Multi-modal evidence fusion and directional verification method, device, equipment and medium

Abstract

The present application relates to the field of artificial intelligence, and discloses a multi-modal evidence fusion and directional verification method, device, equipment and medium, comprising: mapping entities, time and actions of multi-modal evidence materials to event anchor points to obtain an evidence set; constructing an evidence relationship graph and obtaining source reliability; for support edges and refutation edges of the same proposition, calculating conflict consumption based on proposition importance, reliability of mutually contradictory evidence, conflict coverage range and conclusion decision risk, and accumulating total conflict budget; when the total conflict budget exceeds a preset uncertainty threshold, locating the target proposition with the largest contribution and its target evidence subgraph and verifying; incrementally writing the verification result back to the evidence relationship graph, updating the relationship edges and the source reliability, and outputting a verification conclusion containing certain facts, conditional conclusions and unresolved conflicts. The present application solves the problems of multi-modal heterogeneous evidence conflict positioning difficulty and insufficient fact extraction accuracy, and is suitable for financial technology and medical health scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for multimodal evidence fusion and targeted verification. Background Technology

[0002] When intelligent agents perform complex tasks such as accident investigation, contract review, operational diagnosis, and intelligent risk control in the fintech field and assisted diagnosis in the healthcare field, the original evidence materials they acquire often span multiple modalities. Text reports, on-site photos, system logs, and video clips are intertwined, and different modalities express the core elements of the event in significantly different ways. This heterogeneity poses a great challenge to the intelligent agent's evidence understanding and key fact extraction.

[0003] Currently, conventional techniques typically encode various unstructured materials into high-dimensional vectors, then use similarity calculations or models to generate a single factual conclusion. However, this approach, based on a unified feature space, fails to distinguish the inherent differences in time references and entity representations among modalities, making the system highly susceptible to mismapping acquisition time discrepancies as substantive factual contradictions. Furthermore, regarding evidence source management, existing systems lack effective tracing and merging mechanisms for copies or derivations of the same original material in different formats. This results in a large amount of homologous information being repeatedly weighted as evidence, thus interfering with the credibility assessment of the final conclusion.

[0004] Due to the aforementioned shortcomings, current intelligent agents are prone to deviations in their internal reasoning logic when faced with cross-modal evidence due to inaccurate alignment of basic data. This is especially true in the process of retrospective analysis of financial transaction disputes or review of multimodal medical records of patients, where lengthy manual intervention or cumbersome secondary verification processes are often required to offset such fundamental interference, which significantly restricts the effectiveness and response efficiency of intelligent agents in real-world scenarios. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for multimodal evidence fusion and targeted verification to solve the following technical problem: how to reliably anchor and structure multimodal heterogeneous evidence in order to effectively locate substantive factual conflicts and improve the accuracy of fact extraction and response efficiency.

[0006] Firstly, a multimodal evidence fusion and targeted verification method is provided, including: In response to obtaining multimodal evidence materials, the entities, time, and actions in each piece of evidence material are mapped to event anchor points to obtain an evidence set associated with the event anchor points; Based on the evidence set, an evidence relationship graph is constructed, consisting of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and the source reliability associated with the relationship edges is obtained. For the supporting edge and the rebuttal edge of the same proposition to be verified, the conflict cost is calculated based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and the total conflict budget is obtained by summing up the conflict costs of each edge. When the total conflict budget exceeds a preset uncertainty threshold, locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph, verify the target evidence subgraph, and obtain the verification result. The verification result is incrementally written back to the evidence relationship graph, the affected relationship edges and the source reliability are updated, and the verification conclusion containing the established facts, conditional conclusions and unresolved conflicts is output based on the updated evidence relationship graph.

[0007] Secondly, a multimodal evidence fusion and targeted verification device is provided, including: A mapping unit is used to map the entities, time, and actions in each piece of evidence to an event anchor point in response to the acquisition of multimodal evidence materials, thereby obtaining an evidence set associated with the event anchor point; The construction unit is used to construct an evidence relationship graph based on the evidence set, which consists of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and the source reliability associated with the relationship edges is obtained. The calculation unit is used to calculate the conflict cost for the supporting edge and the rebuttal edge of the same proposition to be verified, based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and to sum up the conflict costs to obtain the total conflict budget. The verification unit is used to locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph when the total conflict budget exceeds a preset uncertainty threshold, and to verify the target evidence subgraph to obtain the verification result. The update unit is used to write the verification result incrementally back to the evidence relationship graph, update the affected relationship edges and the source reliability, and output a verification conclusion containing established facts, conditional conclusions and unresolved conflicts based on the updated evidence relationship graph.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described multimodal evidence fusion and targeted verification method.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described multimodal evidence fusion and targeted verification method.

[0010] In the aforementioned scheme implemented by the multimodal evidence fusion and targeted verification method, device, computer equipment, and storage medium, cross-modal event anchor points are constructed to uniformly map entities, times, and actions in various evidence materials to these event anchor points. This fundamentally eliminates pseudo-conflicts caused by differences in the expression benchmarks of different modalities, achieving reliable anchoring of multimodal heterogeneous evidence. Furthermore, based on the evidence set, an evidence relationship graph containing supporting and rebuttal edges is constructed, and the source reliability of the associations is obtained. This graph structure systematically represents the logical adversarial relationships between evidence, completing the structured organization of scattered evidence. Further, based on this, the conflict consumption is calculated and accumulated by comprehensively considering the importance of the proposition, the reliability of contradictory evidence, the scope of conflict coverage, and the risk of conclusion decision, thus obtaining the total conflict budget. This not only transforms qualitative contradictions into quantifiable indicators that can be ranked, but also leverages a preset uncertainty threshold. The method accurately locates the target proposition and its target evidence subgraph that contributes the most to the total conflict budget, avoiding blind and comprehensive verification from the perspective of verification space. Furthermore, it performs targeted verification on the located target evidence subgraph and uses an incremental write-back method to update the affected relation edges and the source reliability. While maintaining the real-time consistency of the evidence graph, it effectively reduces the computational overhead of a single verification, thereby avoiding the processing waiting time caused by full graph recalculation and significantly improving the overall response efficiency of verification conclusion generation. The final output verification conclusion provides a hierarchical credibility basis for the fact extraction results by clearly distinguishing between established facts, conditional conclusions, and unresolved conflicts. This ensures the accuracy of decision-making while accelerating the delivery of credibility information, effectively solving the technical problems of difficult conflict localization, insufficient accuracy of fact extraction, and limited response efficiency in multimodal heterogeneous evidence fusion. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the 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.

[0012] Figure 1 This is a schematic diagram of an application environment for a multimodal evidence fusion and targeted verification method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a multimodal evidence fusion and targeted verification method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a multimodal evidence fusion and targeted verification device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The multimodal evidence fusion and targeted verification method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain multimodal evidence materials from the client, including text reports, on-site images, voice recordings, video clips, system logs, and structured business records. The server maps the entities, times, and actions in each piece of evidence material to event anchors, obtaining an evidence set associated with the event anchors. Based on the evidence set, an evidence relationship graph is constructed, where the relationship edges include at least supporting and rebuttal edges, and the source reliability associated with each relationship edge is obtained. For supporting and rebuttal edges of the same proposition, conflict consumption is calculated based on proposition importance, reliability of contradictory evidence, conflict coverage, and decision risk, and the total conflict budget is accumulated. When the total conflict budget exceeds a preset uncertainty threshold, the target proposition and its corresponding target evidence subgraph are located, and the target evidence subgraph is verified to obtain a verification result. The verification result is incrementally written back to the evidence relationship graph, the affected relationship edges and source reliability are updated, and a verification conclusion containing established facts, conditional conclusions, and unresolved conflicts is output based on the updated evidence relationship graph. The verification conclusion is then fed back to the client. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, or it can be an intelligent agent application deployed on the terminal device. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0015] Please see Figure 2 , Figure 2This is a flowchart illustrating a multimodal evidence fusion and targeted verification method provided by an embodiment of the present invention. This method can be widely applied in complex evidence review and decision support scenarios executed by intelligent agents, and is particularly suitable for accident investigation, contract review, identification and compliance review of transaction fraud in the fintech field, and auxiliary diagnosis and clinical pathway analysis based on multimodal medical records in the healthcare field. In these application scenarios, intelligent agents typically need to simultaneously process multimodal heterogeneous evidence materials such as text reports, on-site images, voice recordings, video clips, system logs, and structured business records. Data from different modalities naturally differ in their information expression methods. For example, an eyewitness text report might use natural language to describe "equipment failure occurred around 10:00," while the system log records it as "09:58:23," and image metadata (such as EXIF ​​information) shows the shooting time as "10:03:17." If these numerical values ​​from different sources are simply compared, non-substantial differences caused by clock errors in the acquisition equipment or ambiguities in natural language descriptions can easily be misjudged as factual conflicts. Furthermore, when faced with conflicting evidence, existing systems often adopt a simple majority voting strategy or blindly invoke all verification tools for comprehensive verification without considering the importance of the conflict itself, the execution cost of the verification tools, and the expected benefits after invocation. This results in high verification costs and a lack of interpretable selection criteria. Therefore, this embodiment addresses the issues of misjudgment and high cost in multimodal fusion by constructing cross-modal event anchors to unify and align dispersed multimodal evidence into a single structured framework and driving targeted verification through a conflict budget mechanism. Specifically, the method includes the following steps: S1, in response to obtaining multimodal evidence materials, the entities, time and actions in each piece of evidence material are mapped to event anchor points to obtain an evidence set associated with the event anchor points.

