An artificial intelligence assisted inspection method and system based on evidence binding and double-track review

Abstract

The application discloses an artificial intelligence auxiliary examination method and system based on evidence binding and double-track review, relates to the technical field of artificial intelligence, and comprises the following steps: performing multi-modal analysis on to-be-examined data to extract physical coordinate information and semantic fingerprints of content segments, and constructing an elastic evidence anchor point library; through the construction of the elastic evidence anchor point library fusing semantic fingerprints and physical coordinates, the accurate binding of multi-modal evidence is realized in combination with adaptive coupling confidence calculation; the cross-document evidence relationship graph is constructed by using topological constraint rules and information entropy measurement, and the checking intensity is dynamically configured, so that the logical path uncertainty is effectively quantified; the heterogeneous antagonistic double-track review mechanism of semantic consistency and physical accessibility, constraint satisfaction and loop detection is adopted, the adaptive iterative optimization of the anchor point library is realized through semantic drift feedback correction closed loop while ensuring the reliability of the examination, and the positioning accuracy, logical robustness and intelligent level of the complex document examination are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-assisted verification method and system based on evidence binding and dual-track review. Background Technology

[0002] With the deepening of digital transformation, massive amounts of unstructured documents have been generated in fields such as finance, law, government affairs, and corporate compliance. In business scenarios such as auditing, compliance review, and judicial evidence collection, auditors need to extract key evidentiary units from a large amount of cross-document data and verify their logical connections and factual accuracy. Traditional auditing methods rely heavily on manual reading and comparison, which is not only inefficient and cognitively burdensome, but also prone to omissions and misjudgments due to fatigue or subjective bias.

[0003] In recent years, although optical character recognition (OCR) and natural language processing (NLP) technologies have been gradually introduced into document processing workflows, existing AI-assisted verification solutions still have the following problems: Traditional solutions often process layout analysis and character recognition as separate pipelines, lacking a joint verification mechanism for geometric projection and character confidence, resulting in fragmentation or misalignment of content segments under complex layouts. At the same time, existing retrieval and positioning methods mostly rely on single semantic matching or keyword search, lacking effective utilization of physical layout features, and are unable to cope with noise interference such as OCR recognition errors, version differences, and minor adjustments in expression, resulting in inaccurate anchor point positioning or misassociation across paragraphs.

[0004] Existing technologies often focus on extracting single-point evidence, neglecting the structured modeling of logical relationships such as causality, parallelism, and referencing between evidence units. Even when knowledge graphs are constructed, they often lack topological constraint encoding for business rules, making it impossible to quantify the reasoning risks of different logical paths. This results in the average distribution of verification resources and an inability to accurately focus on high-risk evidence links.

[0005] Existing systems typically use a single model for result verification, resulting in blind spots due to systematic biases. Furthermore, confidence thresholds are mostly statically preset and cannot be dynamically calibrated based on historical false alarm rates and business evolution trends. When the model experiences semantic drift or changes in data distribution, the system lacks heterogeneous adversarial verification methods and automatic correction mechanisms, leading to a gradual degradation of verification quality over long-term operation and making it difficult to maintain a stable and reliable human-machine collaborative balance.

[0006] Therefore, how to construct an evidence anchor library that integrates multimodal features, achieve high-precision evidence binding and logical graph construction, and ensure the robustness and interpretability of the review results through a dual-track heterogeneous adversarial verification and dynamic threshold calibration mechanism has become a key technical problem that urgently needs to be solved in the field of AI-assisted review. Summary of the Invention

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-assisted verification method based on evidence binding and dual-track review, comprising: Multimodal parsing is performed on the materials to be examined to extract the physical coordinate information and semantic fingerprint of the content fragments, and a flexible evidence anchor point library is constructed. Evidence units are extracted using the first processing track, and a hybrid retrieval and positioning is performed based on the flexible evidence anchor point library to obtain a set of candidate anchor points. Anchor confidence is calculated based on semantic similarity, content matching degree and layout features to determine the target anchor point, and the evidence unit, target anchor point and anchor confidence are encapsulated into a structured binding object. Logical relationships are inferred based on structured binding objects and topological constraint rules are injected. A cross-document evidence relationship graph is constructed with evidence units as nodes, anchor confidence associated with nodes, and logical relationships carrying topological constraint rules as edges. The information entropy of the anchor confidence of each node on the logical path is calculated as a measure of path uncertainty, and a verification strength configuration strategy is generated accordingly. The second review track is used to perform heterogeneous adversarial verification on the structured binding objects and cross-document evidence relationship graphs according to the verification strength configuration strategy, so as to obtain the binding verification results and topology verification results, and generate the review status label accordingly; based on the review status label, structured binding objects and flexible evidence anchor point library, a dual-view review interface is rendered to display evidence units, logical relationship lines and original evidence fragments. The heterogeneous adversarial verification further includes calculating the semantic fingerprint residual vector between the target anchor and the evidence unit as a semantic drift representation when the binding verification result indicates that the target anchor has semantic drift, and feeding it back to the first processing track. The first processing track performs offset correction on the semantic fingerprint of the corresponding content fragment in the elastic evidence anchor library, and re-executes hybrid retrieval and positioning with the corrected semantic fingerprint to update the target anchor and anchor confidence, until the change in anchor confidence is lower than the convergence threshold or the iteration limit is reached, and then outputs the final verification result.

[0009] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps for constructing the flexible evidence anchor point library are as follows: The documents to be reviewed are divided into content segments according to page numbers and paragraph levels; Semantic encoding is performed on the content fragment units to obtain semantic fingerprint vectors, and the page number, line number, and bounding box coordinates of the content fragment units in the original document are extracted as physical coordinate information; Using the unique identifier of the content fragment unit as the primary key, the semantic fingerprint vector is associated with the physical coordinate information and stored as an anchor record; Calculate the cosine similarity between the semantic fingerprint vectors corresponding to any two anchor point records. When the cosine similarity is higher than the preset semantic aggregation threshold, the two corresponding anchor point records are grouped into the same anchor point cluster. For all anchor records within the same anchor cluster, calculate the Euclidean distance between each pair based on the bounding box coordinates in the physical coordinate information. Establish bidirectional adjacency pointers for anchor record pairs whose Euclidean distance is less than a preset spatial proximity threshold, thus forming an adjacency index within the anchor cluster.

