A method and system for matching and verifying legal constitutive elements sub-graph based on multi-modal causal knowledge graph

Abstract

This invention discloses a method and system for matching and verifying subgraphs of legal constituent elements based on a multimodal causal knowledge graph, belonging to the field of data processing. The method includes: acquiring multi-source evidence data of a case to be verified and extracting multiple event units; establishing causal edges pointing from the preceding event unit to the following event unit when two event units satisfy temporal preconditions and spatiotemporal reachability conditions, thus obtaining a fact-causal graph; determining the number of mutually independent supporting sources and accordingly determining the corroboration weight of the causal edges; determining the slot alignment degree between the normative slots of constituent nodes in the constituent template and the semantic slots of the corresponding fact subgraphs in the fact-causal graph, thus obtaining candidate constituent element subgraphs; matching the candidate constituent element subgraphs with the target constituent element subgraph, and outputting verification conclusions based on the matching results. This invention can provide auditable conclusions line by line, avoiding the direct substitution of subjective perceptions with objectively occurring results.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method and system for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs. Background Technology

[0002] Existing multimodal evidence element extraction and element matching methods based on case knowledge graphs typically first perform named entity recognition and relation extraction on text materials, and object detection, speech transcription, and image annotation on audio, video, and image materials respectively, uniformly transcribing all types of materials into triples composed of entities and relations. Then, all triples are merged into a single-layer flat knowledge graph, where nodes mix real-life factual entities and normative concepts, and edges are uniformly denoted as broad relations such as association, causation, and inclusion. Next, the constituent elements are pre-encoded into several target subgraph templates, and the parts that match the target subgraph templates are found in the graph through subgraph isomorphism retrieval or graph embedding vector similarity retrieval. Finally, each retrieved matching item is weighted and scored according to preset weights, and if the score exceeds a unified threshold, the element is determined to be valid and a hit list is output.

[0003] However, existing methods have the following problems: 1) Real-life facts and normative concepts are placed in the same topological layer, separated only by a criterionless association edge. The mapping relationship is determined solely by textual surface similarity, leading to the same material being mapped to different normative concepts depending on the extracted wording; 2) Causal judgments on the graph degenerate into transitive closure calculations of whether directed paths are connected. As long as a connected path exists, a causal relationship is determined to be valid, while the effectiveness allocation of exogenous intervention events along the path remains completely invisible; 3) Flat weighted summation is a continuous, monotonous, and compensable scalar operation. Even if the veto-restraining element takes the most unfavorable value, it still... The weight of the item is subtracted from the total score and can be compensated by the high scores of the other items. The hard prohibition is smoothed out by the continuous probability. 4) Subjective elements that cannot be directly observed can only rely on self-acknowledged keywords or overall scores. Without self-acknowledgement, it is impossible to judge. When the objective result deviates from the subject's cognition, the former will be directly equated with the latter. 5) Directed edges take effect as soon as they are extracted. A single source can support key causal edges. Contradictory materials coexist in the graph and are scored separately. The scoring is only for a single target subgraph and never calculates the explanatory power of competing explanations in parallel. There is no correspondence between the output score and the statutory proof standard.

[0004] Therefore, existing technologies suffer from problems such as the inability to recalculate and trace the correspondence between multimodal evidence data and legal elements, and the inconsistencies between the conclusions output and the hierarchical logic and proof standards of the normative system when there are exogenous intervention events, veto events, single-source evidence, and mutually incompatible evidence. Summary of the Invention

[0005] One of the objectives of this invention is to provide a method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs, in order to solve the problems in the prior art where the correspondence between multimodal evidence data and legal constituent elements is not recalculated or traceable, and the output conclusions are inconsistent with the hierarchical logic and proof standards of the normative system.

[0006] This invention is achieved through the following technical solution: a method for matching and verifying subgraphs of legal constituent elements based on a multimodal causal knowledge graph, comprising the following steps: acquiring multi-source evidence data of a case to be verified; extracting multiple event units from the multi-source evidence data, wherein each event unit includes time information, spatial information, and semantic slots; when two event units satisfy temporal preconditions and spatiotemporal reachability conditions, establishing a causal edge pointing from the preceding event unit to the following event unit, thereby obtaining a factual causal graph; determining mutually independent supports based on whether the sources supporting the causal edge are of the same origin in the evidence derivation relationship. The number of sources is determined, and the corroboration weight of the causal edge is determined accordingly. The corroboration weight of a causal edge with fewer than two mutually independent supporting sources is zero. The slot alignment between the canonical slot of the component node in the component template and the semantic slot of the corresponding fact subgraph in the fact causal graph is determined. The projection value of the component node is determined based on the product of the slot alignment and the corroboration weight of the causal edge in the fact subgraph. The component node is activated when the projection value is not less than the activation threshold to obtain a candidate component subgraph. The candidate component subgraph is matched with the target component subgraph, and the verification conclusion is output based on the matching result.

[0007] Further, matching the candidate requirement subgraph with the target requirement subgraph includes: matching sequentially according to the order of appropriateness level, illegality level, and responsibility level; in the appropriateness level, weighted summing and normalizing the projection values ​​of the activated requirement nodes corresponding to the requirement nodes of the target requirement subgraph in the candidate requirement subgraph with normative weights, and setting them to zero when any of the indispensable requirement nodes in the target requirement subgraph is not activated, to obtain the appropriateness matching value; in the illegality level and the responsibility level, respectively determining the exclusion gate value of each illegality exclusion reason and each responsibility exclusion reason, wherein the exclusion gate value is zero or one; multiplying the appropriateness matching value by the difference obtained by subtracting each exclusion gate value from one, respectively, to obtain the level matching value.