[0016] In specific implementation, in response to the acquisition of multimodal evidence materials, the entities, times, and actions in each piece of evidence material are mapped to event anchors. The event anchor is a standardized structured data object used to carry multiple core elements of an event, specifically including subject fields, object fields, action fields, time range fields, location fields, and status version fields, etc., which are not specifically limited in this invention. Specifically, the subject field records the core individuals, equipment, or organizational entities involved in the event; the object field records the entity objects that the event affects or influences; the action field describes the specific behavior or process that occurred during the event; the time range field records the precise time point or fuzzy time interval of the event; the location field records the geographical coordinates, network address, or physical space identifier of the event; and the status version field records the current status version or snapshot identifier of the event, used to distinguish the evolution of the same event at different stages.

[0017] Furthermore, by parsing the metadata and text / content information in the evidence materials, and using a data parsing engine and natural language processing tools, relevant entities (such as personnel and equipment), time descriptions (such as "10:00" and "09:58"), and action descriptions (such as "failure" and "startup") are extracted from carriers such as text, logs, and image metadata. The extracted information is then populated into the corresponding fields of the event anchor points according to their semantics, thus forming structured evidence records and ultimately obtaining an evidence set associated with the event anchor points. This evidence set contains one or more anchor-aligned evidence records, providing standardized data input for subsequent relationship graph construction.

[0018] In some preferred embodiments, the event anchor point includes a subject field, an object field, an action field, a time range field, a location field, and a status version field. Mapping the entities, times, and actions in each piece of evidence material to the event anchor point includes: performing structured parsing on the evidence material to extract structured metadata carried in the evidence material, the structured metadata including device identifier, collection timestamp, and geographic location coordinates; performing semantic recognition on the content of the evidence material to extract unstructured entities recorded in the evidence material, the unstructured entities including people, objects, and action descriptions; mapping the collection timestamp in the structured metadata to the time range field, mapping the geographic location coordinates to the location field, mapping the device identifier and people in the unstructured entities to the subject field, mapping objects in the unstructured entities to the object field, and mapping action descriptions in the unstructured entities to the action field. If the structured metadata and the unstructured entity have semantic conflicts in the subject field, the time range field, or the location field, the corresponding field values ​​are set as a candidate set, which includes the corresponding values ​​of the structured metadata and the unstructured entity. If the structured metadata and the unstructured entity do not have semantic conflicts in the subject field, the time range field, or the location field, the corresponding value of the structured metadata is mapped to the corresponding field in the event anchor. If the collection timestamp in the structured metadata conflicts with the time point pointed to by the action description in the unstructured entity, the collection timestamp is retained as a time interval, which includes the time point pointed to by the action description. If the collection timestamp in the structured metadata does not conflict with the time point pointed to by the action description in the unstructured entity, the collection timestamp is mapped to the time range field.

[0019] In practice, the structures of evidence materials from different modalities differ significantly. By setting up semantic conflict detection and time point conflict detection mechanisms between structured metadata and unstructured entities, the accuracy and robustness of event anchor field filling can be greatly improved. The specific implementation method is as follows: First, the evidence materials are subjected to structured parsing to extract the structured metadata carried within them. This structured metadata is standard-format data automatically created and embedded in the file by the device or software when generating the evidence. It typically exists as standard fields in the file header, attribute page, or specific metadata areas. For example, image files may contain Exif (Exchangeable Image File Format) information, system logs contain timestamps, process IDs, and hostnames, and business databases contain standard field values. From this structured metadata, the device identifier, acquisition timestamp, and geographic coordinates are extracted. The device identifier is used to characterize the serial number of the hardware device that generated the material or the globally unique identifier of the software process; the acquisition timestamp records the system clock time when the material was generated; and the geographic coordinates record the spatial longitude and latitude information of the acquisition device. These data typically possess high objectivity and accuracy.

[0020] Furthermore, semantic recognition is performed on the content of the evidence materials to extract unstructured entities recorded in them. These unstructured entities refer to natural language descriptive information existing in text, audio transcriptions, video subtitles, or image content descriptions, such as names of people, names of objects, and descriptions of actions. For text reports, the system can accurately locate people (e.g., "Operator A"), objects (e.g., "Reaction Vessel B"), and action descriptions (e.g., "Overpressure") from longer paragraphs using natural language processing techniques such as named entity recognition and relation extraction. For audio or video materials, the system first converts them into text and then performs the aforementioned NLP techniques for extraction. The extraction of unstructured entities relies on contextual semantic understanding, thus offering greater flexibility, but may also result in ambiguity or polysemy in expression.

[0021] Furthermore, after acquiring the structured metadata and the unstructured entities, the structured metadata and the unstructured entities are mapped to corresponding fields. The event anchor includes a subject field, an object field, an action field, a time range field, a location field, and a status version field. Specifically, the collection timestamp in the structured metadata is mapped to the time range field, and the geographic location coordinates are mapped to the location field. Further, the device identifier and the person in the unstructured entity are mapped to the subject field, the object in the unstructured entity is mapped to the object field, and the action description in the unstructured entity is mapped to the action field. After the field mapping is completed, the system integrates multi-source data into a unified anchor object, achieving standardization of heterogeneous information.

[0022] Furthermore, since the structured metadata and the unstructured entity may originate from different information channels and have different generation mechanisms, their content may be inconsistent. To address this, a semantic conflict detection branch is set up. If the structured metadata and the unstructured entity have a semantic conflict in the subject field, the time range field, or the location field, the corresponding field value is set as a candidate set. The candidate set simultaneously includes the corresponding value of the structured metadata and the corresponding value of the unstructured entity. For example, if the device identifier (D123) recorded in the system log and the operator name (Zhang San) described in the text report point to different entities, and it cannot be determined whether they are consistent, then both "D123" and "Zhang San" are retained in the candidate set, instead of forcibly selecting one, thus preserving the ambiguity of objective facts. Furthermore, if the structured metadata and the unstructured entity do not have a semantic conflict in the subject field, the time range field, and the location field, then the corresponding value of the structured metadata is mapped to the corresponding field in the event anchor. This approach is based on the principle that structured metadata typically has higher objectivity and is less susceptible to subjective descriptions.

[0023] Furthermore, to address the different time representation methods, an additional time point conflict detection is implemented. If the collection timestamp in the structured metadata conflicts with the time point pointed to by the action description in the unstructured entity, the collection timestamp is retained as a time interval, which includes the time point pointed to by the action description. If there is no conflict, the collection timestamp is directly mapped to the time range field. This mechanism avoids false conflicts caused by simple numerical comparisons and significantly improves the system's ability to handle fuzzy time.

[0024] This embodiment extracts structured metadata and unstructured entities separately, and fills in the structured data of event anchors based on field correspondences. This standardizes and aligns the expressive features of the same event across different modalities, improving the fusion capability of cross-modal data. Furthermore, by setting a semantic conflict detection branch, when conflicts exist between the two in the subject or location fields, the system uses a candidate set retention method instead of forcing a binary choice, effectively avoiding misjudgments caused by information asymmetry and preserving the ambiguity and uncertainty of the original data. Furthermore, by additionally setting time point conflict detection, when there is a deviation between the collected timestamp and the time point described in natural language, the system accommodates the difference using a time interval rather than simply judging it as contradictory, avoiding false conflicts caused by minor clock deviations. This multi-level mapping and conflict inclusion mechanism enables event anchors to objectively carry time and entity information of different precision and sources, providing a more reliable data foundation for subsequent relationship graph construction and conflict calculation.

[0025] In some preferred embodiments, after obtaining the set of evidence associated with the event anchor, the method further includes: obtaining the source file path, acquisition device serial number, and hash value of the content digest of each piece of evidence; sequentially comparing whether the source file paths of any two pieces of evidence are the same; if the source file paths of the two pieces of evidence are the same, then it is determined that the two pieces of evidence have a direct common source relationship; if the source file paths of the two pieces of evidence are different, then it is compared whether the acquisition device serial numbers of the two pieces of evidence are the same, and also compared whether the hash values ​​of the content digests of the two pieces of evidence are the same; if the acquisition device serial numbers of the two pieces of evidence are the same... If the hash values ​​of the content digests of two pieces of evidence are the same, then the two pieces of evidence are determined to have an indirect homology relationship; if the serial numbers of the acquisition devices of the two pieces of evidence are different or the hash values ​​of the content digests of the two pieces of evidence are different, then the two pieces of evidence are determined not to have a homology relationship; based on the direct homology relationship or the indirect homology relationship, a source chain identifier is constructed for multiple pieces of evidence with a homology relationship, and multiple pieces of evidence with the same source chain identifier are merged into the same evidence group, wherein the evidence group retains only one contribution weight when associating the event anchor point, and the contribution weight is equal to the normalized value of the weight of each piece of evidence in the evidence group.

[0026] In practical applications, the same original source (such as the same device or the same original file) often results in multiple copies through duplication, dissemination, and renaming. If these multiple copies are mistakenly recognized as independent evidence by the system, the weight of the evidence will be incorrectly amplified, severely affecting the credibility of the final conclusion. This embodiment effectively solves this problem of duplicate weighting by establishing a structured mechanism for identifying and merging sources of similarity. The specific implementation method is as follows: After obtaining the set of evidence associated with the event anchor, the system executes a separate process for identifying evidence from the same source. First, it acquires the source file path, acquisition device serial number, and hash value of the content digest for each piece of evidence. The source file path refers to the complete storage path of the evidence material in the underlying file system. Further, the acquisition device serial number is a unique identifier for the physical device (e.g., camera, sensor) or software process (e.g., log collection agent) that generated the evidence, used to trace the direct producer of the material. Further, the hash value of the content digest is a fixed-length (e.g., 256-bit) digital fingerprint calculated using a secure hash algorithm (e.g., SHA-256). This fingerprint is highly sensitive to minute changes in content; when two files have identical content, their hash values ​​will be identical; when there are any subtle differences in content (e.g., a difference of one bit), the hash values ​​will be completely different, thus serving as a strong check for the uniqueness of the evidence content.