[0010] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps for obtaining the candidate anchor point set by performing hybrid retrieval and positioning are as follows: Using the semantic fingerprint vector of the evidence unit as the query input, a semantic vector nearest neighbor search is performed in the elastic evidence anchor library, and anchor records with cosine similarity higher than the preset semantic retrieval threshold are retained as semantic candidate sets. Using the text content of the evidence unit as the query input, perform keyword fuzzy matching in the elastic evidence anchor point library, and retain anchor point records with edit distances lower than the preset content retrieval threshold as the content candidate set; Using the layout location of the evidence unit as the query input, spatial range filtering is performed in the flexible evidence anchor point library, and anchor point records with a bounding box intersection-union ratio higher than the preset layout retrieval threshold are retained as layout candidate sets. Deduplication and merging are performed on anchor records with the same primary key in the semantic candidate set, content candidate set, and layout candidate set. The content fragment units corresponding to the merged anchor records are then used as the candidate anchor set.

[0011] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps for calculating the anchor confidence level to determine the target anchor point are as follows: For each candidate anchor in the candidate anchor set, the cosine similarity between the semantic fingerprint vector of the evidence unit and the candidate anchor is calculated as the semantic score. The inverse of the normalized edit distance between the evidence unit and the candidate anchor point is used as the content score. The intersection-union ratio (IUU) of the bounding boxes of the evidence cells and the candidate anchor points is calculated as the layout score. ; Identify the document type of the materials to be reviewed and determine the semantic baseline weights based on the document type. Content benchmark weight and layout baseline weights ; Among them, the adaptive coupling confidence is calculated. The expression is: ; In the formula, This represents the semantic enhancement factor of the layout. This represents the semantic enhancement factor for the layout, and and It is dynamically fitted based on the historical binding error distribution corresponding to the document type; The candidate anchor point with the highest adaptive coupling confidence is selected as the target anchor point, and the adaptive coupling confidence is used as the anchoring confidence.

[0012] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps of injecting topological constraint rules are as follows: Identify causal, parallel, and referential logical relationships between structured bound objects; encode causal logical relationships as directed edges and attach temporal sequence constraints, encode parallel logical relationships as undirected edges and attach mutual exclusion constraints, and encode referential logical relationships as directed edges and attach anchor consistency constraints. Write the encoded constraint attributes into the corresponding edges of the cross-document evidence relationship graph; When calculating the path uncertainty metric, the constraint attribute type is read for each edge. The information entropy contribution weight of the causal logical relationship edge is set to the first multiple of the benchmark weight, the information entropy contribution weight of the parallel logical relationship edge is set to the second multiple of the benchmark weight, and the information entropy contribution weight of the referential logical relationship edge is set to the third multiple of the benchmark weight.

[0013] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps for obtaining the binding verification result and the topology verification result by performing heterogeneous adversarial verification are as follows: Binding verification is performed using two heterogeneous methods: semantic consistency verification and physical coordinate reachability verification. The semantic consistency test encodes the evidence unit and the target anchor point separately using a pre-trained language model independent of the first processing track, and calculates the KL divergence of the output probability distributions of the two. When the KL divergence is higher than the preset drift judgment threshold, semantic drift is judged to exist. The physical coordinate reachability check calls the document rendering engine to sample pixels in the bounding box region corresponding to the physical coordinate information of the target anchor point, calculates the proportion of non-blank pixels, and determines that the positioning is invalid when the proportion of non-blank pixels is lower than the preset reachability threshold. The binding verification result is obtained by performing a logical AND operation between the semantic consistency test result and the physical coordinate reachability test result. Topology verification is performed using two heterogeneous methods: constraint satisfaction testing and loop detection. The constraint satisfaction test traverses the constraint attributes of all edges in the cross-document evidence relationship graph and verifies whether the constraint attributes of each edge are violated by the node state. Loop detection uses a depth-first search algorithm to traverse the cross-document evidence relationship graph to determine whether there are illegal circular dependency paths where nodes are accessed repeatedly; The topology verification result is obtained by performing a logical AND operation between the constraint satisfaction test result and the loop detection result.

[0014] As a preferred embodiment of the AI-assisted verification method based on evidence binding and dual-track review described in this invention, the specific steps for generating verification status tags are as follows: When both the binding verification result and the topology verification result are passed and the anchor confidence level is higher than the high confidence threshold, a confirmation pass label is generated; When both the binding verification result and the topology verification result are passed, but the anchor confidence level is between the low confidence threshold and the high confidence threshold, a label awaiting manual review is generated. An exception tag is generated when either the binding verification result or the topology verification result fails. When the anchor confidence level is lower than the low confidence threshold, a location failure label is generated regardless of the binding verification result and the topology verification result. Collect historical verification data to confirm the false alarm rate of samples corresponding to the labels. When the false alarm rate is higher than the preset calibration trigger threshold, dynamically lower the high confidence threshold and low confidence threshold based on the false alarm rate deviation.

[0015] Secondly, the present invention provides an artificial intelligence-assisted verification system based on evidence binding and dual-track review, comprising: Anchor point construction module, graph deduction module, and adversarial verification module; The anchor point construction module is used to perform multimodal analysis on the materials to be examined to construct a flexible evidence anchor point library, extract evidence units using the first processing track, and perform hybrid retrieval and positioning based on the anchor point library to determine the target anchor point, and encapsulate the evidence units, target anchor points and anchor confidence into structured binding objects. The graph inference module is used to infer logical relationships based on structured binding objects and inject topological constraint rules to construct a cross-document evidence relationship graph, calculate the information entropy of the anchor confidence of each node on the logical path as a measure of path uncertainty, and generate a verification strength configuration strategy accordingly. The adversarial verification module is used to perform heterogeneous adversarial verification according to the verification strength configuration strategy of the second verification track to obtain binding and topology verification results and generate verification status labels, and render a dual-view verification interface; when semantic drift exists, the semantic fingerprint residual vector is calculated and fed back to the first processing track to correct the anchor point library and reposition until convergence and the final verification result is output.

[0016] The beneficial effects of this invention are as follows: by constructing an elastic evidence anchor point library that integrates semantic fingerprints and physical coordinates, and combining adaptive coupling confidence calculation, accurate binding of multimodal evidence is achieved; by using topological constraint rules and information entropy measurement to construct a cross-document evidence relationship graph and dynamically configure the verification strength, the uncertainty of logical paths is effectively quantified; by adopting a heterogeneous adversarial dual-track review mechanism such as semantic consistency and physical reachability, constraint satisfaction and loop detection, while ensuring the reliability of the review, the anchor point library is adaptively iteratively optimized through semantic drift feedback correction closed loop, thereby improving the positioning accuracy, logical robustness and intelligence level of complex document review. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-assisted verification method based on evidence binding and dual-track review in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of an AI-assisted verification system based on evidence binding and dual-track review in an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Reference Figure 1 and Figure 2 This is one embodiment of the present invention, which provides an AI-assisted verification method based on evidence binding and dual-track review, specifically including: S1. Perform multimodal parsing on the materials to be examined to extract the physical coordinate information and semantic fingerprint of the content fragments, and construct a flexible evidence anchor point library; extract evidence units using the first processing track, perform hybrid retrieval and positioning based on the flexible evidence anchor point library to obtain a set of candidate anchor points, calculate the anchor confidence based on semantic similarity, content matching degree and layout features to determine the target anchor point, and encapsulate the evidence unit, target anchor point and anchor confidence as a structured binding object.

