Dynamic retrieval and credibility evaluation method for reducing large model illusion

Abstract

The invention relates to the field of data information processing, and discloses a dynamic retrieval and credibility evaluation method for reducing large model illusion, which comprises the following steps of: S1, setting a semantic segmentation condition for an input text and executing semantic unit analysis; s2, setting an uncertainty calculation rule and calculating an uncertainty value of each semantic element; s3, setting a retrieval priority based on the target element set; s4, setting an element extension condition according to a retrieval result and introducing an associated element; s5, generating an association strength graph according to the credibility of the extended evidence, and executing causal reasoning according to the association strength graph; and S6, setting a correction range and judging a correction area according to the reasoning credible factor. Semantic fragment division, dependency direction judgment and conflict reconstruction mechanisms are introduced in a semantic analysis stage, and uncertainty quantification and hierarchical retrieval trigger strategies are combined, so that the effect of effectively preventing the model from generating unreferenced inference when information is lost or understanding deviation occurs is achieved, and the occurrence of illusion is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of data information processing, specifically to a dynamic retrieval and credibility assessment method for reducing large model illusion. Background Technology

[0002] Currently, large models have achieved remarkable success in natural language processing tasks, but they have also exposed the "illusion" problem, where the content generated by the model is inconsistent with the facts or logically flawed. Addressing this issue is crucial for improving the reliability of model output and user experience.

[0003] Existing solutions typically rely on increasing the amount of training data for the model, introducing external knowledge bases, or adjusting the model architecture, semantic parsing, and retrieval triggering strategies. However, most methods fail to accurately identify the dependency direction and conflict relationship between semantic segments during the input semantic structure analysis stage. This leads to the model making unfounded logical inferences when information is missing or semantics are ambiguous, resulting in factual errors or fictitious information. Traditional uncertainty judgment mechanisms often use a single threshold to trigger retrieval, lacking dynamic grading and adaptive adjustment capabilities, making it difficult to perform fine-grained control based on semantic complexity and contextual risks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a dynamic retrieval and credibility assessment method for reducing large model illusions. This method solves the problem that introducing external knowledge bases increases model complexity and real-time response latency, thus limiting the application of large models in real-world scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A dynamic retrieval and credibility assessment method for reducing large model illusion includes the following steps:

[0007] S1. Set semantic segmentation conditions for the input text and perform semantic unit parsing. Calculate the correlation between semantic units and construct an initial correlation structure to determine the basic semantic range.

[0008] S2. Set uncertainty calculation rules and calculate the uncertainty value of each semantic element. Determine the set of target elements that need to be triggered by external retrieval based on the uncertainty value.

[0009] S3. Set the search priority based on the target element set, trigger dynamic search scheduling according to the priority, set the search source and search method according to the scheduling result, and execute the search operation.

[0010] S4. Based on the search results, set the element expansion conditions and introduce related elements to form an expanded element set, construct the evidence structure, and set the evidence credibility according to the reliability rules.

[0011] S5. Generate a correlation strength map based on the credibility of extended evidence, and perform causal inference based on the correlation strength map to calculate the inference credibility factor of the candidate generated content.

[0012] S6. Set the correction range and determine the correction area based on the reasoning credibility factor, perform text correction operation and audit the correction content, and record the audit information for traceability.

[0013] By adopting the above technical solutions, a multi-source knowledge base is suddenly established, integrating encyclopedias, professional literature, and internet information to form a comprehensive knowledge system. During the text generation process of the large model, the generated content is compared with information in the multi-source knowledge base in real time. Through semantic similarity calculation, the most relevant knowledge fragments are selected. A credibility assessment algorithm is designed, combining dynamic retrieval results and model generation probabilities to comprehensively evaluate the credibility of the generated content. Low-credibility content will be marked and corrected or regenerated. Adaptive learning, through a feedback mechanism, continuously optimizes the knowledge base and assessment algorithm, enabling the model to gradually reduce illusions and improve output quality.

[0014] As a further description of the above technical solution: S1 includes the following steps:

[0015] Define the granularity of semantic segmentation;

[0016] Determine the dependency direction between semantic segments based on changes in the correlation quantity;

[0017] When a conflict occurs in the dependency direction, a related structure reconstruction operation is triggered.

[0018] The reconstruction is used to maintain the stability and consistency of the semantic parsing results.

[0019] By adopting the above technical solutions, semantic fragment segmentation and dependency identification can be completed in a more detailed and controllable manner, so that the semantic structure remains stable during the parsing process. When a conflict in the dependency direction occurs, the structure can be adjusted in time, thereby avoiding semantic understanding bias, improving the accuracy and coherence of the overall semantic parsing, and providing reliable basic semantic structure support for subsequent uncertainty judgment, retrieval triggering and causal reasoning.

[0020] As a further description of the above technical solution: S2 includes the following steps:

[0021] Calculate the missing, ambiguity, and conflict measures of semantic elements;

[0022] Uncertainty levels are set based on three types of metrics;

[0023] The search decision is triggered based on the correspondence between uncertainty level and threshold.

[0024] Different levels correspond to different search trigger strengths.