[0027] Furthermore, after obtaining the above three attributes, the source file paths of any two pieces of evidence are compared sequentially to see if they are the same. If the source file paths of the two pieces of evidence are the same, it is determined that the two pieces of evidence have a direct common source relationship, indicating that they are actually the same file stored in the same physical location or logical path at the system's underlying level, but have been repeatedly read or imported by the system multiple times.

[0028] Furthermore, if the source file paths of the two pieces of evidence are different, the serial numbers of the acquisition devices and the hash values ​​of the content digests of the two pieces of evidence are further compared. If the serial numbers of the acquisition devices are the same and the hash values ​​of the content digests are the same, it is determined that the two pieces of evidence have an indirect common source relationship. This situation usually occurs when the same scene is captured by the same device, but the image files are renamed, have their metadata changed, or are stored in different directories, resulting in different paths but identical substantive content.

[0029] Furthermore, if the serial numbers of the acquisition devices are different or the hash values ​​of the content digests are different, then the two pieces of evidence are determined to be not related by origin, that is, they belong to two independent sources and should be regarded as independent evidence for evaluation.

[0030] Furthermore, after completing all pairwise comparisons, a source chain identifier is constructed for multiple pieces of evidence with a source relationship based on the direct or indirect homology. The source chain identifier is a unified number or string used to associate all evidence materials originating from the same original source, forming a tracing chain. Further, multiple pieces of evidence with the same source chain identifier are merged into the same evidence group. When associating the evidence group with the event anchor point, only one contribution weight is retained. The contribution weight is calculated by normalizing the original confidence level or weight of each piece of evidence within the evidence group to obtain a comprehensive weight value.

[0031] Specifically, assuming an evidence set contains n pieces of evidence from the same source, each with an original confidence level of c1, c2, ..., cn, the normalized composite weight is W = (c1 + c2 + ... + cn) / n. In this way, regardless of the number of pieces of evidence from the same source within the evidence set, the total influence weight in the evidence diagram is only a single composite value, rather than a simple summation of the weights of each piece of evidence. This objectively reflects the true credibility of the source and avoids weight imbalances caused by evidence duplication.

[0032] This embodiment establishes a structured and traceable same-source identification mechanism by acquiring the source file path, acquisition device serial number, and hash value of the content summary of the evidence materials, and determining direct and indirect same-source relationships based on pairwise comparisons. Furthermore, by constructing a unified source chain identifier for same-source materials and merging multiple same-source materials into an evidence group, the system effectively identifies and eliminates redundant information generated through copying from the same original source, avoiding the accumulation of false evidence credibility due to duplicate weighting. Furthermore, by retaining only one normalized contribution weight, the system can objectively reflect the true credibility level of the source in the evidence association stage, rather than blindly weighting based on the number of copies, significantly improving the rationality of the evidence relationship graph weight allocation and ensuring the fairness of subsequent conflict calculations and verification decisions.

[0033] S2. Based on the evidence set, construct an evidence relationship graph consisting of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and obtain the source reliability associated with the relationship edges.

[0034] In specific implementation, after obtaining the evidence set, an evidence relationship graph is constructed based on the evidence set, consisting of event anchors, evidence materials, propositions to be verified, and verification results. The evidence relationship graph is a graph data structure used to represent the complex logical connections between evidence, between evidence and propositions to be verified, and between event anchors and verification results. Further, in the evidence relationship graph, the relationship edges include at least supporting edges, rebuttal edges, temporal edges, and common-origin edges. Specifically, a supporting edge indicates that a piece of evidence supports the truth of a proposition to be verified; a rebuttal edge indicates that a piece of evidence opposes the proposition to be verified; a temporal edge indicates the chronological order of events or evidence, including strict earlier, later, or simultaneous relationships, which have strong temporal constraints and directionality; and a common-origin edge indicates that different pieces of evidence originate from the same original source, used to identify the derivational relationship between evidence. By introducing these various types of relationship edges, the multi-dimensional factual connections and logical adversarial relationships in the evidence network can be comprehensively represented.

[0035] Furthermore, it is also necessary to obtain the source reliability associated with the aforementioned relationship edge. The source reliability is a dynamic quantitative assessment of the trustworthiness of the evidence provider (such as a log system, sensor device, or human reporter), with a value falling within the range of [0,1]. This value is adjusted in real time based on the source's historical performance, anomalies in the current event, and subsequent verification results to ensure the objectivity of the weight allocation.

[0036] In some preferred embodiments, obtaining the reliability of the sources associated with the relationship edges includes: obtaining the historical accuracy rate of each source in historical verification tasks, where the historical accuracy rate is the proportion of evidence output by the source that is subsequently verified as true; for each source, obtaining a material integrity score of the evidence material corresponding to the source based on the completeness of fields in the evidence material and the signal quality corresponding to the modality type of the evidence material; for each source, retrieving independent evidence that is associated with the evidence material corresponding to the source from the evidence relationship graph, comparing the consistency between the evidence material of the current source and the independent evidence on each field of the event anchor point, and according to the... The consistency degree is obtained by scoring the consistency degree. The independent evidence comes from other sources different from the current source and has a temporal or causal relationship with the evidence materials of the current source. The system detects whether there are data anomalies in the evidence materials associated with the source in the current event. The data anomalies include logical conflicts in the collection timestamps of the evidence materials, failure to verify the integrity check value of the evidence materials, or values ​​in the evidence materials exceeding a preset normal range. The system performs a weighted summation of the historical accuracy, the material integrity score, the consistency degree score, and the detection results of the data anomalies based on predetermined weights, and normalizes the weighted summation result to obtain the source reliability of the source.

[0037] In specific implementation, this embodiment introduces a dynamic multi-factor calibration mechanism, making the source reliability a dynamic indicator that can adapt to new data environments in real time, thereby ensuring that the weight allocation always conforms to the latest evidence status. The specific implementation method is as follows: First, obtain the historical accuracy rate for each source in past verification tasks. The historical accuracy rate refers to the ratio between the number of pieces of evidence verified as genuine by that source in past verification tasks and the total amount of evidence output by that source. For example, if a log system outputs records that were subsequently verified as genuine by humans or third-party systems in 85 out of the past 100 verifications, then the historical accuracy rate of that source is 0.85. This ratio reflects the long-term historical credibility of the source and is an important basis for reliability assessment.

[0038] Furthermore, for each source, a material integrity score is obtained based on the completeness of the fields in the evidence material and the signal quality corresponding to the modality type of the evidence material. Field completeness refers to whether the evidence record provided by the source contains complete core fields, such as explicit timestamps, subject identifiers, or checksums. If a field is missing, the score will decrease accordingly. The signal quality corresponding to the modality type is evaluated based on the specific characteristics of the data type, such as the resolution, signal-to-noise ratio, and compression artifact level of image materials; the sampling rate and dynamic range of audio materials; and the frame rate and bit rate of video materials. Higher completeness and better signal quality result in a higher score.

[0039] Further, for each source, independent evidence that is related to the evidence material corresponding to that source is retrieved from the evidence relationship graph. Independent evidence refers to evidence originating from sources other than the current source and having a temporal (e.g., occurring at the same time) or causal (e.g., action and result) relationship with the evidence material of the current source. A consistency score is obtained by comparing the consistency between the evidence material of the current source and the independent evidence in each field of the event anchor point. The consistency degree is determined by comparing the corresponding values ​​of the evidence material of the current source and the independent evidence in each field of the event anchor point field by field. Specifically, consistency checks are performed on the subject field, the time range field, the location field, and the action field using methods such as entity alignment judgment and comparison of time difference with a preset tolerance threshold. Based on this, the proportion of consistent fields is statistically analyzed, and a weighted sum is obtained by combining the preset weights of each field. This weighted consistency ratio is then mapped to a preset scoring range to obtain the consistency score. This score is used to quantify the logical consistency between the evidence of the current source and the independent evidence in key elements. The higher the consistency, the better the evidence from that source matches independent objective facts; the higher the score, the stronger its credibility.

[0040] Furthermore, the system detects whether there are any data anomalies in the evidence materials associated with the source in the current event. These data anomalies include, but are not limited to: logical conflicts in the collection timestamps of the evidence materials, such as timestamps exceeding the normal operating time range of the system or containing future times; failure to verify the integrity of the evidence materials, such as MD5 checksums of files not matching the original records, indicating that the files may have been tampered with; and values ​​in the evidence materials exceeding preset normal ranges, such as sensor readings exceeding the upper and lower limits of the device's physical measurement range, indicating potential data collection errors.

[0041] Furthermore, the data anomaly detection result refers to the judgment conclusion obtained after detecting the evidence materials associated with the source in the current event regarding whether the evidence materials have any abnormalities, and the numerical score converted based on the judgment conclusion. For example, when a logical conflict is detected in the collection timestamp of the evidence materials, the integrity verification value fails, or the value exceeds the preset normal range, the system determines that the source has data anomalies and quantifies such anomalies as a low confidence value (such as 0.2 or 0.5); when no anomalies are detected, they are quantified as a high confidence value (such as 1.0), thereby obtaining the data anomaly score.