[0024] Furthermore, the materials to be reviewed are segmented into content fragment units according to page numbers and paragraph levels.

[0025] The process of segmenting the documents to be reviewed into content units by page number and paragraph level also includes performing a text-image alignment verification step: using a layout analysis model to identify the set of text line bounding boxes in the document. With paragraph area collection Simultaneously, it utilizes an OCR engine to output character-level recognition results with coordinates. For each paragraph area The bounding boxes of all text lines whose center points fall within the region are selected as candidate line sets. The vertical projection overlap rate of adjacent text lines in the candidate line sets is then calculated. Horizontal spacing ;when Higher than the preset alignment threshold and When the text is below the preset break threshold, adjacent text lines are merged into a single content segment unit. For the merged content segment unit, the weighted average of the confidence scores of all character-level recognition results contained therein is used as the text confidence score of the segment. Subsequent semantic encoding and anchor point recording storage are only performed when the text confidence score is higher than the preset quality threshold.

[0026] By introducing dual alignment verification based on geometric projection and character confidence, the fragmentation and misalignment problems caused by the independent operation of layout detection and character recognition in the traditional parsing process are solved. This ensures that each content fragment unit entering the anchor point library is continuous and complete in physical space and accurate and reliable in semantic carrier, providing a high-quality data base for subsequent hybrid retrieval.

[0027] Semantic encoding is performed on the content fragment unit to obtain a semantic fingerprint vector, and the page number, line number, and bounding box coordinates of the content fragment unit in the original document are extracted as physical coordinate information; the semantic fingerprint vector and physical coordinate information are associated and stored as anchor point records using the unique identifier of the content fragment unit as the primary key.

[0028] Calculate the cosine similarity between the semantic fingerprint vectors corresponding to any two anchor records. When the cosine similarity is higher than the preset semantic aggregation threshold, the two corresponding anchor records are grouped into the same anchor cluster. For all anchor records in the same anchor cluster, calculate the Euclidean distance between each pair based on the bounding box coordinates in the physical coordinate information.

[0029] Euclidean distance The calculation expression is: ; In the formula, and They represent the first Article and No. Each anchor point records the coordinates of the center point of the corresponding bounding box. and These represent the width and height of the corresponding bounding boxes, respectively. The purpose of introducing this calculation is to further distinguish spatial layout relationships within the same semantic cluster. This is because semantic aggregation alone may group similar content scattered across different pages or sections of a document into one category. By measuring geometric distance in four-dimensional coordinate space (center point plus dimensions), we can accurately identify visually adjacent content blocks, providing strict spatial constraints for constructing bidirectional adjacency pointers and preventing semantically identical but physically unrelated segments from being incorrectly associated.

[0030] For anchor point records whose Euclidean distance is less than a preset spatial proximity threshold, establish bidirectional adjacency pointers to form an adjacency index within the anchor point cluster.

[0031] Among them, cosine similarity The calculation expression is: ; In the formula, and These represent the semantic fingerprint vectors of the two anchor records, This represents the vector dot product operation. The Euclidean norm of the vector is used because semantic fingerprint vectors are usually in high-dimensional space, and their magnitude is often affected by the text length or the output characteristics of the encoding model. Directly calculating the distance is easily affected by the magnitude. However, cosine similarity, through normalization, only focuses on the angle between the vectors in the direction, which can purely reflect the topic relevance of content fragments in the semantic space, thus providing a metric for subsequent semantic aggregation that is not affected by the text size.

[0032] Using the semantic fingerprint vector of the evidence unit as the query input, a semantic vector nearest neighbor search is performed in the flexible evidence anchor point library, retaining anchor point records with a cosine similarity higher than a preset semantic retrieval threshold as a semantic candidate set; using the text content of the evidence unit as the query input, keyword fuzzy matching is performed in the flexible evidence anchor point library, retaining anchor point records with an edit distance lower than a preset content retrieval threshold as a content candidate set; using the layout position of the evidence unit as the query input, spatial range filtering is performed in the flexible evidence anchor point library, retaining anchor point records with a bounding box intersection-union ratio higher than a preset layout retrieval threshold as a layout candidate set.

[0033] The process of fusing semantic candidate sets, content candidate sets, and layout candidate sets is aimed at those that simultaneously exist in... For records with the same anchor point in each candidate set, a weighted inverse ranking fusion algorithm is used to calculate their comprehensive retrieval score. The calculation formula is: ; In the formula, Indicates the anchor record identifier. This indicates that the record is in the [number]th [year]. The ranking position in the candidate set. For the first Preset weighting coefficients for each search channel To prevent the smoothing constant from having a denominator of zero, for anchor records that exist only in a single candidate set, their original retrieval scores, after normalization, are directly used as the standard value. ;in accordance with Sort all candidate anchor points in descending order and extract the top-ranked ones. The records constitute the final set of candidate anchor points.

[0034] By introducing a reciprocal ranking fusion mechanism, the incomparability of the three heterogeneous measures of semantic similarity, edit distance and intersection-union ratio in terms of numerical dimensions is eliminated, enabling multimodal retrieval results to be fairly compared and complementary under a unified scale, ensuring that the selection of high-confidence target anchors is based on the collaborative support of multi-dimensional evidence.

[0035] Deduplication and merging are performed on anchor records with the same primary key in the semantic candidate set, content candidate set, and layout candidate set. The content fragment units corresponding to the merged anchor records are then used as the candidate anchor set.

[0036] Each candidate anchor in the candidate anchor set retains its corresponding comprehensive retrieval score when entering the confidence calculation stage. As a priori confidence factor; used in subsequent calculations of adaptive coupling confidence. At that time, Introduced as a normalization scaling factor, the corrected adaptive coupling confidence. The calculation expression is: ; In the formula, Represents the current set of candidate anchor points. This is a safety net weighting coefficient for the prior confidence level, used to prevent factors from causing... Too low a score completely suppresses the contribution of fine ranking score; by embedding the RRF fusion score as a prior factor into the final confidence calculation, the mathematical coherence of the coupling between coarse ranking fusion and fine ranking is achieved, ensuring that the evidence strength in the multimodal retrieval stage can be losslessly transferred to the final positioning decision stage, and avoiding the logical separation between fusion score and confidence calculation.

[0037] For each candidate anchor in the candidate anchor set, the cosine similarity between the semantic fingerprint vector of the evidence unit and the candidate anchor is calculated as the semantic score. The inverse of the normalized edit distance between the evidence unit and the candidate anchor point is used as the content score. .