[0025] By adopting the above technical solutions, the uncertainty of semantic elements can be more precisely quantified from three dimensions: missing, ambiguous, and conflict, and a graded uncertainty judgment mechanism can be established accordingly. Different levels correspond to different search trigger intensities, so that the search behavior will not over-trigger and waste resources, nor will it miss key missing information. Thus, through the accuracy and adaptability of search decisions, it helps to significantly improve the reliability and information completeness of subsequent content generation.

[0026] As a further description of the above technical solution: S3 includes the following steps:

[0027] Set up a retrieval priority scheduling table and match priorities based on the type of target element;

[0028] The selection of search sources is dynamically adjusted based on search feedback.

[0029] The search method is switched based on changes in the search source.

[0030] Priority scheduling is used to improve retrieval coverage and consistency.

[0031] By adopting the above technical solutions, the importance of target semantic elements can be used to achieve hierarchical scheduling of retrieval tasks, making the investment of retrieval resources more focused. Furthermore, the retrieval source and retrieval method can be dynamically adjusted according to real-time retrieval feedback, thereby ensuring that the retrieval path has flexibility and adaptability in different information environments.

[0032] As a further description of the above technical solution: S4 includes the following steps:

[0033] Set evidence mapping rules;

[0034] Calculate the matching degree between evidence and semantic elements based on mapping rules;

[0035] The credibility of the evidence is adjusted by weighting based on changes in the matching degree.

[0036] The weighted correction is used to highlight the contribution of highly relevant evidence.

[0037] By adopting the above technical solutions, a quantifiable matching relationship can be established between evidence and semantic elements, enabling the information obtained from external retrieval to be accurately evaluated based on relevance. By weighted correction of changes in matching degree, the contribution of highly relevant evidence can be effectively highlighted and the influence of low-value or noisy evidence can be suppressed. This not only significantly improves the accuracy of evidence selection but also enhances the reliability and stability of the evidence structure in subsequent causal reasoning, thereby providing a more solid basis for the final generated content.

[0038] As a further description of the above technical solution: S5 includes the following steps:

[0039] Set a threshold for association strength and filter strongly associated paths;

[0040] Construct causal chains based on strongly related paths;

[0041] Detect conflict nodes in the causal chain and perform conflict resolution;

[0042] Conflict resolution is used to maintain the monotonicity and consistency of the reasoning path.

[0043] By adopting the above technical solutions, the relationships between semantic elements can be hierarchically filtered, enabling the reasoning process to prioritize the construction of causal chains based on strongly related paths, thereby improving the effectiveness and reliability of the reasoning chain. On this basis, by automatically detecting and resolving conflict nodes in the causal chain, the reasoning process can be prevented from breaking or deviating due to contradictory information, thus maintaining the monotony and logical consistency of the reasoning path as a whole.

[0044] As a further description of the above technical solution: S6 includes the following steps:

[0045] Set the correction strategy table;

[0046] The correction method of replacement, insertion, or expansion is selected based on the reasoning credibility factor;

[0047] The revised content was subjected to a consistency review in accordance with auditing rules;

[0048] Among them, consistency review is used to ensure that the revised content and the evidence structure remain logically consistent.

[0049] By adopting the above technical solution, the inference results can be modified in a targeted manner based on the preset correction strategy table, so that the correction methods such as replacement, insertion or expansion are matched with the credibility factor of the inference. This ensures the necessary correction while avoiding excessive changes. Combined with the audit rules, the correction content is reviewed for consistency. It can verify in real time whether the correction is consistent with the constructed evidence structure and prevent the introduction of new contradictions or content that deviates from the evidence support due to the correction.

[0050] As a further description of the above technical solution: S6 also includes the following steps:

[0051] Define the confidence decay function for the inference results;

[0052] Dynamically smooth the inference confidence factor based on the decay function;

[0053] A secondary search is triggered when the smoothed confidence factor falls below the target threshold.

[0054] Secondary retrieval is used to improve the reliability of low-confidence regions.

[0055] By adopting the above technical solution, the confidence decay function can be used to dynamically smooth the changing trend of the reasoning confidence factor, enabling the system to more accurately identify the areas where the confidence decreases in the reasoning chain. When the smoothed confidence factor is lower than the target threshold, a secondary retrieval is automatically triggered, which can timely supplement the low-confidence segment with external information support, thereby effectively improving the reliability and stability of weak reasoning links.

[0056] As a further description of the above technical solution: Based on the audit information records, a structured format for the audit information records is set, including:

[0057] The input information, evidence information, and reasoning information during the correction process are recorded in a structured manner according to the format.

[0058] A timeline is set in the structured record to trace the sequential relationship of the generation steps;

[0059] The time link is used for tracking and management.

[0060] By adopting the above technical solutions, the input information, evidence information, and reasoning information in the correction process can be systematically recorded in a structured format, making the audit process clearer and more efficient. Setting a timeline to record the sequential relationship of the generation steps allows for precise tracing of the decision-making process and data changes at each stage, ensuring the transparency and verifiability of the audit trail.

[0061] A dynamic retrieval and credibility assessment system for reducing large model illusion, the system comprising:

[0062] The semantic parsing module is used to perform semantic segmentation and construct related structures;

[0063] The uncertainty determination module is used to calculate the uncertainty and trigger the retrieval.

[0064] The retrieval scheduling module is used to set the retrieval method and execute dynamic retrieval.

[0065] The evidence construction module is used to generate an expanded set of elements and evidence credibility.

[0066] The inference module is used to generate association strength maps and perform causal inference.