[0042] Furthermore, after obtaining the four indicators—historical accuracy, material integrity score, consistency score, and data anomaly detection result—the four indicators are weighted and summed based on predetermined weights, and the weighted summation result is normalized to obtain the source reliability. Before participating in the weighted calculation, the historical accuracy, material integrity score, consistency score, and data anomaly score are all uniformly mapped to dimensionless values ​​in the interval [0,1]. Specifically, the historical accuracy is directly converted into a score based on historical statistical proportions; the material integrity score is assessed based on the completeness of data fields and signal quality; the consistency score is assessed based on the logical consistency with independent evidence; and the data anomaly score is assessed based on whether the current evidence has real-time logical conflicts or verification failures. These four indicators comprehensively measure the credibility level of the same source from four different dimensions: historical reputation, material quality, logical consistency, and real-time health. Each value represents the contribution to the goal of "source credibility" under that dimension, and therefore can be weighted and summed according to preset weights. The system normalizes the weighted result back to the [0,1] interval, thus obtaining the source reliability characterizing the overall credibility level of the source. Specifically, if the weights of historical accuracy, material integrity score, consistency score, and data anomaly detection result are w1, w2, w3, and w4, respectively, and the scores for each are v1, v2, v3, and v4, then the source reliability R = normalize(w1v1 + w2v2 + w3v3 + w4v4). The normalization process ensures that the output value of the source reliability always falls within the [0,1] interval, facilitating its direct use as a weighting factor in subsequent conflict consumption calculations. This dynamic calculation process is triggered in each verification cycle or when specific conditions are met, ensuring that the source reliability is always up-to-date and most accurate.

[0043] This embodiment dynamically calibrates source reliability by integrating four dimensions: historical accuracy, material integrity score, consistency with independent evidence, and the presence of data anomalies in the current event. This transforms reliability assessment from a static, fixed value into a dynamic indicator reflecting the source's real-time performance in the current specific task. Furthermore, by introducing a data anomaly detection mechanism, the reliability of a source decreases when it experiences temporal logic conflicts or verification failures in the current event. This reduces the source's influence weight in subsequent conflict consumption calculations and verification location, preventing a single anomalous piece of evidence from misleading the entire system's decision-making. This multi-dimensional dynamic calibration mechanism allows the weights of supporting and rebuttal sides to more accurately reflect the objective credibility level of the evidence, significantly improving the system's sensitivity and accuracy in perceiving source credibility.

[0044] S3. For the supporting edge and the rebuttal edge of the same proposition to be verified, calculate the conflict cost based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and sum up the conflict costs to obtain the total conflict budget.

[0045] In practice, after constructing the evidence relationship graph, for the supporting and rebuttal edges of the same proposition to be verified, the conflict consumption is calculated based on the proposition's importance, the reliability of contradictory evidence, the scope of conflict coverage, and the risk of the conclusion decision. The conflict consumption is used to quantify the specific destructive power of the contradictions involved in the proposition on the final decision. The following precise mathematical formula is used to calculate the conflict consumption: C_j=Importance_j×Σ(Rel_support×Rel_refute)×Scope_j×DecisionRisk_j; In this formula, C_j represents the conflict cost calculated for the j-th proposition. The larger the value, the more severe the impact of the conflict on the global judgment. Importance_j represents the importance of the proposition to the final decision. Its value is determined by the task type, domain security level, or user-preset threshold. Propositions with higher importance have higher weight even if their conflicts are small. Rel_support and Rel_refute represent the source reliability corresponding to the supporting edge and the refuting edge, respectively. The values ​​of both are derived from the aforementioned dynamic evaluation results, and their range is [0,1]. Σ(Rel_support×Rel_refute) e) represents the summation of the products of all supporting and opposing sides of the proposition. This summation reflects the intensity of the antagonistic relationship between the "evidence hardness" of the conflicting parties. Specifically, when both the supporting and opposing sides are highly reliable, this product term increases significantly, indicating the most intense conflict and extremely high trust levels. Scope_j represents the number of evidence nodes or the length of the decision chain covered by the conflict, i.e., the scope of the conflict's impact. The wider the scope, the more profound its impact on the system. DecisionRisk_j represents the severity of the consequences that might result from ignoring the conflict and making an incorrect decision, such as equipment damage, economic loss, personal safety risks, or legal compliance risks. This formula, through the multiplication of four factors, transforms qualitative semantic contradictions into quantitative computational indicators, thus providing an objective mathematical basis for subsequent automated verification.

[0046] In some preferred embodiments, calculating conflict consumption based on proposition importance, the reliability of contradictory evidence, the scope of conflict coverage, and the risk of conclusion decision includes: obtaining the proposition importance, the scope of conflict coverage, and the risk of conclusion decision for the same proposition to be verified; determining the first reliability corresponding to the supporting edge and the second reliability corresponding to the rebuttal edge of the same proposition to be verified; calculating the product of the first reliability and the second reliability, and summing the products of all supporting edges and rebuttal edges of the same proposition to be verified to obtain the total reliability product; and multiplying the proposition importance, the total reliability product, the scope of conflict coverage, and the risk of conclusion decision to obtain the conflict consumption of the same proposition to be verified.

[0047] In specific implementation, for the supporting and rebuttal edges of the same proposition to be verified, the importance of the proposition, the scope of conflict coverage, and the risk of the conclusion decision are first obtained. The importance of the proposition is a weight set by the system or user based on its core position in the entire event to be verified. For example, in an accident investigation, a proposition concerning "whether the equipment was operating improperly" usually has higher importance than a proposition concerning "whether the report author was correctly signed," because the former directly affects the determination of the root cause of the accident, involving equipment safety and human safety. The scope of conflict coverage refers to the number of evidence nodes or the length of the decision chain affected by the proposition. For example, if a conflict affects half of the nodes in the entire evidence graph, its coverage is wide and its impact is large; conversely, if it only involves a very small number of evidence nodes, its scope is small. The risk of the conclusion decision refers to the severity of the adverse consequences that may result if the conflict is ignored and an incorrect decision is made directly, such as equipment damage, economic loss, personal safety risks, or legal compliance risks.

[0048] Furthermore, a first reliability corresponding to the supporting edge and a second reliability corresponding to the rebuttal edge of the same proposition to be verified are determined respectively. Here, the reliability directly refers to the aforementioned dynamically calibrated source reliability. The reliability of each supporting and rebuttal edge represents the degree of credibility of the evidence supporting and opposing the proposition.

[0049] Furthermore, after determining the first reliability and the second reliability, the product of the first reliability and the second reliability is calculated, and the products of all supporting and opposing sides of the same proposition to be verified are summed to obtain the total reliability product. This total reliability product reflects the antagonistic strength between the evidence strength of the conflicting parties. When both supporting and opposing evidence are highly reliable, the product is larger, and the conflict is more intense; when either side's evidence is unreliable, the product decreases, and the conflict intensity weakens. This term is the core regulating term in conflict consumption, allowing subtle differences in evidence strength to be amplified or reduced.

[0050] Furthermore, the conflict consumption of the same proposition to be verified is obtained by multiplying the sum of the products of the proposition importance, the reliability, the conflict coverage, and the decision risk of the conclusion. This calculated result, C_j, is a single quantitative measure of the intensity of the contradiction formed by the proposition in the current evidence diagram. The larger this value, the more severe the impact of the contradiction of the proposition on the final decision, and the more it needs to be addressed first. Through this multiplicative structure, the four factors can form a combined force; only when all four factors are high will the final conflict consumption be very high, which actually reflects a comprehensive constraint.

[0051] This embodiment establishes a complete, quantifiable, and interpretable conflict assessment model by decomposing conflict consumption into the product of four factors: proposition importance, the sum of products of the reliability of contradictory evidence, the scope of conflict coverage, and the risk of the conclusion decision. By introducing the sum of reliability products, conflict consumption dynamically reflects the adversarial strength of evidence, avoiding judgment based solely on the strength of one side of the evidence. Furthermore, the calculation structure of multiplying the four factors makes the degree of conflict between different propositions comparable, allowing the system to objectively rank conflicts and prioritize the core contradictions that have the greatest impact on decision-making. By comparing with a preset uncertainty threshold, the system can accurately locate the key contradictions that truly affect the decision, providing an objective mathematical basis for subsequent targeted verification.

[0052] S4. When the total conflict budget exceeds a preset uncertainty threshold, locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph, verify the target evidence subgraph, and obtain the verification result.

[0053] In practice, after obtaining the conflict costs of each proposition, the conflict costs are summed to obtain the total conflict budget. The total conflict budget reflects the overall level of contradiction currently existing in the entire evidence graph. Further, it is determined whether the total conflict budget exceeds a preset uncertainty threshold. The preset uncertainty threshold is a dynamic parameter determined based on the security level of the task's domain, historical experience, or user-input tolerance. For example, in high-security scenarios such as financial transaction fraud investigation, this threshold may be set lower to tolerate less controversy and uncertainty; while in routine operation and maintenance diagnosis scenarios, this threshold can be relatively relaxed to reduce unnecessary computational costs. If the total conflict budget exceeds the preset uncertainty threshold, it indicates that there are significant contradictions in the current evidence graph that cannot be covered by existing knowledge, requiring targeted external verification. Further, the system will locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph. The target evidence subgraph contains all evidence nodes, relationship edges, and related evidence sets supporting or refuting the proposition directly related to the target proposition. Further, targeted verification is performed on the target evidence subgraph. Verification methods include, but are not limited to, retrieving original database records, rereading images with metadata, calibrating audio timelines, or querying trusted third-party systems, etc. This invention does not specifically limit these methods, thereby obtaining verification results. This verification is highly focused and low-cost because it targets only localized hotspots of conflict, rather than the entire dataset.