[0038] The inverse of the normalized edit distance is used as the content score. The calculation expression is: ; In the formula, Representing the text of the evidence unit With candidate anchor text The Levenstein edit distance between strings is the minimum number of single-character edit operations required to transform one string into another. and These represent the lengths of the two strings, respectively. This algorithm, rather than simple keyword matching, is used because the materials to be reviewed may have OCR recognition errors, version differences, or minor adjustments in expression. Exact matching is very easy to fail, while edit distance can tolerate local noise and distortion at the character sequence level. Then, the influence of text length on the score is eliminated by maximum length normalization. Finally, the value is inverted so that the larger the value, the higher the matching degree, thus providing a quantitative indicator for content scoring that is robust to text perturbations.

[0039] The intersection-union ratio (IUU) of the bounding boxes of the evidence cells and the candidate anchor points is used as the layout score. Identify the document type of the materials to be reviewed, and determine the semantic baseline weights based on the document type. Content benchmark weight and layout baseline weights .

[0040] Among them, the adaptive coupling confidence is calculated. The expression is: ; In the formula, This represents the semantic enhancement factor of the layout. This represents the semantic enhancement factor for the layout, and and The anchor point is dynamically fitted based on the historical binding error distribution corresponding to the document type; the candidate anchor point with the highest adaptive coupling confidence is selected as the target anchor point, and the adaptive coupling confidence is used as the anchoring confidence.

[0041] The dynamic fitting process of α and β mentioned above specifically includes: collecting manually corrected binding samples from historical review tasks under the same document type, and extracting the original layout score of each sample before correction. Original semantic score And the offset of the corrected anchor point relative to the original target anchor point after manual correction. To minimize the ranking loss between the predicted confidence level and the manually confirmed results, a multiple regression model is constructed. and As the independent variable, the offset The reciprocal of the equation is used as the dependent variable weight for weighted fitting; when the fitting residuals converge, the cross coefficients of the layout term and the semantic term in the regression equation are extracted as the current α and β, respectively; every preset period or when the cumulative number of newly added corrected samples reaches a preset number, the regression model parameters are updated using an incremental learning strategy to ensure that the enhancement coefficients continuously adapt to the layout evolution trend of the current document type; through the above fitting process, α and β are no longer empirical constants, but statistical quantities that directly reflect the actual coupling strength between layout and semantics in a specific document type, ensuring that the cross enhancement term in the confidence calculation formula has a clear data source and physical meaning.

[0042] The incremental learning strategy is implemented using a double-buffered hot update mechanism: the system maintains two independent parameter storage spaces for the primary model instance and the backup model instance; when the update condition is triggered, a mixed dataset consisting of the historical full samples and the newly added corrected samples is loaded onto the backup model instance to perform fine-tuning training, and elastic weights are used to consolidate the loss function during training. Constraint key parameter offset: ; In the formula, For the fitting loss on the new samples, For the first The importance weights of each parameter (estimated by the diagonal elements of the Fisher information matrix) The parameter values ​​corresponding to the main model. To determine the regularization strength; after the backup model is trained and passes performance verification on the offline validation set, it is switched to the primary model instance via atomic pointer swapping, and the original primary model is downgraded to the base for the next round of updates; through the synergy of the double-buffered architecture and EWC regularization, the zero-disruption continuity of online services is ensured, while effectively suppressing catastrophic forgetting during incremental updates, making... and Its dynamic evolution possesses deployability and stability at the engineering level.

[0043] It should be noted that by performing multi-level segmentation of the materials to be examined and constructing a dual index structure that integrates semantic fingerprints and physical coordinates, it is possible to achieve efficient semantic clustering and spatial association while preserving the original layout features of the documents. This provides a refined data foundation that takes into account both semantic relevance and location proximity for subsequent hybrid retrieval, thereby improving the construction quality and retrieval response efficiency of the flexible evidence anchor library.

[0044] In particular, the dual index structure is designed to address the disconnect between semantic understanding and layout structure in traditional document parsing. By first aggregating based on semantic similarity and then building an adjacency index based on physical proximity, it simulates the cognitive habits of human reviewers who focus on both content association and visual location cues. This allows the anchor library to not only possess semantic retrieval capabilities but also to quickly narrow down the search scope for subsequent positioning by utilizing spatial contextual relationships. This effectively avoids cross-paragraph misassociations or missed detections caused by relying solely on semantic matching, laying a data foundation that aligns with actual review cognitive patterns for high-precision evidence binding.

[0045] S2. Based on structured binding objects, infer logical relationships and inject topological constraint rules. Construct a cross-document evidence relationship graph with evidence units as nodes, anchor confidence associated with nodes, and logical relationships carrying topological constraint rules as edges. Calculate the information entropy of the anchor confidence of each node on the logical path as a measure of path uncertainty, and generate a verification strength configuration strategy accordingly.

[0046] Furthermore, it identifies causal, parallel, and referential logical relationships between structured bound objects.

[0047] The process of identifying the logical relationship between structured bound objects involves extracting the text content and metadata fields of the evidence unit corresponding to each pair of structured bound objects and inputting them into a pre-trained logical relationship classification model.

[0048] The construction process of the pre-trained logical relationship classification model includes: firstly, extracting evidence unit pairs labeled with logical relationships from historical examination archives as positive samples; and then constructing negative samples using a hard negative sample mining strategy. Specifically, evidence unit pairs with high semantic similarity but no actual logical connection are selected as hard negative samples, and evidence unit pairs from adjacent paragraphs within the same document but belonging to different topics are selected as positional interference negative samples. Positive and negative samples are then mixed in a preset ratio to form a fine-tuned dataset. Based on a general language model, the underlying parameters are frozen, and only the top-level classification head and the last two Transformer blocks are fully fine-tuned. The loss function uses label-smoothed cross-entropy loss. : ; In the formula, The predicted probability of the true label category. The total number of categories, The smoothing coefficient is used to effectively alleviate the overfitting problem caused by the sparsity of logical relationship samples and class imbalance in the review scenario by introducing hard negative sample mining and label smoothing loss. This enables the model to accurately distinguish interference terms that are semantically similar but logically unrelated, ensuring that the accuracy of graph edge generation meets the requirements of industrial-grade review.

[0049] The logical relationship classification model employs a dual-tower encoder to extract contextual representation vectors E1 and E2 from two evidence units, concatenating them and then using a multilayer perceptron to output probability distributions for four categories: causal, parallel, referential, and irrelevant. Simultaneously, the timestamp, chapter number, and explicit conjunctions of the document containing the evidence unit are extracted as auxiliary features and fused with the model's output probability distributions using Bayesian methods to obtain the final logical relationship classification result. When a causal relationship is determined, the edge direction is further determined based on the timestamp or chapter order. When a referential relationship is determined, a referencing identifier is extracted from the evidence unit text using regular expressions and matched against the unique identifier of the target anchor point for verification; the referencing relationship is confirmed only when a match is successful.