[0067] The Correction and Audit module is used to perform text corrections and record audit information.

[0068] By adopting the above technical solutions, a collaborative closed loop can be formed in multiple stages such as semantic parsing, uncertainty determination, dynamic retrieval, evidence construction, causal reasoning, and correction auditing. This ensures that the generation process of the large model is always controlled by quantifiable uncertainty and credibility constraints. By introducing dynamic retrieval when key information is missing or confidence decreases, and combining evidence credibility assessment and causal reasoning verification, inferences without evidence support can be effectively suppressed, thereby significantly reducing the probability of large model illusions.

[0069] This invention provides a dynamic retrieval and credibility assessment method for reducing large model illusions. It has the following beneficial effects:

[0070] 1. In this invention, by introducing semantic segmentation, dependency direction determination and conflict reconstruction mechanisms in the semantic parsing stage, and combining uncertainty quantification and hierarchical retrieval triggering strategies, the invention effectively avoids unfounded inferences by the model when information is missing or there is a misunderstanding, thereby significantly reducing the occurrence of hallucinations.

[0071] 2. In this invention, by adopting a dynamic retrieval scheduling and evidence credibility weighted evaluation mechanism, and combining a causal reasoning system of association strength graph construction, strong association path screening and conflict resolution, the key knowledge is supplemented in real time during the generation process and the consistency of the reasoning path is maintained, thereby improving the reliability and logical rigor of the generated content.

[0072] 3. In this invention, by setting a correction strategy table, a credibility factor decay function, and a secondary retrieval triggering mechanism, and by implementing structured audit records and consistency checks on the correction content, the effects of continuously optimizing retrieval strategies and inference rules using feedback data, system adaptive learning, and long-term performance improvement are achieved. Attached Figure Description

[0073] Figure 1 This is a flowchart of the method of the present invention;

[0074] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0075] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0076] To better understand the above technical solutions, the technical solutions of the present invention will be clearly and completely described below in conjunction with embodiments.

[0077] Example 1

[0078] Reference Figure 1 In a first embodiment of the present invention, the present invention provides a dynamic retrieval and credibility assessment method for reducing large model illusions, comprising the following steps:

[0079] S1. Set semantic segmentation conditions for the input text and perform semantic unit parsing. Calculate the correlation between semantic units and construct an initial correlation structure to determine the basic semantic range.

[0080] S1 includes the following steps:

[0081] Define the granularity of semantic segmentation;

[0082] Determine the dependency direction between semantic segments based on changes in the correlation quantity;

[0083] When a conflict occurs in the dependency direction, a related structure reconstruction operation is triggered.

[0084] Reconstruction is used to maintain the stability and coherence of semantic parsing results.

[0085] Specifically, firstly, the granularity of semantic segmentation is set. After the input text is processed by sentence segmentation, word segmentation, and entity recognition, an initial set of segments is obtained. By setting a minimum length threshold With maximum length threshold For excessively short segments, a merging process is performed, and for excessively long segments, a splitting process is performed, so that each semantic segment maintains a balance between information density and semantic integrity, thereby forming a set of standard semantic units for subsequent association calculations.

[0086] Determine the dependency direction between semantic segments based on changes in the correlation quantity, for any semantic segment Its vector representation is obtained through the encoding model. And after normalization, we obtain The correlation between any two semantic segments It consists of a similarity term and a statistical term, and its calculation formula is as follows:

[0087] in, Represents semantic fragments point to The strength of the association; This is the normalized vector; This represents the number of times the two segments co-occur. Let be the weighting coefficient, satisfying To prevent the removal of the zero constant.

[0088] In obtaining bidirectional correlation volume and Then, the dependency direction is determined using the difference threshold method, and the determination formula is as follows:

[0089] ;

[0090] in, For the direction of dependence; This is the threshold for direction determination. When... When, it means Depends on when -1 indicates Depends on when When the direction is uncertain, it indicates that the direction is uncertain.

[0091] Finally, a restructured association operation is triggered when a dependency conflict occurs. A directed association structure is constructed based on the direction determination results. If a directed cycle or a high-strength bidirectional dependency is detected, the current association structure is deemed unstable. Conflicts are identified through the following conditions:

[0092] ;

[0093] in, The average edge weight of the conflict cycle; This represents the number of edges in the ring; ; This is the conflict threshold. When the condition is met, the associated structure is restructured.

[0094] During the reconstruction process, the conflict edges are re-evaluated in weight. The complete formula is as follows:

[0095] ;

[0096] in, The updated edge weights; For memory coefficient; The recalculated value is based on the joint calculation of semantic similarity and syntactic strength. If the semantic entropy of a semantic segment exceeds a set threshold and its length is greater than the splitting threshold, then the segment is semantically split. If multiple segments exhibit high consistency clustering characteristics, then node merging is performed. The reconstruction operation is used to maintain the stability and coherence of the semantic parsing results. By adaptively adjusting the conflicting dependencies, the final semantic association structure satisfies the requirements of directional consistency and structural acyclicity, thereby ensuring that the subsequent semantic reasoning process is carried out on the basis of a stable structure and avoiding the drift of the overall parsing results due to local misjudgment.

[0097] S2. Set uncertainty calculation rules and calculate the uncertainty value of each semantic element. Determine the set of target elements that need to be triggered by external retrieval based on the uncertainty value.

