicting sets at the claim item level, and for each... Evidence of the target object that is marked as supported will be added to the support set. Evidence that is marked as a target of conflict will be included in the conflict set. When the outcome of the evidence determination is unknown, it is not included in the calculation. The same evidence is deduplicated to prevent it from being counted repeatedly in the same set.

[0116] right and The calculation time weight for each piece of evidence related to the target object within the calculation When the claim item's time is a point in time. At that time, adopt Time interval for evidence The shortest distance is used as the time distance, when the claimed entry time is the time window. When the time frame is exceeded, the nearest boundary distance between two time windows is used as the time distance. Based on a preset time segmentation threshold, evidence with the shortest distance segment is assigned a higher weight, evidence with the middle distance segment is assigned a medium weight, and evidence with the longest distance segment is assigned a lower weight. During the time window, the weight is reset to zero to ensure... It falls within the range of zero to one, so that it can form a bounded metric together with the weighting coefficients;

[0117] Set supported weight coefficients Conflict weighting coefficient ,constraint Greater than Before calculation Internal evidence undergoes a matching integrity check. Only evidence where the subject, predicate, object, and location all match and are of the same type is retained for its time weighting and added to the accumulation. Evidence that fails the check is removed from the list. and Removed from the valid count, not participating in contribution and penalty, to avoid incomplete matching causing a shift in the item-level consistency score;

[0118] Furthermore, first in the claim item Calculate the item-level consistency score, and then apply it to the set of claim items involved in the calculation. The external consistency value is obtained by averaging. The item-level consistency score is calculated using a bounded ratio formed by the difference and sum of the time weights for support and conflict, and a linear mapping from zero to one is applied. To ensure the denominator is positive and computable, when... and When both are empty, from Removed from:

[0119] ;

[0120] in, For the set of claim items participating in the calculation, For the claim entry Supported collection, For the claim entry The set of conflicts To determine the time weights based on the time windows of the condition set and the time of the claim list, To support the weighting coefficients, Here are the conflict weighting coefficients, and , for The cardinality;

[0121] Through the above steps, we complete the screening of target object evidence under the constraints of the distinguishing condition set, grouping of claim items, construction of support and conflict sets, calculation of time weights and weighted normalization aggregation, obtain item-level consistency scores, and calculate the average to obtain the external consistency value. ,Will Used for subsequent thresholding judgments, ensuring that the output only applies to target objects defined by the distinguishing condition set, and can be directly aligned with the evidence list and evidence judgment results to form a verifiable quantitative basis.

[0122] In this embodiment, step S5 specifically includes:

[0123] semantic graph List of claims As input, the semantic graph generates internal contradiction markers. The list of claims includes time sequence constraints and relational edges. It consists of several claim entries, each claim entry contains subject, predicate, object, time, location, numerical value and type fields. The claim entry is the smallest processing unit, and it runs through the entire process of uniqueness check and time sequence check.

[0124] The semantic graph and the claim list are subjected to field extraction and normalization. For each claim item, the subject, predicate, object, time, location, value, and type are extracted, and the time is unified as a time point. or time window When time is expressed in a relative manner, the reference time in the semantic graph is used for anchoring, and it is transcribed into a time point or time window. The location is converted into a binary representation of hierarchical coding and coordinate boundaries, the numerical value is converted into a closed interval representation, and the type is aligned to a controlled vocabulary to form a set of claim entries for verification.

[0125] The claim set is grouped by subject, predicate, object, and type. A uniqueness check is performed on any two claims within each group. First, the intersection of time windows is checked for non-emptiness. If the intersection is empty, no uniqueness conflict is triggered. If the intersection is not empty, the location and value are further checked. Location inconsistency is determined using a dual rule of hierarchical inclusion and coordinate boundary. Inconsistent locations are determined if the hierarchical levels are different and there is no parent-child relationship, or if the coordinate boundary intersection is empty. Value inconsistency is determined using a value tolerance threshold. Compare absolute differences; the absolute difference is greater than... If the two claims are determined to be inconsistent, and the time windows overlap and either the location or the numerical values ​​are inconsistent, then the two claims are marked as having a unique conflict and associated with the corresponding claim entry within the group.