[0054] In some preferred embodiments, verifying the target evidence subgraph includes: determining key conflict fields from the target evidence subgraph, wherein the key conflict fields are fields where the confidence difference between the supporting edge and the rebuttal edge is less than a preset difference threshold; obtaining the range of verification fields supported by each candidate verification tool, and selecting candidate verification tools whose verification field range covers the key conflict fields to form a candidate tool list; for each candidate verification tool in the candidate tool list, predicting the amount of conflict consumption that the candidate verification tool can reduce for the target evidence subgraph, as verification. The system calculates the benefits and predicts the execution cost, time delay, and side effects of the candidate verification tools, using the sum of these as the comprehensive cost. It then calculates the utility value of each candidate verification tool, where the utility value is the verification benefit divided by the comprehensive cost. The candidate verification tools in the list are sorted from highest to lowest utility value, and each tool is selected sequentially until the selected set covers all the key conflict fields. The selected set of verification tools is then identified as the target verification tool, and the target verification tool is invoked to verify the target evidence subgraph.

[0055] In specific implementation, this embodiment introduces the concept of "utility value" and establishes a quantitative model for the input-output ratio of verification tool selection. This ensures that when faced with multiple optional verification schemes, the system can reasonably allocate limited intelligent agent computing resources, solving the problem of excessively high costs associated with blindly calling all tools for comprehensive verification. The specific implementation method is as follows: After locating the target proposition and its corresponding target evidence subgraph, the key conflict field is first determined from the target evidence subgraph. The key conflict field refers to the specific data field or attribute that causes the contradiction. In this embodiment, the key conflict field is defined as the field where the confidence difference between the supporting edge and the opposing edge is less than a preset difference threshold. The preset difference threshold is an empirical value, such as 0.1. When the confidence levels of the supporting and opposing sides are close, neither side can provide an overwhelming advantage. In this case, verifying this field is most likely to eliminate the conflict, and therefore it is the field most in need of verification.

[0056] Furthermore, the range of verification fields supported by each candidate verification tool is obtained. The candidate verification tools refer to various external verification or reprocessing mechanisms that the agent can currently invoke, such as raw database query tools, image metadata re-parsing tools, audio timeline calibration tools, or third-party trusted system interfaces; this invention does not specifically limit these. Different tools can verify different ranges of fields, which determines the applicable scenarios and effectiveness of the tools. For example, database query tools can verify the accuracy of timestamps, image parsing tools can verify the metadata of the capturing device, and audio calibration tools can align the timeline.

[0057] Furthermore, candidate verification tools are selected from the candidate verification tools whose verification field range covers the key conflicting field, forming a list of candidate tools. This step eliminates invalid tools that cannot verify the key field, narrows the candidate range, and allows subsequent calculations to focus on the truly relevant set of tools.

[0058] Furthermore, after obtaining the list of candidate tools, a dual prediction is performed for each candidate verification tool in the list. First, the amount of conflict consumption that the candidate verification tool can reduce for the target evidence subgraph is predicted, as the verification benefit. The verification benefit is calculated as: the conflict consumption of the current target evidence subgraph minus the expected conflict consumption of the target evidence subgraph after verification, i.e., the expected reduction in the C_j value. Second, the execution cost, time delay, and side effects of the candidate verification tool are predicted. The execution cost refers to the computing resources, network overhead, or API call economic cost required to call the tool; the time delay refers to the time span between initiating the call and obtaining the result; the side effects refer to the negative impacts that calling the tool may bring, such as increased system load, additional data interference, triggering unnecessary alarms, or introducing new uncertainties.

[0059] Furthermore, the sum of the execution cost, the time delay, and the side effects is taken as the comprehensive cost, i.e., Comprehensive Cost = Cost_k + Delay_k + SideEffect_k. Then, the utility value of each candidate verification tool is calculated, where the utility value is the verification benefit divided by the comprehensive cost. A higher utility value indicates that the tool achieves a greater conflict resolution effect at a lower cost. This model accurately reflects the input-output ratio of each verification operation, enabling the system to make optimal resource allocation decisions in an economic sense.

[0060] Further, the candidate verification tools in the list of available tools are sorted from highest to lowest according to their utility values, and the candidate verification tools are selected sequentially until the set of selected verification tools covers all the key conflict fields. The selected set of verification tools is determined as the target verification tool, and the target verification tool is invoked to verify the target evidence subgraph.

[0061] This embodiment identifies fields where the difference in confidence between supporting and opposing sides is below a preset threshold as key conflict fields, ensuring that the verification process focuses on the core points of contention where the contradictions are most intense, rather than conducting a general search. By introducing a utility value (Utility = Verification Benefit / Overall Cost) and selecting tools according to their utility values ​​from high to low, the system prioritizes tools with the lowest cost, least latency, fewest side effects, and the greatest expected reduction in conflict. Furthermore, this utility-ranked targeted selection mechanism ensures that each invoked verification operation has a clear return on investment, effectively avoiding the high cost of blindly calling all tools for comprehensive verification and significantly improving verification efficiency in complex task scenarios.

[0062] S5, write the verification result incrementally back to the evidence relationship graph, update the affected relationship edges and the source reliability, and output the verification conclusion containing the established facts, conditional conclusions and unresolved conflicts based on the updated evidence relationship graph.

[0063] In practice, after obtaining the verification result, the verification result is incrementally written back to the evidence relationship graph, updating the affected relationship edges and the source reliability. The incremental write-back refers to modifying only the attributes (such as confidence value and source reliability score) of the nodes and relationship edges directly corresponding to the verification result, without reconstructing the entire evidence relationship graph, thereby significantly reducing the global computational overhead caused by local verification. Furthermore, based on the updated evidence relationship graph, a verification conclusion containing established facts, conditional conclusions, and unresolved conflicts is output. The verification conclusion not only includes the final conclusion text and applicable event anchors, but also explicitly lists the supporting evidence set, retained counterexamples, remaining conflict budget, effective time range, and fields requiring manual confirmation. Through different state labels, the agent can accurately distinguish between verified established facts, conditional conclusions with minor uncertainties, and unresolved conflicts requiring manual intervention, thus providing a clear and hierarchical basis for credibility in the final decision.

[0064] This embodiment constructs cross-modal event anchors to uniformly map evidence materials from different modalities and time-based representation systems onto the same structured framework, thus avoiding false conflict judgments caused by directly comparing asynchronous time information (such as log timestamps and natural language descriptions). By constructing an evidence relationship graph containing supporting edges, rebuttal edges, temporal edges, and common-origin edges, the complex logical adversarial relationships between evidence are accurately represented, laying a structured foundation for subsequent conflict identification and impact assessment. By utilizing a conflict consumption calculation formula, comprehensively considering proposition importance, evidence reliability, impact scope, and decision risk, the system can accurately quantify the destructive power of contradictions, effectively preventing low-importance conflicts from consuming excessive computing resources. Furthermore, by introducing a comparison mechanism between the total conflict budget and a preset uncertainty threshold, targeted verification is ensured only when the contradiction scale exceeds the limit, effectively avoiding the high computational costs caused by blindly calling all tools in traditional solutions. Furthermore, through an incremental write-back mechanism, only affected nodes and edges are updated, maintaining the overall consistency of the evidence graph and significantly reducing subsequent maintenance overhead. The final structured conclusions include supporting evidence, reserved counterexamples, and fields requiring human verification, ensuring the traceability and transparency of the decision-making process.

[0065] In some preferred embodiments, the step of incrementally writing the verification result back to the evidence relationship graph and updating the affected relationship edges and the source reliability includes: obtaining the target evidence material identifier and target proposition identifier corresponding to the verification result; locating the target relationship edge and target source node related to the verification result in the evidence relationship graph based on the target evidence material identifier and target proposition identifier; extracting a new confidence value corresponding to the target relationship edge from the verification result, and extracting an event anomaly elimination indication corresponding to the target source node from the verification result; updating the [relationship graph] based on the new confidence value. The confidence level of the target relation edge is determined; based on the event anomaly elimination indication, multiple reliability evaluation indicators of the target source node are obtained, and the multiple reliability evaluation indicators are weighted, summed, and normalized according to predetermined weights to obtain the updated source reliability of the target source node. The multiple reliability evaluation indicators include the historical accuracy, material integrity score, consistency score, and data anomaly detection result of the target source node; for other relation edges and other source nodes in the evidence relation graph other than the target relation edge and the target source node, their original confidence level and original source reliability remain unchanged.

[0066] In practice, this embodiment effectively solves the computational redundancy problem caused by the need to recalculate the entire evidence graph after verification through a partial update mechanism, ensuring that the maintenance of the evidence graph is always in the most efficient state. The specific implementation method is as follows: After obtaining the verification result, the target evidence material identifier and target proposition identifier corresponding to the verification result are first acquired. The target evidence material identifier is an ID or path used to uniquely identify the evidence material being verified; the target proposition identifier refers to the ID of the specific proposition to be verified involved in the verification. These two identifiers determine the direct scope of influence of the verification result.

[0067] Furthermore, in the evidence relationship graph, based on the target evidence material identifier and the target proposition identifier, target relationship edges and target source nodes related to the verification result are located. The target relationship edge refers to the supporting or rebuttal edge that directly connects the target evidence material and the target proposition, and the change in its weight directly determines the structural change of the evidence graph; the target source node refers to the source entity that generates the target evidence material, and the change in its reliability will affect the weight of all subsequent outputs from that source.