[0098] S2 includes the following steps:

[0099] Calculate the missing, ambiguity, and conflict measures of semantic elements;

[0100] Uncertainty levels are set based on three types of metrics;

[0101] The search decision is triggered based on the correspondence between uncertainty level and threshold.

[0102] Different levels correspond to different search trigger strengths.

[0103] Specifically, firstly, the missing, ambiguity, and conflict measures of semantic elements are calculated. For semantic units... Calculate missing metrics Its formula is as follows:

[0104] ;

[0105] in, To find the fragment in the semantic knowledge base or the current text The number of valid semantic elements matched; This represents the upper limit of matchable elements. To prevent division by zero of constants. The higher the value, the more semantic information is missing from the segment.

[0106] Quantifying semantic ambiguity, ambiguity measurement Based on the definition of the discreteness of semantic vector distribution, its complete formula is:

[0107] ;

[0108] in, For fragment vectors; The cluster center vector; Belongs to cluster probability weights; The number of clusters; This is the normalization factor. If... A large value indicates that the fragment may have multiple semantic interpretations, i.e., significant ambiguity. A method for calculating the conflict metric is given. The formula for measuring the degree of contradiction in the semantic structure of a segment is as follows:

[0109] ;

[0110] in, To and A set of adjacent segments that are semantically related; and For bidirectional correlation; add to the denominator To prevent numerical anomalies, if the bidirectional correlations around a segment differ significantly, the conflict metric of that segment will increase.

[0111] Secondly, uncertainty levels are set based on the three types of metrics. The three indicators are combined to define the fragment-level uncertainty. The calculation formula is as follows:

[0112] ;

[0113] in, For the weighting coefficients, satisfying This can be adjusted according to the data characteristics. Subsequently, based on the uncertainty value, it is mapped to different levels: low, medium, relatively high, and high. The level boundaries are determined by parameters. Decision, when This indicates the highest level of uncertainty; when For higher-level cases, and so on, the significance of this hierarchical approach is that it enables the system to cope with different types of uncertainty in a more flexible way, not just relying on a single threshold, but using interval-based judgment, making semantic compensation processing more resolution.

[0114] The retrieval decision is triggered based on the correspondence between uncertainty levels and thresholds. When the uncertainty level is high or above, the system will invoke the retrieval decision module to trigger external information retrieval, background knowledge completion, or internal cache query. The decision criteria can be expressed as follows:

[0115] ;

[0116] in, Indicates whether to trigger a search. It can be equal to the high uncertainty threshold, or it can be set to a slightly lower value based on experience to enhance the sensitivity of the system;

[0117] Different levels correspond to different retrieval trigger intensities. The highest level can trigger multi-source parallel retrieval, higher levels trigger single-source retrieval, medium levels trigger lightweight queries, and low levels can be left untriggered. This controls retrieval resource consumption while ensuring compensation capabilities in critical scenarios. The above processing enables the system to promptly identify and adopt adaptation strategies when faced with complex situations such as semantic missing, ambiguity, or conflict, thereby maintaining the overall effectiveness of semantic parsing.

[0118] S3. Set the search priority based on the target element set, trigger dynamic search scheduling according to the priority, set the search source and search method according to the scheduling result, and execute the search operation;

[0119] S3 includes the following steps:

[0120] Set up a retrieval priority scheduling table and match priorities based on the type of target element;

[0121] The selection of search sources is dynamically adjusted based on search feedback.

[0122] The search method is switched based on changes in the search source.

[0123] Priority scheduling is used to improve retrieval coverage and consistency.

[0124] Specifically, firstly, a retrieval priority scheduling table is set up, and a priority vector set is constructed.

[0125] ;

[0126] Used to describe the order and weight of calls to different search sources. Any search source Priority score Its calculation formula can be expressed as:

[0127] ;

[0128] in, Assess the matching quality of this search source in historical tasks; For the availability of the search source (e.g., latency, stability, etc.); To assess the relevance of the content to the current task requirements; For the weighting coefficients, satisfying The calculated priority vector is used to build a scheduling table, which plays a crucial scheduling role throughout the retrieval phase.

[0129] Matching priority is based on the type of target element. Different semantic elements, including entities, events, attributes, and background knowledge, have different search preferences. Define its type vector Priority matching is based on the following formula:

[0130]

[0131] in, For retrieval source capability feature vector; The similarity function can be cosine similarity. Finally, the priorities in the scheduling table are updated with weights based on the matching scores to obtain the new priorities:

[0132] in, This is an adjustment factor used to control the magnitude of the matching effect.

[0133] When the system retrieves from the source When feedback is received, the quality of the results is evaluated and recorded as a feedback score. This feedback score is used to promptly correct the scheduling table, causing it to converge to a retrieval path more suitable for the current context. The formula is adjusted as follows:

[0134] ;

[0135] in, This is a memory factor used to balance historical priority with real-time feedback. If the retrieval feedback remains low-quality, the priority of that source will gradually decrease and it may be removed from the main retrieval path; conversely, it will be promoted to a higher position. It also utilizes a retrieval consistency metric. Auxiliary scheduling update:

[0136] ;

[0137] in, This represents the degree of alignment between a single retrieval and the target semantics. This is the length of the statistical window. If consistency is low, the system tends to reduce the frequency of calls to that search source.