[0126] Based on the time and relationships in the semantic graph, the preceding and subsequent claims are determined, and a set of sequential constraints is established. Priority is given to using the preceding and inclusion relationships in the semantic graph to determine the order. When two claims have a clear preceding relationship, the former is marked as the preceding claim and the latter as the subsequent claim. When two claims have an inclusion relationship, the order is derived from the start and end relationship of the inclusion window. When the semantic graph does not give a clear relationship and the time windows of the two claims do not intersect, the order is derived from the position of the time window. The above results constitute a set of sequential constraints for subsequent time order checks.

[0127] The time sequence of the priority constraint set is checked. When the priority claim is a time point, its time must not be later than the time point or start time window of the subsequent claim. If it is later, it is judged as a time sequence conflict. When the priority claim is a time window, its end time must not be later than the start time of the subsequent claim. If the end time is later, it is judged as a time sequence conflict. In the case of equal boundaries, the principle of not being later is applied and the claim is approved. Claims with missing time fields are not included in the priority constraint judgment to avoid mislabeling caused by unfounded inference.

[0128] The uniqueness conflict and the time sequence conflict are merged at the claim item level to form an internal contradiction mark. For each claim item, an internal contradiction mark is generated that includes the contradiction type, the index of the claim item involved, the conflict trigger field and the corresponding time window. The internal contradiction mark is attached to the claim list and output uniformly, so that subsequent steps can directly make threshold-based decisions and locate based on the mark.

[0129] Through the above process, uniqueness and time sequence checks based on semantic graphs and claim lists are achieved, making internal contradiction markers interpretable at the field level, verifiable at the grouping granularity, and possessing objective judgment criteria in both the time and numerical dimensions.

[0130] In this embodiment, step S6 specifically includes:

[0131] External consistency values internal contradiction markers As input, in conjunction with the first threshold With the second threshold List of claims List of evidence Distinguishing condition set Generate the final judgment, location information, and chain of evidence. Organize the subject, predicate, object, time, place, value, and type by claim item. The fields and their document locations are stored in a structured manner according to the entries. Define the boundaries of ownership of evidence related to the target object;

[0132] load and The relationship between the two thresholds is checked. Not less than Record the larger one as The smaller one is recorded as To ensure that the threshold relationship meets the requirement that the second threshold is less than the first threshold, after completing the threshold verification, the threshold is frozen for subsequent adjudication.

[0133] in accordance with and Generate the final decision, empty and Not less than The situation is judged as supported, and non-empty or Below The situation is judged as conflict, and the other situations are judged as unknown. The final judgment is a single conclusion at the level of the claim list, without multi-level label superposition, to avoid duplicate judgment with subsequent thresholding calculations.

[0134] Location information is generated based on the final judgment and internal contradiction markers. When the final determination is to support, the location information will be directed to all participants in the claim list. Calculation of claim entries And its time window; when a conflict is ultimately determined, it is resolved first. Add the index of inconsistent claim entries, the triggering time window, and the corresponding time, location, or numerical fields to the list. When multiple conflict sources exist simultaneously, the field paths and values ​​are listed item by item. When the final determination is unknown, based on the screening results of the distinguishing condition set on the evidence list, the claim entries with a zero evidence count for the target object are identified, and the index of the claim entry and its time window are added to the list. The reason for the missing evidence was noted as "the target object's evidence is missing."