[0050] By integrating a dual inference mechanism of deep learning semantic understanding and rule-based metadata verification, the problem of low recognition rate of single models in long-tail logical relationships is overcome. This enables the generation of graph edges to have both semantic generalization ability and meet the strict constraints of business logic, ensuring that the topological structure of the cross-document evidence relationship graph truly reflects the inherent relationship between evidence.

[0051] Encode causal logical relationships as directed edges and attach temporal sequence constraints; encode parallel logical relationships as undirected edges and attach mutual exclusion checks; encode referential logical relationships as directed edges and attach anchor point consistency constraints; and write the encoded constraint attributes into the corresponding edges of the cross-document evidence relationship graph.

[0052] When calculating the path uncertainty metric, the constraint attribute type is read for each edge. The information entropy contribution weight of the causal logical relationship edge is set to the first multiple of the benchmark weight, the information entropy contribution weight of the parallel logical relationship edge is set to the second multiple of the benchmark weight, and the information entropy contribution weight of the referential logical relationship edge is set to the third multiple of the benchmark weight.

[0053] Path uncertainty measurement The calculation expression is: ; In the formula, It represents a complete logical path. Indicates the first path Anchor confidence of each node This represents the information entropy contribution weight multiple corresponding to the constraint attribute type of the edge connected to the node. The binary information entropy function is chosen because the anchor confidence level is essentially a probability value. When the confidence level is close to 0 or 1, the system has high certainty and low uncertainty, while when the confidence level is close to 0.5, the system is in the most ambiguous state. The information entropy curve can non-linearly amplify the uncertainty of this intermediate state. Multiplying by the edge weight multiple adds business logic risk to the mathematical metric, so that the final path uncertainty includes both the binding credibility at the data level and the reasoning risk level at the logical level, providing an interpretable quantitative basis for the subsequent differentiated allocation of verification resources.

[0054] The specific process of generating a verification strength configuration strategy based on path uncertainty metrics involves pre-setting a set of uncertainty classification thresholds. ,Will The range is divided into low uncertainty regions. Uncertainty in the medium region With high uncertainty region ;when When configuring a lightweight verification strategy, that is, only performing a single semantic consistency check and not triggering loop detection; when When configuring a standard verification strategy, semantic consistency checks and physical coordinate reachability checks are performed sequentially, and local subgraph loop detection is triggered when constraint satisfaction checks fail; when At the same time, an enhanced verification strategy is configured, which adds a secondary semantic encoding comparison of independent heterogeneous models on the basis of standard verification, performs depth-first search loop detection on the entire graph, and tightens the judgment threshold of the verification results according to a preset ratio; by establishing An explicit mapping table between verification resource consumption enables the system to accurately allocate limited computing power to the evidence link with the highest logical risk, avoiding redundant verification on low-risk paths and maximizing overall processing throughput while ensuring verification quality.

[0055] It should be noted that by encoding different types of logical relationships into graph edges with specific constraint attributes and setting information entropy contribution weights accordingly, abstract business logic rules can be transformed into quantifiable topological constraints. This allows path uncertainty measurement to accurately reflect the degree of influence of different logical connections on verification reliability, thereby providing a scientific basis for adaptive configuration of verification strength and enhancing the logical rigor of cross-document evidence reasoning.

[0056] In particular, the core intention of introducing a layout and semantic mutual enhancement mechanism and document type adaptive weights is to overcome the inherent problem of large differences in the contribution of multimodal features in different document types. For example, layout information in table documents is often more discriminative than pure text semantics, while the opposite is true for contract documents. By dynamically fitting enhancement coefficients based on historical binding errors instead of using fixed parameters, the system can automatically adjust the coupling strength of features in each dimension according to the actual characteristics of the document. The adaptive design avoids the generalization bottleneck caused by manually preset weights, ensuring that the best evidence matching effect can be achieved in diverse document scenarios, and improving the robustness and practicality of the localization results.

[0057] The underlying logic of constraining and encoding logical relationships and setting information entropy weights differently stems from the fact that different logical connections have fundamentally different impacts on the reliability of the verification conclusion. Causal relationships involve time sequence and conditional dependence, and once an error occurs, the entire reasoning chain will collapse. Therefore, higher uncertainty weights are needed to trigger stronger verification. Parallel relationships, although independent, have the risk of mutual exclusion, while referential relationships require anchor point consistency. By internalizing these business rules into computable attributes of graph edges, the path uncertainty measurement is no longer a simple addition of confidence levels, but a quantitative indicator that truly reflects the business logic risk. This enables the intelligent tilting of verification resources towards high-risk paths, improving overall verification efficiency and logical rigor.

[0058] S3. Using the second review track, heterogeneous adversarial verification is performed on the structured binding objects and cross-document evidence relationship graphs according to the verification strength configuration strategy. The binding verification results and topology verification results are obtained, and the review status labels are generated accordingly. Based on the review status labels, structured binding objects and flexible evidence anchor point library, a dual-view review interface is rendered to display evidence units, logical relationship connections and original evidence fragments.

[0059] The dual-view verification interface linkage mechanism consists of an evidence logic graph view on the left and an original document image view on the right. These two views establish a real-time mapping channel through a shared anchor point unique identifier. When a user selects an evidence node in the graph view, the system automatically scrolls the right-hand image view to the corresponding page number and highlights the corresponding bounding box area based on the physical coordinates of the target anchor point associated with that node. Simultaneously, other content fragments pointed to by adjacent indexes within the same anchor point cluster are presented in a semi-transparent overlay for contextual reference. When a user clicks on a bounding box area in the image view, the system performs a reverse query to locate the corresponding evidence node in the flexible evidence anchor point library and automatically expands the upstream and downstream logical paths of that node in the graph view, while simultaneously measuring the path uncertainty. Edges exceeding the preset attention threshold are marked with differentiated colors; in addition, the review status labels are directly mapped to the color codes of map nodes and image highlight boxes, with green corresponding to the confirmed pass label, yellow corresponding to the label awaiting manual review, red corresponding to the abnormal mark label, and gray flashing corresponding to the positioning failure label.

[0060] Through the aforementioned two-way interactive rendering and state visualization encoding, the abstract binding confidence and topological uncertainty are transformed into visual cues that users can intuitively perceive, reducing the cognitive load and operation path length during manual review, and making the interface itself a technical means to improve review efficiency rather than a simple information display window.