[0138] The search method will be switched based on changes in the search source. Search methods include lightweight search, deep search, cross-source search, and multi-round search. The system will switch search methods when one of the following conditions is met:

[0139] ;

[0140] or

[0141] ;

[0142] in, This refers to the change in priority. This is the trigger threshold; This is the consistency threshold. If any condition is met, the system determines that the retrieval source has changed significantly and needs to switch retrieval strategies, from single-source lightweight retrieval to multi-source parallel retrieval; from vector retrieval to knowledge graph retrieval; or from single-hop retrieval to multi-hop reasoning retrieval. The retrieval method switch triggered after the change makes the system have higher coverage when facing semantic gaps, while maintaining the semantic consistency of the output results and avoiding deviations caused by a single path or a single knowledge source.

[0143] S4. Based on the search results, set the element expansion conditions and introduce related elements to form an expanded element set, construct the evidence structure, and set the evidence credibility according to the reliability rules.

[0144] S4 includes the following steps:

[0145] Set evidence mapping rules;

[0146] Calculate the matching degree between evidence and semantic elements based on mapping rules;

[0147] The credibility of the evidence is adjusted by weighting based on changes in the matching degree.

[0148] The weighted correction is used to highlight the contribution of highly relevant evidence.

[0149] Specifically, by setting evidence mapping rules, evidence information retrieved externally or internally is mapped to specific semantic elements. Then, the matching degree of the evidence to the corresponding semantic elements is calculated according to the mapping rules, and the credibility of the evidence is weighted and adjusted according to the dynamic changes in the matching degree, so as to highlight the evidence that is highly relevant to the target semantics, thereby reducing the influence of noisy evidence and enhancing the reliability and directionality of the final reasoning link;

[0150] First, establish the evidence mapping rules. The evidence set is denoted as... The set of target semantic elements is denoted as The mapping rule consists of a semantic similarity term, a structural consistency term, and a contextual association term, and its mapping function can be expressed as:

[0151] ;

[0152] in, For semantic similarity, it is usually based on vector cosine similarity; Structural consistency score (e.g., whether they appear in the same event frame or dependency subtree structure similarity); For context co-occurrence; For weighted parameters, satisfying The above mapping rules are used to determine the basic association strength between evidence and semantic elements, providing a unified entry point for subsequent matching degree calculation;

[0153] Secondly, the matching degree between evidence and semantic elements is calculated based on mapping rules. With semantic elements Matching degree Defined as:

[0154] ;

[0155] The denominator is used for normalization. To prevent division by zero, the formula ensures that the distribution of each piece of evidence is comparable across all target elements. The system also incorporates a score reflecting the completeness of the evidence itself. The adjusted matching degree is obtained as follows:

[0156] ;

[0157] in, This is an adjustment coefficient used to incorporate consideration of the quality of the evidence itself, so that higher-quality evidence receives a higher base weight in the matching degree calculation.

[0158] Finally, a weighted adjustment is performed on the credibility of the evidence based on the change in matching degree. The credibility of the evidence is denoted as... Its update method is as follows:

[0159] ;

[0160] in, For memory factors, This represents the match value of evidence to the most relevant semantic element. This allows the credibility of evidence to be correlated with its contribution to the current semantic interpretation, reducing the involvement of weakly relevant evidence.

[0161] To highlight highly relevant evidence, a weighted enhancement mechanism is further introduced. When the matching degree of a piece of evidence to any semantic element exceeds an enhancement threshold... At that time, the strengthening weight is applied to the evidence:

[0162] ;

[0163] in, For the Heaviside step function, when The value is 1 if the condition is met, otherwise it is 0. To enhance the coefficient, this enhancement operation is used to create a significant relevance gradient within the evidence set, giving truly valid evidence greater influence, thereby improving the stability and directional orientation of the overall reasoning chain.

[0164] S5. Generate a correlation strength map based on the credibility of extended evidence, and perform causal inference based on the correlation strength map to calculate the inference credibility factor of the candidate generated content.

[0165] S5 includes the following steps:

[0166] Set a threshold for association strength and filter strongly associated paths;

[0167] Construct causal chains based on strongly related paths;

[0168] Detect conflict nodes in the causal chain and perform conflict resolution;

[0169] Conflict resolution is used to maintain the monotonicity and consistency of the reasoning path.

[0170] Specifically, firstly, a threshold for association strength is set, and strongly associated paths are filtered for any semantic unit. and The strength of the correlation between Set a threshold Used to filter out strongly related paths, specifically defined as:

[0171] path ;

[0172] in, This represents the set of all strongly associated paths. This set includes all paths with an association strength greater than a threshold. The paths represent significant dependencies between semantic units. These paths will be used to construct subsequent causal chains.

[0173] Construct causal chains based on strongly associated paths, and then use the selected set of strongly associated paths... Constructing causal chains Each of the causal chains Represents a directed path, indicating from arrive The causal relationship is established by traversing all strongly correlated paths, gradually building a causal chain for any two segments. and ,fruit Then establish The causal chain is established, and more related fragments are connected through graph traversal algorithms, including depth-first search or breadth-first search.