[0135] Based on the distinguishing condition set In the list of evidence Internal screening extracts evidence of the target object corresponding to the final judgment, constructing a chain of evidence. When the final determination is support, select those that meet the criteria. All evidence constrained and consistent with the claim list in terms of subject, predicate, object, and location is used to establish a mapping from claim items to evidence items, with the addition of time windows and numerical fields. When a conflict is ultimately determined, evidence that meets the criteria is selected. For all target object evidence that meets the constraints of the claim list dimension and triggers any of the following conditions: opposite predicate meaning, unequal values, or inconsistent locations, establish a corresponding mapping and attach trigger fields and values. When the final determination is unknown, retain the evidence that has passed the test. Claims with a count of zero after screening are identified, and placeholder links are used to create empty evidence prompts. Ambiguity exclusion records are then linked to the aforementioned mapping, creating an exclusion branch in the evidence chain for each excluded candidate and its mismatch condition. Within the boundaries, a complete attribution path is given, and for each piece of evidence in the chain of evidence, an evidence list is attached to point to the source of the literature and the paragraph position;

[0136] Output final judgment and location information With the chain of evidence including the exclusion record and the list of evidence cited. Keep the output consistent with the claim list. The correspondence allows the final judgment to be directly reviewed at the level of the claim items, allows location information to be directly verified at the level of time and field, and allows the chain of evidence to be established. The boundary is fully traced. Through the above steps, the decision-making loop under the constraints of the first and second thresholds is completed, realizing the joint judgment of external consistency value driving and internal contradiction mark constraint. The evidence range is converged by the distinguishing condition set, ensuring that the final judgment, location information and evidence chain output have a unique landing point on the target object.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A large-scale hallucination detection method based on deep learning and semantic structural information, characterized in that, Includes the following steps: S1. Receive the generated text from the language model, perform semantic structure parsing, and generate a semantic graph and a list of claims; S2. Extract the identification feature set based on the semantic graph, retrieve the candidate object set in the knowledge base, determine the distinguishable conditions of each candidate object based on the identification feature set, select the minimum combination of conditions that can be uniquely pointed to in the claim list to form the distinguishable condition set, generate a semantic query plan based on the distinguishable condition set, and record the ambiguity exclusion records of excluded candidates and corresponding conditions. S3. Query the knowledge base and document library according to the semantic query plan, generate an evidence list, divide the evidence into target object evidence and excluded object evidence according to the distinguishing condition set, form the evidence judgment result for the claim list, and output the evidence list and evidence judgment result. S4. Based on the evidence list and the evidence judgment results, calculate the external consistency value on the target object defined by the distinguishing condition set. S5. Perform uniqueness and time sequence checks on the claim list based on the semantic graph, and generate internal contradiction markers; S6. Based on the external consistency value and the internal contradiction marker, generate the final judgment and location information under the constraints of the first threshold and the second threshold, and output the evidence chain including the ambiguity exclusion record and the evidence list reference.

2. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S1 specifically refers to: The generated text from the language model is segmented into sentences and the text boundaries are identified to obtain sentence sequences and paragraph orders; Perform part-of-speech tagging and dependency analysis on the sentence sequence to determine the positions of predicates and arguments, and generate intra-sentence relations; Perform named entity recognition and event extraction, extracting roles, organizations, locations, times, and values ​​to form entity and event candidates; Performs referential resolution and cross-sentence linking, merges the same entity and the same event, and normalizes time expressions into time points or time windows; Construct a semantic graph with entities and events as nodes and relationships and roles as edges, and add time, location, numerical and type attributes, and write time order constraints; Set a confidence threshold to remove nodes and relationships in the semantic graph with a confidence level below the threshold, and retain components with a confidence level not lower than the threshold. Extract a list of claims from the semantic graph, organize the subject, predicate, object, time, location, numerical value and type constraints into claim items, remove duplicate claims, and output the semantic graph and the list of claims.

3. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S2 specifically refers to: Using semantic graphs and claims lists as input, we extract identification feature sets and standardize the representations of roles, organizations, locations, times, values, and types to form an alignable identification feature set. Based on the identification feature set, a candidate object set is retrieved from the knowledge base. Field matching and time window overlap determination are used for initial screening to remove objects that do not overlap in time or are inconsistent in type, thus obtaining a candidate object set that meets the semantic boundaries of the claim list. For each candidate object in the candidate object set, determine the distinguishable conditions, align the role, organization, location, time, value and type with the identification feature set, and set the value range and matching method for each condition. Among them, time and type are mandatory constraints, and if they are not met, they are marked as unusable conditions. Perform minimum condition combination filtering on the candidate object set. With the goal of minimizing the number of conditions, add time, type, location, organization, role, and value in a fixed priority order. Once added, it is determined to be unique. The determination of uniqueness is based on the number of objects in the candidate object set that meet all the added conditions. If the condition is not met, continue to increment the combination until it is met. The minimum combination of conditions that achieves unique pointing is summarized into a distinguishing condition set. For candidate objects that are not uniquely pointed to, ambiguity exclusion records are generated based on the mismatch relationship with the distinguishing condition set. The ambiguity exclusion record includes the excluded candidate, the name of the mismatch condition, the mismatch value, and the mismatch time window, which is used to clarify the reason for exclusion. Based on the distinguishing condition set, a semantic query plan is generated, which is then mapped to executable fields and filtering conditions for knowledge base query and document retrieval. These include time window filtering, location range filtering, type consistency filtering, numerical range filtering, and role matching rules, forming an executable field list and filtering logic. Perform field alignment and boundary checks on the semantic query plan to ensure that the fields and filtering conditions correspond to the distinguishing condition set, that the time window does not exceed the time range of the claim list, and that the types are consistent. This completes the execution of the semantic query plan and outputs the distinguishing condition set, the semantic query plan, and the ambiguity exclusion record.

4. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S3 specifically refers to: The semantic query plan is mapped to a retrieval request for the knowledge base and document library, the retrieval is executed, the returned content is parsed, the subject, predicate, object, time, location, numerical and type fields are extracted, and the evidence content in the same paragraph as the field is located to form candidate entries. Based on the semantic query plan, candidate items are filtered by time window and type constraints. Candidate items whose time falls within the time window of the semantic query plan and whose type is consistent are included in the evidence list. The field representation is unified so that the evidence list and the claim list are associated with each other in terms of subject, predicate, object, time, location, value and type. The evidence list is classified according to the differentiation criteria set. Items that meet all the criteria in the differentiation criteria set are classified as target object evidence, while items that do not meet any of the mandatory criteria of time or type are classified as excluded object evidence. The remaining mismatched items of location, institution, role, and value are compared with the remaining criteria in the differentiation criteria set and classified as excluded object evidence. Perform consistency comparison on the evidence of the target object, match the subject, predicate, object and location of the claim list with the evidence of the target object item by item, and perform an equality judgment on the value and the value of the claim list under the premise that the time of the evidence is included in the time window of the distinguishing condition set and the type is consistent. If all the above matches are satisfied, it is determined to be supported. Conflict detection is performed on the evidence of the target object. Under the premise that the evidence time is included in the time window of the distinguishing condition set and the type is consistent, if any of the following situations occur, such as opposite predicate meanings, unequal values, or inconsistent locations, it is determined to be a conflict. When both supporting and conflicting items appear in the evidence of the target object, the conflict shall prevail. When the evidence for the target object is empty and the evidence list only contains the evidence for the excluded object, it is determined to be unknown. When the evidence for the target object is empty and the evidence list is empty, it is determined to be unknown. The evidence for the excluded object does not participate in the generation of support and conflict. The supporting, conflicting, and unknown claims are summarized in the claim list dimension to form the evidence judgment result, and the evidence list and evidence judgment result are output.

5. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S4 specifically refers to: The evidence list, evidence judgment results, and differentiation condition set are used as input. The evidence attribution is limited according to the differentiation condition set. Items that meet all constraints of the differentiation condition set are selected from the evidence list as target object evidence. Items that do not meet any of the time and type constraints are removed to obtain the target object evidence set used for calculation. Based on the claim list, the evidence of the target object is grouped, a mapping relationship between claim items and evidence of the target object is established, a set of evidence of the target object is determined for each claim item for calculation, and time, location and numerical fields are reserved for each piece of evidence for subsequent weight calculation; Based on the evidence determination results, the support set and conflict set are determined at the claim item level. When the evidence determination result is supportive, the target object evidence of the claim item is recorded in the support set. When the evidence determination result is conflicting, the target object evidence of the claim item is recorded in the conflict set. When the evidence determination result is unknown, it is not included in the calculation. For each piece of evidence of the target object in the support set and conflict set, a time weight is calculated. The time distance is determined by the time window that distinguishes the condition set and the time of the claim list. The corresponding weight is given according to the preset time segment threshold. A higher weight is given for a shorter time distance and a lower weight is given for a longer time distance. Set support weight coefficient and conflict weight coefficient. The conflict weight coefficient is greater than the support weight coefficient. Before calculation, check the matching completeness of subject, predicate, object and location. Only when all the above fields match, contribution and penalty are included. Incomplete matching of target object evidence is not included in the calculation. Calculate an item-level consistency score for each claim item. The time weights of the supporting set are accumulated according to the support weight coefficient, and the time weights of the conflict set are accumulated according to the conflict weight coefficient. The difference between the two is mapped to an item-level consistency score according to a preset standardization rule. The item-level consistency score ranges from zero to one. The external consistency value is obtained by averaging the item-level consistency scores of all claim items with equal weights, and then output on the target object defined by the distinguishing condition set.

6. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S5 specifically refers to: Using semantic graphs and a list of claims as input, the subject, predicate, object, time, location, value, and type of each claim item are extracted, and the time is unified into a time point or time window to form a set of claim items for verification. The claim set is grouped by subject, predicate, object and type. A uniqueness check is performed within each group. If any two claims have overlapping time windows and the locations or values ​​are inconsistent, they are judged to be in uniqueness conflict. Based on the time and relationships in the semantic graph, the preceding and subsequent claims are determined for each group, and a set of sequential constraints is established. The determination of the preceding and subsequent claims is based on the start and end relationships of the claim time and the semantic graph relationships. Perform a time order check on the set of sequential constraints. If the preceding claim is a time point, its time must not be later than the time of the subsequent claim. If the preceding claim is a time window, its end time must not be later than the start time of the subsequent claim. If the conditions are not met, it is determined to be a time order conflict. The uniqueness conflict and the chronological order conflict are merged at the claim item level to generate an internal contradiction marker, and the internal contradiction marker is output on the claim list.

7. The large-model hallucination detection method based on deep learning and semantic structural information according to claim 1, characterized in that, S6 specifically refers to: Using external consistency values ​​and internal contradiction markers as inputs, load the first threshold and the second threshold, check the relationship between the two thresholds, and confirm that the second threshold is less than the first threshold. The final judgment is generated based on the external consistency value and the internal contradiction mark. When the external consistency value is not lower than the first threshold and the internal contradiction mark is empty, it is judged as support. When the external consistency value is lower than the second threshold or the internal contradiction mark is not empty, it is judged as conflict. All other cases are judged as unknown. Location information is generated based on the final judgment and internal contradiction markers. For support, the claim item and time window are located; for conflict, the claim item where the inconsistency occurred and the time, location or numerical field are located; for unknown, the claim item with missing evidence is located. Based on the set of distinguishing conditions, the evidence of the target object corresponding to the final judgment is screened from the evidence list. The evidence chain is constructed by combining the ambiguity elimination record, and the final judgment, location information and evidence list references are established to correspond. Output the final judgment, location information, and the chain of evidence including the ambiguity exclusion record and the evidence list cited by the evidence chain.

Images