[0068] Further, a new confidence value corresponding to the target relation edge is extracted from the verification result, and an event anomaly elimination indication corresponding to the target source node is extracted from the verification result. The new confidence value is a new evaluation of the relation edge after verification, typically within the range [0,1]; the event anomaly elimination indication indicates whether the current data anomaly has been eliminated by the verification result. For example, if the verification finds that the original data anomaly was caused by equipment failure and the equipment has been repaired, the anomaly elimination indication is true, meaning that the bad records from that source should be erased or mitigated.

[0069] Furthermore, based on the new confidence value, the confidence of the target relation edge is updated, that is, the old confidence on the relation edge is directly replaced with the new confidence obtained after verification, so that the local dependency relationship is updated in real time.

[0070] Further, based on the event anomaly elimination indication, multiple reliability assessment indicators of the target source node are obtained, and these indicators are weighted, summed, and normalized according to predetermined weights to obtain the updated source reliability of the target source node. The multiple reliability assessment indicators include the historical accuracy, material integrity score, consistency score, and data anomaly detection result of the target source node. Since the current source anomaly may have been eliminated by the verification result, the updated source reliability should be greater than the source reliability before the event anomaly elimination indication. Specifically, if the anomaly elimination indication is true, the data anomaly detection result changes from "anomaly exists" to "no anomaly," and the weight of this factor in the weighted summation will no longer be used as a penalty, thereby improving the overall source reliability.

[0071] Furthermore, for all other relationship edges and source nodes in the evidence graph besides the target relationship edge and the target source node, their original confidence and source reliability remain unchanged. This approach ensures that the update operation is local and does not affect the entire evidence graph, thereby greatly reducing the computational burden on the system.

[0072] This embodiment achieves precise local updates to the evidence graph by locating target relationship edges and target source nodes based on the identifiers in the verification results. It replaces the old confidence values ​​of relationship edges with new confidence values ​​provided by the verification results and recalculates the source reliability based on anomaly elimination indicators, ensuring that the evidence graph reflects the latest factual state promptly after each verification. Furthermore, it preserves the original attributes of other irrelevant relationship edges and source nodes, avoiding the enormous computational overhead of recalculating the entire evidence relationship graph due to local verification, significantly improving the efficiency and real-time performance of evidence graph maintenance. This incremental update mechanism ensures that the system can continuously maintain the accuracy and consistency of the evidence graph with extremely low computational cost when processing large-scale multimodal evidence.

[0073] In some preferred embodiments, the output based on the updated evidence graph includes a confirmed fact, a conditional conclusion, and a verification conclusion of unresolved conflicts, comprising: traversing all propositions to be verified in the updated evidence graph; for each proposition to be verified, determining whether the current conflict cost of the proposition to be verified is zero; if the current conflict cost of the proposition to be verified is zero, then marking the output state of the proposition to be verified as a confirmed fact; if the current conflict cost of the proposition to be verified is greater than zero and less than the preset uncertainty threshold, then marking the output state of the proposition to be verified as a conditional conclusion ... If the prior conflict cost is greater than or equal to the preset uncertainty threshold, the output state of the proposition to be verified is marked as an unresolved conflict; the edge weights corresponding to the supporting edges and the rebuttal edges in the updated evidence graph are extracted. If the edge weight corresponding to the supporting edge is greater than the edge weight corresponding to the rebuttal edge, the evidence material corresponding to the supporting edge is taken as the supporting evidence set; if the edge weight corresponding to the rebuttal edge is greater than the edge weight corresponding to the supporting edge, the evidence material corresponding to the rebuttal edge is taken as a reserved counterexample; based on the supporting evidence set, the reserved counterexamples, and the output state of each proposition to be verified, the verification conclusion is generated.

[0074] In specific implementation, this embodiment traverses and classifies the updated evidence relationship graph, clearly distinguishing the propositions to be verified into three states: certain facts, conditional conclusions, and unresolved conflicts. This addresses the problem of vague conclusion outputs and the inability to differentiate credibility levels, enabling downstream tasks to make different decision responses based on the credibility of the conclusions. The specific implementation method is as follows: The verification conclusion is output based on the updated evidence relationship graph. First, all propositions to be verified in the updated evidence relationship graph are traversed. For each proposition to be verified, it is determined whether the current conflict cost of the proposition to be verified is zero.

[0075] Furthermore, if the current conflict cost of the proposition to be verified is zero, it indicates that there is no contradictory evidence for the proposition, or that opposing evidence has been completely verified and eliminated. Therefore, the output state of the proposition to be verified is marked as a confirmed fact. A confirmed fact is the strongest of the three types of states and can be directly used as the basis for the final decision, such as directly triggering a security alarm or automatically generating an incident report.

[0076] Furthermore, if the current conflict consumption of the proposition to be verified is greater than zero and less than the preset uncertainty threshold, it indicates that although the proposition has some contradiction, the contradiction intensity is low and insufficient to affect the overall judgment. Therefore, the output state of the proposition to be verified is marked as a conditional conclusion. A conditional conclusion means that the conclusion is valid under certain conditions, but there are still some minor uncertainties that need attention, which can be used by decision-makers as a risk warning or suggestion for adoption.

[0077] Furthermore, if the current conflict consumption of the proposition to be verified is greater than or equal to the preset uncertainty threshold, it indicates that the contradiction intensity of the proposition is still high, exceeding the system's tolerance range. Therefore, the output state of the proposition to be verified is marked as unresolved conflict. Unresolved conflict means that the system cannot make a reliable judgment based on the current knowledge, requiring manual intervention or further investigation. In this case, the system should clearly mark the fields that need attention for review.

[0078] Furthermore, the edge weights corresponding to supporting edges and rebuttal edges are extracted from the updated evidence graph. If the edge weight corresponding to a supporting edge is greater than the edge weight corresponding to a rebuttal edge, the evidence material corresponding to the supporting edge is included in the supporting evidence set; if the edge weight corresponding to a rebuttal edge is greater than the edge weight corresponding to a supporting edge, the evidence material corresponding to the rebuttal edge is included as a reserved counterexample. The supporting evidence set and the reserved counterexamples together constitute the complete picture of the existing evidence for this proposition.

[0079] Furthermore, finally, the supporting evidence set, the retained counterexamples, the output status of each of the propositions to be verified, the conclusion text, the applicable event anchor, the remaining conflict budget, the effective time range, and the fields requiring manual confirmation are encapsulated into the verification conclusion. The conclusion text is a natural language description generated by the system based on the propositions to be verified that are marked as certain facts; the applicable event anchor refers to the structured event anchor object corresponding to the conclusion text, used to limit the scope of the subject, object, time, and location to which the conclusion applies; the remaining conflict budget represents the overall level of uncertainty that still exists after the incremental verification result is written back to the evidence relationship diagram; the effective time range is the boundary of the conclusion's validity determined based on the verified and calibrated time field information in the supporting evidence, used to clarify the applicable period of the conclusion in the time dimension; the fields requiring manual confirmation are a list of key data items that still have ambiguity in the original evidence or have not completely eliminated conflicts between multiple sources of data after verification, resulting in low confidence. The system categorizes each proposition to be verified by traversing the updated evidence relationship graph, extracts core statements based on established facts to generate conclusion text, retrieves relevant anchor points from the evidence graph to lock applicable event anchor points, and deducts the conflict consumption already resolved in this verification from the global total conflict budget to calculate the remaining conflict budget. At the same time, it determines the effective time range based on the calibrated timestamp and time interval, and then filters out fields that still have confidence defects to be reviewed to form fields that need to be manually confirmed. Finally, the aforementioned elements, along with the supporting evidence set, retained counterexamples, and the output status of each proposition, are encapsulated together to form the verification conclusion.

[0080] The verification conclusion is a structured output object containing explicit metadata indications, which can be directly output to users or downstream application systems for in-depth analysis.

[0081] This embodiment classifies the current conflict consumption of each proposition to be verified in the updated evidence relationship graph into three levels, enabling the system to intelligently distinguish between three states: established facts, conditional conclusions, and unresolved conflicts. Furthermore, established facts can be directly adopted by the system, conditional conclusions provide effective boundaries and applicable conditions for decision-making, and unresolved conflicts provide clear entry points and priorities for manual review. Furthermore, by extracting supporting edge weights and rebuttal edge weights, the system can clearly display supporting evidence and existing counterexamples for the conclusion, greatly enhancing the interpretability and traceability of the conclusion. Furthermore, the final encapsulated structured verification conclusion includes the remaining conflict budget, effective time range, and fields requiring manual confirmation, allowing downstream systems or operators to clearly understand the scope and limitations of the conclusion, effectively improving the transparency and reliability of decision-making.

[0082] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0083] In one embodiment, a multimodal evidence fusion and targeted verification device is provided, which corresponds one-to-one with the multimodal evidence fusion and targeted verification methods described in the above embodiments. For example... Figure 3 As shown, the multimodal evidence fusion and targeted verification device includes: A mapping unit is used to map the entities, time, and actions in each piece of evidence to an event anchor point in response to the acquisition of multimodal evidence materials, thereby obtaining an evidence set associated with the event anchor point; The construction unit is used to construct an evidence relationship graph based on the evidence set, which consists of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and the source reliability associated with the relationship edges is obtained. The calculation unit is used to calculate the conflict cost for the supporting edge and the rebuttal edge of the same proposition to be verified, based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and to sum up the conflict costs to obtain the total conflict budget. The verification unit is used to locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph when the total conflict budget exceeds a preset uncertainty threshold, and to verify the target evidence subgraph to obtain the verification result. The update unit is used to write the verification result incrementally back to the evidence relationship graph, update the affected relationship edges and the source reliability, and output a verification conclusion containing established facts, conditional conclusions and unresolved conflicts based on the updated evidence relationship graph.