[0174] Then, conflict nodes are detected and conflict resolution is performed in the causal chain. After the causal chain is constructed, the system will check whether there are conflict nodes in the causal chain. Conflict nodes refer to mutually contradictory dependencies, which are more complex than those mentioned above. and To detect conflicting nodes, this embodiment defines conflict detection metrics. The degree of conflict in the dependency direction is calculated using the following formula:

[0175] ;

[0176] when When a conflict is detected, the system will resolve the conflicting nodes. Resolution operations include adjusting the direction of the causal path, deleting the conflicting path, or reconstructing the causal chain based on semantic reasoning rules. The specific resolution process is as follows:

[0177] Conflicting nodes are sorted, and those that have the greatest impact on the reasoning path are resolved first.

[0178] According to the weighting coefficients Select the optimal path after resolution. If a path does not affect the monotonicity and consistency of the reasoning after resolution, then retain that path;

[0179] If the conflict cannot be resolved, then abandon the causal path and rebuild a new causal chain.

[0180] Finally, conflict resolution is used to maintain the monotonicity and consistency of the reasoning path. During conflict resolution, the monotonicity of the reasoning path is ensured, meaning that forward reasoning in the causal chain always maintains a consistent direction. The resolved causal chain should satisfy the following monotonicity condition:

[0181] if ,thentime time ;

[0182] in, Represents a node in a causal chain The reasoning sequence is timestamped. This ensures that causal relationships do not form loops during the reasoning process, maintaining the monotonicity of the reasoning path.

[0183] S6. Set the correction range and determine the correction area based on the reasoning credibility factor, perform text correction operation and audit the correction content, and record the audit information for traceability.

[0184] S6 includes the following steps:

[0185] Set the correction strategy table;

[0186] The correction method of replacement, insertion, or expansion is selected based on the reasoning credibility factor;

[0187] The revised content was subjected to a consistency review in accordance with auditing rules;

[0188] Among them, consistency review is used to ensure that the revised content and the evidence structure remain logically consistent.

[0189] Specifically, first, a correction strategy table is set up, denoted as [Table content missing]. Each of them Describe a possible semantic correction method: Replace, Insert, or Extend.

[0190] For each strategy Set its strategy factor vector 2 , , in, Policy fit measures the accessibility of the policy with the current inference structure; The risk coefficient represents the additional uncertainty that the strategy may introduce. The policy depth indicates the scope (local / global) of the effect of the correction policy within the semantic structure. Correction Policy Table It provides the rule space for the entire correction system. Secondly, during the reasoning process, each node generates a credibility factor to select the correction method of replacement, insertion, or expansion based on the reasoning credibility factor. This is used to represent the reliability of a node. The reliability factor is mapped to a modified mode preference according to the following formula:

[0191] ;

[0192] in, Node trusted feature vector For strategy features The similarity function includes cosine.

[0193] The final strategy for selecting the highest score:

[0194] ;

[0195] like The corresponding replacement operation will change the node. Replace with nodes of the same type but with higher reliability. If the operation corresponds to insertion, logical completion units are inserted before and after the node; if the operation corresponds to expansion, more auxiliary fragments are introduced from the evidence structure to expand the reasoning chain. Adaptive correction of the reasoning structure and consistency checks are used to ensure that the corrected semantic structure does not compromise the logical validity of the evidence chain. A consistency score is defined. The calculation method is as follows:

[0196] Cons Coher Among them, Cons Represents the correction node and the evidence set Consistency; Coher Represents the node and the current causal chain. Continuity; These are the weighting coefficients.

[0197] If the audit results meet the following:

[0198] ;

[0199] The amendment is then considered passed;

[0200] If the conditions are not met, a rollback mechanism will be implemented to undo the current modification and select a new option. It also includes setting up conflict detection items to review whether the corrected fragment introduces structural conflicts:

[0201] ;

[0202] like If the correction is not valid, it must be rejected or adjusted. The consistency review ensures that the final reasoning structure remains consistent in logic, directionality, and evidentiary relevance, and that the correction operation does not disrupt the monotonicity and stability of the reasoning chain.

[0203] S6 also includes the following steps:

[0204] Define the confidence decay function for the inference results;

[0205] Dynamically smooth the inference confidence factor based on the decay function;

[0206] A secondary search is triggered when the smoothed confidence factor falls below the target threshold.

[0207] Secondary retrieval is used to improve the reliability of low-confidence regions.

[0208] Specifically, firstly, a confidence decay function is defined for the reasoning result. The credibility factor in the reasoning process will naturally decay with the reasoning depth, evidence distance, or link complexity. The confidence decay function is then defined. ,in This represents the inference depth or link span. The decay function can be exponential, linear, or piecewise; this example uses exponential decay. in, This is the decay rate parameter. This function is used to simulate the phenomenon that the greater the inference depth, the more likely the original credibility factor is to be weakened.

[0209] Secondly, the inference reliability factor is dynamically smoothed based on the decay function for any inference node. Its original credibility factor is Smoothed confidence factor Defined as follows:

[0210]

[0211] in, The depth of the node in the inference chain; To smooth out the weights and balance the original, reliable signal with the attenuated signal, the attenuation term will have a larger weight when a node is in a deep path, thus explicitly marking low-confidence regions. Furthermore, a dynamically smoothed gradient can be defined. Used to reflect the magnitude of change before and after smoothing:

[0212] ;

[0213] If a node shows a significant decrease in confidence after smoothing, it indicates a potential reasoning risk in that region. Finally, a secondary search is triggered when the smoothed confidence factor falls below a target threshold. To improve the reliability of low-confidence regions, a trigger threshold is set. When the following conditions are met:

[0214] ;

[0215] That is, the node is considered If the reasoning chain is not stable enough, a secondary retrieval mechanism will be automatically triggered.