[0061] The heterogeneous adversarial verification also includes calculating the semantic fingerprint residual vector between the target anchor and the evidence unit as a semantic drift representation when the binding verification result shows that the target anchor has semantic drift and feeding it back to the first processing track; the first processing track performs offset correction on the semantic fingerprint of the corresponding content fragment in the elastic evidence anchor library.

[0062] The offset correction is performed by adjusting the semantic fingerprint residual vector. Multiply by the correction step size factor Then superimposed on the original semantic fingerprint vector The corrected semantic fingerprint vector is obtained from the above. Among them, the step size factor is corrected. With the number of iterations Increasing and linearly decreasing, i.e. To prevent oscillations near the optimal solution during the correction process; record the current anchoring confidence level after each iteration. If the absolute value of the change in anchor confidence between two consecutive iterations All are below the convergence threshold If convergence is reached, the current anchor confidence level and the corresponding target anchor point are output; if the iteration limit is reached... If the convergence condition is still not met, the target anchor point and semantic fingerprint corresponding to the one with the highest anchor confidence in each iteration are backtracked as the final output, and the semantic fingerprint of this iteration is written back to the elastic evidence anchor point library to replace the original record. By introducing the decay step size and the historical best backtracking mechanism, the mathematical stability of the correction process is guaranteed, and the correction failure caused by local minima is avoided, so that the feedback loop can reliably terminate and produce effective results under various initial bias conditions.

[0063] The hybrid retrieval and localization are re-executed with the corrected semantic fingerprint to update the target anchor and anchor confidence until the change in anchor confidence is lower than the convergence threshold or the iteration limit is reached, at which point the final verification result is output.

[0064] Furthermore, two heterogeneous methods, semantic consistency testing and physical coordinate reachability testing, are used to perform binding verification. Semantic consistency testing encodes the evidence unit and the target anchor point using a pre-trained language model independent of the first processing track.

[0065] The selection and adaptation of the pre-trained language model, independent of the first processing track, follows the principle of orthogonal heterogeneity: if the first processing track uses a semantic embedding model based on the Transformer architecture, then the secondary processing track uses a classification model based on a statistical language model or a convolutional neural network architecture, ensuring that there is an essential difference between the two in terms of feature extraction paradigms; at the same time, the overlap between the training corpus of the secondary processing track model and the training corpus of the first processing track is less than a preset threshold. Prioritize using cross-domain or cross-language transfer-tuned models as verifiers; before deployment, use a standard test set to verify the misclassification correlation coefficient between the two models on the same erroneous samples. Only when Only after confirming that the heterogeneity meets the standards can it be put into use; through the triple constraints of architectural differentiation, data decorrelation and low correlation verification for misjudgment, the cognitive blind spot complementarity between the review track and the main processing track is guaranteed from the engineering level, so that KL divergence detection can truly have the ability to discover systematic semantic drift, rather than just repeating the judgment results of the main track.

[0066] Calculate the KL divergence of the output probability distributions of both; KL divergence The calculation expression is: ; In the formula, This represents the first expression of the independent pre-trained language model's response to the evidence unit output. The probability of each semantic category, This represents the corresponding category probability output by the same model for the target anchor point. The use of KL divergence instead of cosine similarity for semantic consistency verification is because the first processing track has already used vector similarity for localization. If the verification track still uses the same metric, systematic deviations cannot be detected. KL divergence measures the information loss between two complete probability distributions and is extremely sensitive to differences in distribution shape. Even if the main directions of two vectors are similar, as long as the distribution ratio of secondary semantic components shifts, it can be captured, thus truly achieving in-depth detection of semantic drift in heterogeneous verification.

[0067] When the KL divergence is higher than the preset drift determination threshold, semantic drift is determined to exist.

[0068] The physical coordinate reachability check calls the document rendering engine to sample pixels in the bounding box region corresponding to the physical coordinate information of the target anchor point, calculates the proportion of non-blank pixels, and determines that the positioning is invalid when the proportion of non-blank pixels is lower than the preset reachability threshold; the semantic consistency check result and the physical coordinate reachability check result are logically ANDed to obtain the binding check result.

[0069] Topology verification is performed using two heterogeneous methods: constraint satisfaction testing and loop detection. Constraint satisfaction testing involves traversing the constraint attributes of all edges in the cross-document evidence graph to verify whether the constraint attributes of each edge are violated by the node state. Loop detection uses a depth-first search algorithm to traverse the cross-document evidence graph to determine whether there are illegal circular dependency paths where nodes are visited repeatedly. The constraint satisfaction test results and loop detection results are then logically ANDed to obtain the topology verification results.

[0070] Furthermore, when both the binding verification result and the topology verification result are passed and the anchor confidence level is higher than the high confidence threshold, a confirmation pass label is generated; when both the binding verification result and the topology verification result are passed but the anchor confidence level is between the low confidence threshold and the high confidence threshold, a manual review pending label is generated; when either the binding verification result or the topology verification result is failed, an anomaly label is generated; when the anchor confidence level is lower than the low confidence threshold, a positioning failure label is generated regardless of the binding verification result or the topology verification result.

[0071] The false alarm rate of samples corresponding to the labels was confirmed by collecting historical inspection data.

[0072] False alarm rate The calculation expression is: ; In the formula, This indicates the number of samples in historical verification data that were marked as passed but were confirmed as incorrect through manual sampling or subsequent feedback. This represents the total number of samples that automatically generated and confirmed pass labels during the same period. This indicator is defined to provide an unambiguous feedback signal for dynamic threshold calibration. The pass rate or accuracy rate alone cannot distinguish whether the system is too lenient or too strict, while the false alarm rate directly describes the scale of quality loopholes in automated release decisions. Using it as a trigger condition can ensure that the calibration mechanism is only activated when the system actually experiences a quality decline, avoiding frequent parameter adjustments under normal fluctuations and ensuring that threshold adjustments are both based on evidence and remain stable.

[0073] When the false alarm rate exceeds the preset calibration trigger threshold, the high confidence threshold and low confidence threshold are dynamically lowered based on the false alarm rate deviation.

[0074] Dynamically lower the high confidence threshold With low confidence threshold The process also includes providing and Preset safety lower limit values ​​respectively and The threshold after any calibration must not be lower than the corresponding safety lower limit to prevent the automatic release function from becoming substantially ineffective due to an excessively low threshold; a sliding window mechanism is used to calculate the false alarm rate, which is only calculated when the most recent consecutive false alarms occur. Average false alarm rate over a statistical period All are above the calibration trigger threshold The adjustment operation is only performed at specific times; instantaneous fluctuations within a single cycle do not trigger calibration; the magnitude of each adjustment is the false alarm rate deviation. With preset damping coefficient The product of the two thresholds, and the single downward adjustment does not exceed half the difference between the current threshold and the corresponding safety lower limit. When the false alarm rate falls back to [a certain level] in subsequent statistical periods After maintaining the preset observation period, the threshold is gradually increased according to the preset recovery step size until the initial value is restored or the downward adjustment condition is triggered again. By introducing four constraints of safety lower limit, sliding window confirmation, damping limit and gradual recovery, the threshold calibration process can not only respond to the real quality degradation trend, but also effectively filter noise interference and extreme anomalies, ensuring that the system is always in a controllable and stable working state during long-term operation.