[0084] In some preferred embodiments, the event anchor includes a subject field, an object field, an action field, a time range field, a location field, and a status version field. Mapping the entities, times, and actions in each piece of evidence to the event anchor includes: The evidence materials are subjected to structured analysis to extract structured metadata carried in the evidence materials. The structured metadata includes device identifier, collection timestamp, and geographic location coordinates. Semantic recognition is performed on the content of the evidence material to extract unstructured entities recorded in the evidence material, including descriptions of people, objects, and actions; Map the collection timestamp in the structured metadata to the time range field, map the geographic location coordinates to the location field, map the device identifier and the person in the unstructured entity to the subject field, map the object in the unstructured entity to the object field, and map the action description in the unstructured entity to the action field. If the structured metadata and the unstructured entity have a semantic conflict in the subject field, the time range field, or the location field, then the corresponding field value is set as a candidate set, which includes the corresponding value of the structured metadata and the corresponding value of the unstructured entity; if the structured metadata and the unstructured entity do not have a semantic conflict in the subject field, the time range field, and the location field, then the corresponding value of the structured metadata is mapped to the corresponding field in the event anchor. If the collection timestamp in the structured metadata conflicts with the time point pointed to by the action description in the unstructured entity, the collection timestamp is retained as a time interval, which includes the time point pointed to by the action description; if the collection timestamp in the structured metadata does not conflict with the time point pointed to by the action description in the unstructured entity, the collection timestamp is mapped to the time range field.

[0085] In some preferred embodiments, the apparatus further includes a deduplication unit, the deduplication unit being used for: Obtain the source file path, acquisition device serial number, and hash value of the content summary for each of the aforementioned evidence materials; The source file paths of any two pieces of evidence are compared sequentially. If the source file paths of the two pieces of evidence are the same, it is determined that the two pieces of evidence have a direct common source relationship. If the source file paths of the two pieces of evidence are different, then compare whether the serial numbers of the acquisition devices of the two pieces of evidence are the same, and compare whether the hash values ​​of the content digests of the two pieces of evidence are the same; If the serial numbers of the acquisition devices in the two pieces of evidence are the same and the hash values ​​of the content digests in the two pieces of evidence are the same, then it is determined that the two pieces of evidence have an indirect homology relationship. If the serial numbers of the acquisition devices in the two pieces of evidence are different or the hash values ​​of the content digests in the two pieces of evidence are different, then it is determined that the two pieces of evidence do not have a common source relationship. Based on the direct or indirect homology relationship, a source chain identifier is constructed for multiple pieces of evidence with a homology relationship, and multiple pieces of evidence with the same source chain identifier are merged into the same evidence group. When the evidence group is associated with the event anchor, only one contribution weight is retained, and the contribution weight is equal to the normalized value of the weight of each piece of evidence in the evidence group.

[0086] In some preferred embodiments, obtaining the source reliability associated with the relation edge includes: Obtain the historical accuracy rate of each source in the historical verification task, where the historical accuracy rate is the proportion of evidence output by the source that is verified as true by subsequent verifications. For each source, based on the completeness of the fields in the evidence material and the signal quality corresponding to the modality type of the evidence material, a material integrity score for the evidence material corresponding to the source is obtained; For each source, independent evidence that is related to the evidence material corresponding to the source is retrieved from the evidence relationship graph. The consistency between the evidence material of the current source and the independent evidence on each field of the event anchor is compared. A consistency score is obtained based on the consistency. The independent evidence comes from other sources different from the current source and has a temporal or causal relationship with the evidence material of the current source. Detect whether there are data anomalies in the evidence materials associated with the source in the current event. The data anomalies include logical conflicts in the collection timestamps of the evidence materials, failure to verify the integrity check value of the evidence materials, or values ​​in the evidence materials exceeding a preset normal range. The historical accuracy, material integrity score, consistency score, and data anomaly detection results are weighted and summed based on predetermined weights, and the weighted summation result is normalized to obtain the source reliability.

[0087] In some preferred embodiments, calculating conflict consumption based on proposition importance, the reliability of contradictory evidence, the scope of conflict coverage, and the risk of conclusion decision includes: Obtain the proposition importance, conflict coverage, and conclusion decision risk corresponding to the same proposition to be verified; Determine the first reliability corresponding to the supporting edge and the second reliability corresponding to the rebuttal edge for the same proposition to be verified, respectively; Calculate the product of the first reliability and the second reliability, and sum the products of all the supporting edges and the rebuttal edges of the same proposition to be verified to obtain the total reliability product. The conflict consumption of the same proposition to be verified is obtained by multiplying the sum of the products of the proposition importance, the reliability, the conflict coverage, and the decision risk of the conclusion.

[0088] In some preferred embodiments, verifying the target evidence subgraph includes: Key conflict fields are determined from the target evidence subgraph, where the key conflict field is the field whose confidence difference between the supporting edge and the rebuttal edge is less than a preset difference threshold; Obtain the range of verification fields supported by each candidate verification tool, and filter out the candidate verification tools whose range of verification fields covers the key conflicting fields to form a list of candidate tools; For each candidate verification tool in the list of candidate tools, predict the amount of conflict consumption that the candidate verification tool can reduce for the target evidence subgraph, as the verification benefit, and predict the execution cost, time delay and side effects of the candidate verification tool, and take the sum of the execution cost, the time delay and the side effects as the comprehensive cost. Calculate the utility value for each of the candidate verification tools, where the utility value is the verification benefit divided by the overall cost; The candidate verification tools in the list of available tools are sorted from high to low according to their utility values, and the candidate verification tools are selected in turn until the set of selected verification tools covers all the key conflict fields. The set of selected verification tools is then determined as the target verification tool, and the target verification tool is invoked to verify the target evidence subgraph.

[0089] In some preferred embodiments, the step of incrementally writing the verification result back to the evidence relationship graph and updating the affected relationship edges and the source reliability includes: Obtain the target evidence material identifier and target proposition identifier corresponding to the verification result; In the evidence relationship graph, based on the target evidence material identifier and the target proposition identifier, the target relationship edge and target source node related to the verification result are located; Extract a new confidence value corresponding to the target relation edge from the verification result, and extract an event anomaly elimination indication corresponding to the target source node from the verification result; Based on the new confidence value, update the confidence of the target relation edge; Based on the event anomaly elimination indication, multiple reliability evaluation indicators of the target source node are obtained, and the multiple reliability evaluation indicators are weighted, summed and normalized based on predetermined weights to obtain the updated source reliability of the target source node. The multiple reliability evaluation indicators include the historical accuracy, material integrity score, consistency score and data anomaly detection result of the target source node. For other relationship edges and other source nodes in the evidence relationship graph besides the target relationship edge and the target source node, their original confidence and original source reliability remain unchanged.

[0090] In some preferred embodiments, the output based on the updated evidence relationship graph includes verification conclusions of established facts, conditional conclusions, and unresolved conflicts, including: Traverse all the propositions to be verified in the updated evidence relationship graph. For each proposition to be verified, determine whether the current conflict cost of the proposition to be verified is zero. If the current conflict cost of the proposition to be verified is zero, then the output state of the proposition to be verified is marked as a confirmed fact. If the current conflict cost of the proposition to be verified is greater than zero and less than the preset uncertainty threshold, then the output state of the proposition to be verified is marked as a conditional conclusion. If the current conflict consumption of the proposition to be verified is greater than or equal to the preset uncertainty threshold, then the output state of the proposition to be verified is marked as unresolved conflict. Extract the edge weights corresponding to the supporting edges and the rebuttal edges in the updated evidence relationship graph. If the edge weight corresponding to the supporting edge is greater than the edge weight corresponding to the rebuttal edge, then the evidence material corresponding to the supporting edge is taken as the supporting evidence set; if the edge weight corresponding to the rebuttal edge is greater than the edge weight corresponding to the supporting edge, then the evidence material corresponding to the rebuttal edge is taken as a reserved counterexample. The verification conclusion is generated based on the set of supporting evidence, the retained counterexamples, and the output state of each of the propositions to be verified.

[0091] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of a multimodal evidence fusion and targeted verification method on the server side.

[0092] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational 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 in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of a multimodal evidence fusion and targeted verification method.

[0093] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a multimodal evidence fusion and targeted verification method provided in any of the above method embodiments.

[0094] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a multimodal evidence fusion and targeted verification method provided in any of the above method embodiments.

[0095] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0096] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0097] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0098] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0099] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for multimodal evidence fusion and targeted verification, characterized in that, include: In response to obtaining multimodal evidence materials, the entities, time, and actions in each piece of evidence material are mapped to event anchor points to obtain an evidence set associated with the event anchor points; Based on the evidence set, an evidence relationship graph is constructed, consisting of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and the source reliability associated with the relationship edges is obtained. For the supporting edge and the rebuttal edge of the same proposition to be verified, the conflict cost is calculated based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and the total conflict budget is obtained by summing up the conflict costs of each edge. When the total conflict budget exceeds a preset uncertainty threshold, locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph, verify the target evidence subgraph, and obtain the verification result. The verification result is incrementally written back to the evidence relationship graph, the affected relationship edges and the source reliability are updated, and the verification conclusion containing the established facts, conditional conclusions and unresolved conflicts is output based on the updated evidence relationship graph.