[0216] Secondary searches may include the following forms:

[0217] Enhanced search: Retrieving results from external search sources Relevant high-confidence evidence;

[0218] Cross-source retrieval: Extracting structured information from alternative knowledge sources to complete the search results;

[0219] Deep search involves performing multi-hop searches on target elements to uncover potential causal relationships. The goal of secondary search is to increase the evidence density of local links, significantly improve the smoothed confidence factor, and avoid weak areas in the reasoning chain. Based on the above secondary search, it also includes confidence repair rules, as follows:

[0220] ;

[0221] in, A new credibility factor is generated for the secondary retrieval. This rule ensures that the repaired result will not fall below the smoothed credibility level.

[0222] Based on the audit records, setting a structured format for the audit records also includes:

[0223] The input information, evidence information, and reasoning information during the correction process are recorded in a structured manner according to the format.

[0224] A timeline is set in the structured record to trace the sequential relationship of the generation steps;

[0225] The time link is used for tracking and management.

[0226] Specifically, in addition to the aforementioned audit rules and secondary retrieval mechanism, this embodiment also includes a structured management step based on audit records. This involves setting a structured format for audit records. To ensure the traceability and reconfigurability of the correction process, this embodiment maps audit records to a set of structured entries. Each entry It consists of three core fields: input fields, evidence fields, and reasoning fields. Its structured form can be represented as a triple:

[0227] ;

[0228] in, This indicates the input information for the correction step; This represents a snapshot of the evidence information; This indicates the current inference state (causal chain position, confidence factor value, etc.). This structured format ensures that all audit events are recorded in a unified data model, facilitating subsequent indexing and backtracking. Input, evidence, and inference information are structured and recorded according to this format during the correction process. Each operation during the correction process generates a log unit. In this embodiment, it is mapped into structured entries. Input information Includes the correction strategy number, the referenced node identifier, and related environment variables; evidence information. Includes a structural summary of the current evidence set, a snapshot of the credibility of the evidence, and reasoning information. Then record the current causal chain position and node depth. Confidence factors before and after correction To ensure data consistency, this embodiment employs the following recording constraints for key indicators:

[0229] ;

[0230] If the record is incomplete, adjacent contextual information is automatically collected to form a valid entry. Through the above structured record, the system can capture the semantic flow of the correction process completely with low overhead. Finally, a time link is set in the structured record to trace the sequential relationship of the generation steps.

[0231] To establish a reproducible reasoning trajectory, this example introduces a timeline. The timeline provides temporal information for structured records, enabling each entry to... Its time tag Establish corresponding relationships:

[0232] ;

[0233] in, Logical timestamps (Lamport clocks), monotonically increasing sequence numbers, or composite time indices based on actual time can be used. Further define the sequence constraints of the time chain:

[0234] ;

[0235] Ensure that records strictly maintain the order in which they were generated, for use in subsequent reasoning trajectory reconstruction, conflict investigation, and process consistency review. Furthermore, this example introduces a timeline tracing and management function:

[0236] ;

[0237] It is used to quickly backtrack and reconstruct the reasoning context, avoid the misuse of correction operations in the absence of historical state, ensure the controllability and transparency of overall management, and ensure the traceability of the entire correction, reasoning and auditing process by introducing a time link.

[0238] Example 2:

[0239] Reference Figure 2 In a second embodiment of the present invention, the present invention provides a dynamic retrieval and credibility assessment system for reducing large model illusions, the system comprising:

[0240] The semantic parsing module is used to perform semantic segmentation and construct related structures;

[0241] The uncertainty determination module is used to calculate the uncertainty and trigger the retrieval.

[0242] The retrieval scheduling module is used to set the retrieval method and execute dynamic retrieval.

[0243] The evidence construction module is used to generate an expanded set of elements and evidence credibility.

[0244] The inference module is used to generate association strength maps and perform causal inference.

[0245] The Correction and Audit module is used to perform text corrections and record audit information.

[0246] Specifically, the semantic parsing module performs semantic segmentation and relational structure construction. It first decomposes the input text into a multi-level structure based on preset semantic segmentation conditions, forming a set of semantic units. Then, it calculates the correlation between semantic units and constructs a semantic relational graph structure based on the gradient changes of these correlations. This graph is used to determine the direction of semantic dependencies and logical boundaries, ensuring semantic consistency and interpretability in subsequent retrieval and reasoning processes.

[0247] The uncertainty assessment module calculates the uncertainty level of each semantic element based on a combination of missing, ambiguity, and conflict metrics. When the uncertainty exceeds a set threshold, the module triggers an external retrieval signal to mark the semantic element as a region requiring reinforcement. This mechanism can dynamically identify potentially unstable segments in the generation of large models.

[0248] The retrieval scheduling module matches the retrieval priority scheduling table based on the uncertainty level and semantic type of the target element, and determines the retrieval source, including encyclopedia knowledge base, domain database, open network, and other retrieval methods such as keyword retrieval, semantic retrieval, or multi-hop retrieval.