[0075] It should be noted that by generating multi-level verification status labels through comprehensive binding and topological dual verification results and anchored confidence, and dynamically calibrating the confidence threshold in combination with historical false alarm rates, it is possible to achieve hierarchical and refined management and adaptive optimization of verification conclusions. While ensuring the efficiency of automatic passage of high-confidence samples, it can accurately divert and guide low-confidence or abnormal samples with manual intervention, continuously improving the accuracy, robustness and practical value of the overall verification system.

[0076] In particular, the core advantage of heterogeneous adversarial verification lies in eliminating the systematic blind spots of a single detection method by adopting the orthogonal verification principle. Semantic models may make misjudgments due to training data bias, but physical rendering engines are not affected by semantic interference based on pixel-level sampling, and vice versa. Similarly, constraint verification focuses on the compliance of local edge attributes, while loop detection identifies illegal dependencies from the global topology. The two complement each other to cover local and global anomalies. The multi-dimensional cross-verification mechanism reduces the probability of false positives and false negatives, ensuring that the verification results have both depth and breadth, and providing a highly credible basis for subsequent review decisions.

[0077] The introduction of a multi-level classification and historical false alarm rate-driven threshold dynamic calibration mechanism is based on the risk sensitivity and business evolution of the review task itself. Static thresholds cannot adapt to data distribution drift or business standard adjustments during long-term operation. By using the false alarm rate of confirmed samples as a feedback signal to automatically lower the confidence threshold, the system can gradually release automated processing capabilities while ensuring a safety baseline. At the same time, it forces low-confidence or abnormal samples to be diverted to the manual review stage to form a closed-loop learning. The design avoids efficiency losses caused by excessive conservatism and prevents quality risks caused by blind confidence, enabling the system to continuously approach the optimal human-machine collaboration balance point during continuous operation.

[0078] This embodiment also provides an AI-assisted verification system based on evidence binding and dual-track review, including an anchor point construction module, a graph inference module, and an adversarial review module.

[0079] The anchor point construction module is used to perform multimodal parsing on the materials to be examined to build a flexible evidence anchor point library. It uses the first processing track to extract evidence units and performs hybrid retrieval and positioning based on the anchor point library to determine the target anchor point. The evidence units, target anchor points and anchor confidence are encapsulated into structured binding objects.

[0080] The graph inference module is used to infer logical relationships based on structured binding objects and inject topological constraint rules to construct a cross-document evidence relationship graph. It calculates the information entropy of the anchor confidence of each node on the logical path as a measure of path uncertainty and generates a verification strength configuration strategy accordingly.

[0081] The adversarial review module is used to perform heterogeneous adversarial verification according to the verification strength configuration strategy of the second review track to obtain binding and topology verification results and generate review status labels, and render the dual-view review interface; when semantic drift exists, the semantic fingerprint residual vector is calculated and fed back to the first processing track to correct the anchor point library and reposition until convergence and the final review result is output.

[0082] This embodiment also provides a computer device applicable to the AI-assisted verification method based on evidence binding and dual-track review, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the AI-assisted verification method based on evidence binding and dual-track review as proposed in the above embodiment.

[0083] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0084] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the AI-assisted verification method based on evidence binding and dual-track review as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0085] In summary, this invention constructs a flexible evidence anchor library that integrates semantic fingerprints and physical coordinates, and combines adaptive coupling confidence calculation to achieve precise binding of multimodal evidence; it utilizes topological constraint rules and information entropy measurement to construct a cross-document evidence relationship graph and dynamically configure verification strength, effectively quantifying the uncertainty of logical paths; and it adopts a heterogeneous adversarial dual-track verification mechanism, including semantic consistency and physical reachability, constraint satisfaction and loop detection, to ensure the reliability of verification while achieving adaptive iterative optimization of the anchor library through semantic drift feedback correction closed loop, thereby improving the positioning accuracy, logical robustness and intelligence level of complex document verification.

[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-assisted verification method based on evidence binding and dual-track review, characterized in that: include: Multimodal parsing is performed on the materials to be examined to extract the physical coordinate information and semantic fingerprint of the content fragments, and a flexible evidence anchor point library is constructed. Evidence units are extracted using the first processing track, and a hybrid retrieval and positioning is performed based on the flexible evidence anchor point library to obtain a set of candidate anchor points. Anchor confidence is calculated based on semantic similarity, content matching degree and layout features to determine the target anchor point, and the evidence unit, target anchor point and anchor confidence are encapsulated into a structured binding object. Logical relationships are inferred based on structured binding objects and topological constraint rules are injected. A cross-document evidence relationship graph is constructed with evidence units as nodes, anchor confidence levels associated with nodes, and logical relationships carrying topological constraint rules as edges. The information entropy of the anchor confidence of each node on the logical path is used as a measure of path uncertainty, and a verification strength configuration strategy is generated accordingly. The second review track is used to perform heterogeneous adversarial verification on the structured binding objects and cross-document evidence relationship graphs according to the verification strength configuration strategy, so as to obtain the binding verification results and topology verification results, and generate the review status label accordingly; based on the review status label, structured binding objects and flexible evidence anchor point library, a dual-view review interface is rendered to display evidence units, logical relationship lines and original evidence fragments. The heterogeneous adversarial verification also includes, when the binding verification result shows that there is semantic drift at the target anchor point, calculating the semantic fingerprint residual vector between the target anchor point and the evidence unit as a semantic drift representation and feeding it back to the first processing track. The first processing step performs offset correction on the semantic fingerprint of the corresponding content fragment in the elastic evidence anchor library, and then re-executes hybrid retrieval and positioning with the corrected semantic fingerprint to update the target anchor and anchor confidence, until the change in anchor confidence is lower than the convergence threshold or the iteration limit is reached, and then outputs the final verification result.

2. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 1, characterized in that: The specific steps for constructing the flexible evidence anchor point library are as follows: The documents to be reviewed are divided into content segments according to page numbers and paragraph levels; Semantic encoding is performed on the content fragment units to obtain semantic fingerprint vectors, and the page number, line number, and bounding box coordinates of the content fragment units in the original document are extracted as physical coordinate information; Using the unique identifier of the content fragment unit as the primary key, the semantic fingerprint vector is associated with the physical coordinate information and stored as an anchor record; Calculate the cosine similarity between the semantic fingerprint vectors corresponding to any two anchor point records. When the cosine similarity is higher than the preset semantic aggregation threshold, the two corresponding anchor point records are grouped into the same anchor point cluster. For all anchor records within the same anchor cluster, calculate the Euclidean distance between each pair based on the bounding box coordinates in the physical coordinate information. Establish bidirectional adjacency pointers for anchor record pairs whose Euclidean distance is less than a preset spatial proximity threshold, thus forming an adjacency index within the anchor cluster.

3. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 2, characterized in that: The specific steps for obtaining the candidate anchor point set through hybrid retrieval and positioning are as follows: Using the semantic fingerprint vector of the evidence unit as the query input, a semantic vector nearest neighbor search is performed in the elastic evidence anchor library, and anchor records with cosine similarity higher than the preset semantic retrieval threshold are retained as semantic candidate sets. Using the text content of the evidence unit as the query input, perform keyword fuzzy matching in the elastic evidence anchor point library, and retain anchor point records with edit distances lower than the preset content retrieval threshold as the content candidate set; Using the layout location of the evidence unit as the query input, spatial range filtering is performed in the flexible evidence anchor point library, and anchor point records with a bounding box intersection-union ratio higher than the preset layout retrieval threshold are retained as layout candidate sets. Deduplication and merging are performed on anchor records with the same primary key in the semantic candidate set, content candidate set, and layout candidate set. The content fragment units corresponding to the merged anchor records are then used as the candidate anchor set.

4. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 3, characterized in that: The specific steps for calculating the anchoring confidence level to determine the target anchor point are as follows: For each candidate anchor in the candidate anchor set, the cosine similarity between the semantic fingerprint vector of the evidence unit and the candidate anchor is calculated as the semantic score. The inverse of the normalized edit distance between the evidence unit and the candidate anchor point is used as the content score. The intersection-union ratio (IUU) of the bounding boxes of the evidence cells and the candidate anchor points is calculated as the layout score. ; Identify the document type of the materials to be reviewed and determine the semantic baseline weights based on the document type. Content benchmark weight and layout baseline weights ; Among them, the adaptive coupling confidence is calculated. The expression is: ; In the formula, This represents the semantic enhancement factor of the layout. This represents the semantic enhancement factor for the layout, and and It is dynamically fitted based on the historical binding error distribution corresponding to the document type; The candidate anchor point with the highest adaptive coupling confidence is selected as the target anchor point, and the adaptive coupling confidence is used as the anchoring confidence.

5. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 4, characterized in that: The specific steps for injecting topological constraint rules are as follows: Identify causal, parallel, and referential logical relationships between structured bound objects; encode causal logical relationships as directed edges and attach temporal sequence constraints, encode parallel logical relationships as undirected edges and attach mutual exclusion constraints, and encode referential logical relationships as directed edges and attach anchor consistency constraints. Write the encoded constraint attributes into the corresponding edges of the cross-document evidence relationship graph; When calculating the path uncertainty metric, the constraint attribute type is read for each edge. The information entropy contribution weight of the causal logical relationship edge is set to the first multiple of the benchmark weight, the information entropy contribution weight of the parallel logical relationship edge is set to the second multiple of the benchmark weight, and the information entropy contribution weight of the referential logical relationship edge is set to the third multiple of the benchmark weight.

6. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 5, characterized in that: The specific steps for performing heterogeneous adversarial verification to obtain binding verification results and topology verification results are as follows: Binding verification is performed using two heterogeneous methods: semantic consistency verification and physical coordinate reachability verification. The semantic consistency test encodes the evidence unit and the target anchor point separately using a pre-trained language model independent of the first processing track, and calculates the KL divergence of the output probability distributions of the two. When the KL divergence is higher than the preset drift judgment threshold, semantic drift is judged to exist. The physical coordinate reachability check calls the document rendering engine to sample pixels in the bounding box region corresponding to the physical coordinate information of the target anchor point, calculates the proportion of non-blank pixels, and determines that the positioning is invalid when the proportion of non-blank pixels is lower than the preset reachability threshold. The binding verification result is obtained by performing a logical AND operation between the semantic consistency test result and the physical coordinate reachability test result. Topology verification is performed using two heterogeneous methods: constraint satisfaction testing and loop detection. The constraint satisfaction test traverses the constraint attributes of all edges in the cross-document evidence relationship graph and verifies whether the constraint attributes of each edge are violated by the node state. Loop detection uses a depth-first search algorithm to traverse the cross-document evidence relationship graph to determine whether there are illegal circular dependency paths where nodes are accessed repeatedly; The topology verification result is obtained by performing a logical AND operation between the constraint satisfaction test result and the loop detection result.

7. The AI-assisted verification method based on evidence binding and dual-track review as described in claim 6, characterized in that: The specific steps for generating the verification status label are as follows: When both the binding verification result and the topology verification result are passed and the anchor confidence level is higher than the high confidence threshold, a confirmation pass label is generated; When both the binding verification result and the topology verification result are passed, but the anchor confidence level is between the low confidence threshold and the high confidence threshold, a label awaiting manual review is generated. An exception tag is generated when either the binding verification result or the topology verification result fails. When the anchor confidence level is lower than the low confidence threshold, a location failure label is generated regardless of the binding verification result and the topology verification result. Collect historical verification data to confirm the false alarm rate of samples corresponding to the labels. When the false alarm rate is higher than the preset calibration trigger threshold, dynamically lower the high confidence threshold and low confidence threshold based on the false alarm rate deviation.

8. An AI-assisted verification system based on evidence binding and dual-track review, employing the AI-assisted verification method based on evidence binding and dual-track review as described in any one of claims 1 to 7, characterized in that: include: Anchor point construction module, graph deduction module, and adversarial verification module; The anchor point construction module is used to perform multimodal analysis on the materials to be examined to construct a flexible evidence anchor point library, extract evidence units using the first processing track, and perform hybrid retrieval and positioning based on the anchor point library to determine the target anchor point, and encapsulate the evidence units, target anchor points and anchor confidence into structured binding objects. The graph inference module is used to infer logical relationships based on structured binding objects and inject topological constraint rules to construct a cross-document evidence relationship graph, calculate the information entropy of the anchor confidence of each node on the logical path as a measure of path uncertainty, and generate a verification strength configuration strategy accordingly. The adversarial verification module is used to perform heterogeneous adversarial verification according to the verification strength configuration strategy of the second verification track to obtain binding and topology verification results and generate verification status labels, and render a dual-view verification interface; when semantic drift exists, the semantic fingerprint residual vector is calculated and fed back to the first processing track to correct the anchor point library and reposition until convergence and the final verification result is output.

Images