2. The multimodal evidence fusion and targeted verification method according to claim 1, characterized in that, The event anchor point includes a subject field, an object field, an action field, a time range field, a location field, and a status version field. Mapping the entities, times, and actions in each piece of evidence to the event anchor point includes: The evidence materials are subjected to structured analysis to extract structured metadata carried in the evidence materials. The structured metadata includes device identifier, collection timestamp, and geographic location coordinates. Semantic recognition is performed on the content of the evidence material to extract unstructured entities recorded in the evidence material, including descriptions of people, objects, and actions; Map the collection timestamp in the structured metadata to the time range field, map the geographic location coordinates to the location field, map the device identifier and the person in the unstructured entity to the subject field, map the object in the unstructured entity to the object field, and map the action description in the unstructured entity to the action field. If the structured metadata and the unstructured entity have a semantic conflict in the subject field, the time range field, or the location field, then the corresponding field value is set as a candidate set, which includes the corresponding value of the structured metadata and the corresponding value of the unstructured entity; if the structured metadata and the unstructured entity do not have a semantic conflict in the subject field, the time range field, and the location field, then the corresponding value of the structured metadata is mapped to the corresponding field in the event anchor. If the collection timestamp in the structured metadata conflicts with the time point pointed to by the action description in the unstructured entity, the collection timestamp is retained as a time interval, which includes the time point pointed to by the action description; if the collection timestamp in the structured metadata does not conflict with the time point pointed to by the action description in the unstructured entity, the collection timestamp is mapped to the time range field.

3. The multimodal evidence fusion and targeted verification method according to claim 2, characterized in that, After obtaining the set of evidence associated with the event anchor, the method further includes: Obtain the source file path, acquisition device serial number, and hash value of the content summary for each of the aforementioned evidence materials; The source file paths of any two pieces of evidence are compared sequentially. If the source file paths of the two pieces of evidence are the same, it is determined that the two pieces of evidence have a direct common source relationship. If the source file paths of the two pieces of evidence are different, then compare whether the serial numbers of the acquisition devices of the two pieces of evidence are the same, and compare whether the hash values ​​of the content digests of the two pieces of evidence are the same; If the serial numbers of the acquisition devices in the two pieces of evidence are the same and the hash values ​​of the content digests in the two pieces of evidence are the same, then it is determined that the two pieces of evidence have an indirect homology relationship. If the serial numbers of the acquisition devices in the two pieces of evidence are different or the hash values ​​of the content digests in the two pieces of evidence are different, then it is determined that the two pieces of evidence do not have a common source relationship. Based on the direct or indirect homology relationship, a source chain identifier is constructed for multiple pieces of evidence with a homology relationship, and multiple pieces of evidence with the same source chain identifier are merged into the same evidence group. When the evidence group is associated with the event anchor, only one contribution weight is retained, and the contribution weight is equal to the normalized value of the weight of each piece of evidence in the evidence group.

4. The multimodal evidence fusion and targeted verification method according to claim 1, characterized in that, The step of obtaining the source reliability associated with the relation edge includes: Obtain the historical accuracy rate of each source in the historical verification task, where the historical accuracy rate is the proportion of evidence output by the source that is verified as true by subsequent verifications. For each source, based on the completeness of the fields in the evidence material and the signal quality corresponding to the modality type of the evidence material, a material integrity score for the evidence material corresponding to the source is obtained; For each source, independent evidence that is related to the evidence material corresponding to the source is retrieved from the evidence relationship graph. The consistency between the evidence material of the current source and the independent evidence on each field of the event anchor is compared. A consistency score is obtained based on the consistency. The independent evidence comes from other sources different from the current source and has a temporal or causal relationship with the evidence material of the current source. Detect whether there are data anomalies in the evidence materials associated with the source in the current event. The data anomalies include logical conflicts in the collection timestamps of the evidence materials, failure to verify the integrity check value of the evidence materials, or values ​​in the evidence materials exceeding a preset normal range. The historical accuracy, material integrity score, consistency score, and data anomaly detection results are weighted and summed based on predetermined weights, and the weighted summation result is normalized to obtain the source reliability.

5. The multimodal evidence fusion and targeted verification method according to claim 1, characterized in that, The calculation of conflict consumption, based on proposition importance, reliability of contradictory evidence, scope of conflict coverage, and decision-making risk, includes: Obtain the proposition importance, conflict coverage, and conclusion decision risk corresponding to the same proposition to be verified; Determine the first reliability corresponding to the supporting edge and the second reliability corresponding to the rebuttal edge for the same proposition to be verified, respectively; Calculate the product of the first reliability and the second reliability, and sum the products of all the supporting edges and the rebuttal edges of the same proposition to be verified to obtain the total reliability product. The conflict consumption of the same proposition to be verified is obtained by multiplying the sum of the products of the proposition importance, the reliability, the conflict coverage, and the decision risk of the conclusion.

6. The multimodal evidence fusion and targeted verification method according to claim 1, characterized in that, The verification of the target evidence subgraph includes: Key conflict fields are determined from the target evidence subgraph, where the key conflict field is the field whose confidence difference between the supporting edge and the rebuttal edge is less than a preset difference threshold; Obtain the range of verification fields supported by each candidate verification tool, and filter out the candidate verification tools whose range of verification fields covers the key conflicting fields to form a list of candidate tools; For each candidate verification tool in the list of candidate tools, predict the amount of conflict consumption that the candidate verification tool can reduce for the target evidence subgraph, as the verification benefit, and predict the execution cost, time delay and side effects of the candidate verification tool, and take the sum of the execution cost, the time delay and the side effects as the comprehensive cost. Calculate the utility value for each of the candidate verification tools, where the utility value is the verification benefit divided by the overall cost; The candidate verification tools in the list of available tools are sorted from high to low according to their utility values, and the candidate verification tools are selected in turn until the set of selected verification tools covers all the key conflict fields. The set of selected verification tools is then determined as the target verification tool, and the target verification tool is invoked to verify the target evidence subgraph.

7. The multimodal evidence fusion and targeted verification method according to claim 1, characterized in that, The step of incrementally writing the verification result back to the evidence relationship graph and updating the affected relationship edges and the source reliability includes: Obtain the target evidence material identifier and target proposition identifier corresponding to the verification result; In the evidence relationship graph, based on the target evidence material identifier and the target proposition identifier, the target relationship edge and target source node related to the verification result are located; Extract a new confidence value corresponding to the target relation edge from the verification result, and extract an event anomaly elimination indication corresponding to the target source node from the verification result; Based on the new confidence value, update the confidence of the target relation edge; Based on the event anomaly elimination indication, multiple reliability evaluation indicators of the target source node are obtained, and the multiple reliability evaluation indicators are weighted, summed and normalized based on predetermined weights to obtain the updated source reliability of the target source node. The multiple reliability evaluation indicators include the historical accuracy, material integrity score, consistency score and data anomaly detection result of the target source node. For other relationship edges and other source nodes in the evidence relationship graph besides the target relationship edge and the target source node, their original confidence and original source reliability remain unchanged; And / or, the output based on the updated evidence relationship graph includes verification conclusions that establish facts, conditional conclusions, and unresolved conflicts, including: Traverse all the propositions to be verified in the updated evidence relationship graph. For each proposition to be verified, determine whether the current conflict cost of the proposition to be verified is zero. If the current conflict cost of the proposition to be verified is zero, then the output state of the proposition to be verified is marked as a confirmed fact. If the current conflict cost of the proposition to be verified is greater than zero and less than the preset uncertainty threshold, then the output state of the proposition to be verified is marked as a conditional conclusion. If the current conflict consumption of the proposition to be verified is greater than or equal to the preset uncertainty threshold, then the output state of the proposition to be verified is marked as unresolved conflict. Extract the edge weights corresponding to the supporting edges and the rebuttal edges in the updated evidence relationship graph. If the edge weight corresponding to the supporting edge is greater than the edge weight corresponding to the rebuttal edge, then the evidence material corresponding to the supporting edge is taken as the supporting evidence set; if the edge weight corresponding to the rebuttal edge is greater than the edge weight corresponding to the supporting edge, then the evidence material corresponding to the rebuttal edge is taken as a reserved counterexample. The verification conclusion is generated based on the set of supporting evidence, the retained counterexamples, and the output state of each of the propositions to be verified.

8. A multimodal evidence fusion and targeted verification device, characterized in that, include: A mapping unit is used to map the entities, time, and actions in each piece of evidence to an event anchor point in response to the acquisition of multimodal evidence materials, thereby obtaining an evidence set associated with the event anchor point; The construction unit is used to construct an evidence relationship graph based on the evidence set, which consists of event anchors, evidence materials, propositions to be verified, and verification results. The relationship edges of the evidence relationship graph include at least supporting edges and rebuttal edges, and the source reliability associated with the relationship edges is obtained. The calculation unit is used to calculate the conflict cost for the supporting edge and the rebuttal edge of the same proposition to be verified, based on the importance of the proposition, the reliability of the contradictory evidence, the scope of the conflict coverage and the risk of the conclusion decision, and to sum up the conflict costs to obtain the total conflict budget. The verification unit is used to locate the target proposition that contributes the most to the total conflict budget and its corresponding target evidence subgraph when the total conflict budget exceeds a preset uncertainty threshold, and to verify the target evidence subgraph to obtain the verification result. The update unit is used to write the verification result incrementally back to the evidence relationship graph, update the affected relationship edges and the source reliability, and output a verification conclusion containing established facts, conditional conclusions and unresolved conflicts based on the updated evidence relationship graph.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multimodal evidence fusion and targeted verification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the multimodal evidence fusion and targeted verification method as described in any one of claims 1 to 7.

Images