[0249] The evidence construction module system sets evidence mapping rules based on the search results, semantically aligns external knowledge fragments with target semantic elements, calculates the matching degree and forms an extended element set, and then calculates the credibility of the evidence based on reliability rules including source weight, information consistency, cross-validation results, etc., and performs weighted correction to highlight the contribution of highly relevant evidence, thereby forming an evidence structure with a hierarchical confidence distribution.

[0250] The reasoning module constructs an association strength graph based on the credibility of extended evidence, models the causal relationship between semantic elements, and then filters strong association paths, constructs causal chains, and performs conflict resolution operations on conflict nodes in the causal chains to maintain the monotonicity and consistency of the reasoning path. Finally, it generates the reasoning credibility factor of the candidate output content, providing a quantitative basis for subsequent correction and auditing.

[0251] The correction and audit module sets the correction range based on the reasoning credibility factor and corrects low-confidence content through replacement, insertion, or expansion. Then, it performs a consistency review according to audit rules to ensure that the corrected content is logically consistent with the evidence structure. It sets a confidence decay function to dynamically smooth the reasoning credibility factor. When the confidence level after smoothing is lower than the target threshold, it automatically triggers a secondary search to enhance the reliability of the low-confidence area. After the correction is completed, the module records the entire process in a structured audit and establishes a timeline to trace the sequential relationship of the generation steps, making the whole process traceable and monitorable.

[0252] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic retrieval and credibility assessment method for reducing large model illusions, characterized in that, Includes the following steps: S1. Set semantic segmentation conditions for the input text and perform semantic unit parsing. Calculate the correlation between semantic units and construct an initial correlation structure to determine the basic semantic range. S2. Set uncertainty calculation rules and calculate the uncertainty value of each semantic element. Determine the set of target elements that need to be triggered by external retrieval based on the uncertainty value. S3. Set the search priority based on the target element set, trigger dynamic search scheduling according to the priority, set the search source and search method according to the scheduling result, and execute the search operation. S4. Based on the search results, set the element expansion conditions and introduce related elements to form an expanded element set, construct the evidence structure, and set the evidence credibility according to the reliability rules. S5. Generate a correlation strength map based on the credibility of extended evidence, and perform causal inference based on the correlation strength map to calculate the inference credibility factor of the candidate generated content. S6. Set the correction range and determine the correction area based on the reasoning credibility factor, perform text correction operation and audit the correction content, and record the audit information for traceability.

2. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S1 includes the following steps: Define the granularity of semantic segmentation; Determine the dependency direction between semantic segments based on changes in the correlation quantity; When a conflict occurs in the dependency direction, a related structure reconstruction operation is triggered. The reconstruction is used to maintain the stability and consistency of the semantic parsing results.

3. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S2 includes the following steps: Calculate the missing, ambiguity, and conflict measures of semantic elements; Uncertainty levels are set based on three types of metrics; The search decision is triggered based on the correspondence between uncertainty level and threshold. Different levels correspond to different search trigger strengths.

4. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S3 includes the following steps: Set up a retrieval priority scheduling table and match priorities based on the type of target element; The selection of search sources is dynamically adjusted based on search feedback. The search method is switched based on changes in the search source. Priority scheduling is used to improve retrieval coverage and consistency.

5. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S4 includes the following steps: Set evidence mapping rules; Calculate the matching degree between evidence and semantic elements based on mapping rules; The credibility of the evidence is adjusted by weighting based on changes in the matching degree. The weighted correction is used to highlight the contribution of highly relevant evidence.

6. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S5 includes the following steps: Set a threshold for association strength and filter strongly associated paths; Construct causal chains based on strongly related paths; Detect conflict nodes in the causal chain and perform conflict resolution; Conflict resolution is used to maintain the monotonicity and consistency of the reasoning path.

7. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S6 includes the following steps: Set the correction strategy table; The correction method of replacement, insertion, or expansion is selected based on the reasoning credibility factor; The revised content was subjected to a consistency review in accordance with auditing rules; Among them, consistency review is used to ensure that the revised content and the evidence structure remain logically consistent.

8. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, S6 further includes the following steps: Define the confidence decay function for the inference results; Dynamically smooth the inference confidence factor based on the decay function; A secondary search is triggered when the smoothed confidence factor falls below the target threshold. Secondary retrieval is used to improve the reliability of low-confidence regions.

9. The dynamic retrieval and credibility assessment method for reducing large model illusions according to claim 1, characterized in that, Based on the audit information records, a structured format for the audit information records is established, including: The input information, evidence information, and reasoning information during the correction process are recorded in a structured manner according to the format. A timeline is set in the structured record to trace the sequential relationship of the generation steps; The time link is used for tracking and management.

10. A dynamic retrieval and credibility assessment system for reducing large model illusion, applied to the dynamic retrieval and credibility assessment method for reducing large model illusion as described in any one of claims 1-9, characterized in that, The system includes: The semantic parsing module is used to perform semantic segmentation and construct related structures; The uncertainty determination module is used to calculate the uncertainty and trigger the retrieval. The retrieval scheduling module is used to set the retrieval method and execute dynamic retrieval. The evidence construction module is used to generate an expanded set of elements and evidence credibility. The inference module is used to generate association strength maps and perform causal inference. The Correction and Audit module is used to perform text corrections and record audit information.

Images