[0008] Further, determining the exclusion threshold value for each unlawful exclusion reason includes: locating the event unit corresponding to the unlawful infringement in the candidate element subgraph, setting its infringement intensity to zero and cutting off all its incoming edges to obtain a counterfactual truncated graph; recalculating the amount of legal interest damage caused by the event unit corresponding to the defender on the counterfactual truncated graph according to topological order; using the ratio of the difference between the original amount of legal interest damage and the recalculated amount of legal interest damage to the original amount of legal interest damage as the limit contribution rate; when the limit contribution rate is not less than the limit tolerance, determining the exclusion threshold value of the unlawful exclusion reason to be one; otherwise, determining the exclusion threshold value to be zero and transitioning to the excessive defense branch.

[0009] Furthermore, the determination of the exclusion gate value for each reason for exclusion of responsibility includes at least one of the following: when the projection value of the element node corresponding to the assessment opinion on age and mental state at the time of the act is not less than the activation threshold, the exclusion gate value corresponding to the lack of criminal capacity is determined to be one; when all causal edges representing the actor's accessibility to normative information are missing and the actor has fulfilled the obligation of reasonable inquiry, the exclusion gate value corresponding to the unavoidable error of illegality is determined to be one; when there is no reachable event unit pointing to a legal alternative behavior in the factual causal graph, the exclusion gate value corresponding to the lack of expected possibility is determined to be one.

[0010] Furthermore, during the sequential matching process, when the matching result of any level is zero, the matching of subsequent levels is terminated, and the identifier of that level, the name of the criterion that caused it to be zero, the measured value of the criterion, and the threshold are pushed onto the tracing stack; the top content of the tracing stack is output as a rejection reason.

[0011] Furthermore, the multi-source evidence data includes evidence materials in at least two modalities from video, audio, images, text, physical evidence, and electronic data; the time information is a timestamp interval defined by the earliest start point and the latest end point; the spatial information is spatial location coordinates in a unified spatial coordinate system; the semantic slots include action predicate slots, agent slots, and recipient slots whose values ​​are taken from a factual vocabulary list; the event unit also includes a modal identifier and modal reliability, the modal reliability being determined based on the statutory probative value level of the evidence type and the compliance of the evidence collection procedure; in the event unit, the event unit corresponding to the act performed by the actor is an action node, and the event unit corresponding to the result of the infringement of legal interests is a result node; the normative slots include: normative predicate slots, normative subject slots, and normative object slots; determining the projection value of the element node includes: projecting the normative predicate into the normative object slot; The predicate slot, the normative subject slot, and the normative object slot correspond to the action predicate slot, the agent slot, and the recipient slot of the fact subgraph, respectively. For each pair of corresponding slots, it is determined whether the slot value of the fact subgraph falls into the synonym domain of the normative slot according to a preset synonym domain table. The synonym domain is a collection of factual phrases that are sufficient to serve as subordinate facts of the normative slot in terms of norms. The ratio of the number of slot pairs falling into the synonym domain to the total number of normative slots of the element node is used as the slot alignment degree. The corroboration weights of each causal edge in the fact subgraph are weighted and summed using the normative correlation coefficient of the element node as the weight, and normalized using the sum of the normative correlation coefficients to obtain the evidence corroboration item. The product of the slot alignment degree and the evidence corroboration item is used as the projection value. The activation threshold is set according to the type of element node.

[0012] Furthermore, the synonym domain table is organized using standardized slots as keys and factual phrase aggregations as values. The content of the synonym domain table is determined based on at least one of legal provisions, judicial interpretations, and guiding cases and is updated according to version. Each time the projection value is determined, the version number of the synonym domain table used is recorded. Before obtaining the factual causal graph, the method further includes: for any two event units, determining the intersection-union ratio of their timestamp intervals and the Euclidean distance between their spatial coordinates, and judging whether the action predicate slots, agent slots, and recipient slots of the two events fall into the same synonym domain according to the synonym domain table; when the intersection-union ratio is not lower than the temporal identity threshold, the Euclidean distance does not exceed the spatial identity distance threshold, and all three slots fall into the same synonym domain, the two events are merged into one event unit. The merged event unit takes the slot value and spatial coordinates of the one with higher modal reliability, takes the intersection of their timestamp intervals, takes the union of their modal identifiers, and merges the corresponding evidence materials.

[0013] Further, establishing the causal edge from the preceding event unit to the following event unit includes: designating the two event units as the first event unit and the second event unit, respectively; when the earliest start point of the timestamp interval of the first event unit is not later than the earliest start point of the timestamp interval of the second event unit, and the ratio of the overlap duration of their timestamp intervals to the length of the shorter interval does not exceed the overlap tolerance, it is determined that the timing precondition is met, and the first event unit is the preceding event unit; the duration between the earliest start point of the first event unit and the latest end point of the second event unit is used as the available time window. When the available time window is negative, it is set to zero. The maximum displacement rate of the scene is the sum of the displacement that can be covered within the available time window and the position error tolerance as the displacement budget. When the distance between the spatial coordinates of the two does not exceed the displacement budget, it is determined that the spatiotemporal reachability condition is met. The maximum displacement rate of the scene is determined according to the scene type of the case scene. The method also includes: performing a depth-first search on the fact-cause graph. For each loop detected, the causal edge with the smallest temporal pre-margin in the loop is deleted until the fact-cause graph does not contain loops. The temporal pre-margin is the difference between the earliest starting point of the endpoint and the starting point of the causal edge.

[0014] Further, determining the number of mutually independent supporting sources includes: constructing an evidence source graph for the case to be verified, wherein the evidence source graph uses evidence carrier nodes, evidence collection behavior nodes, and statement subject nodes as nodes, and the evidence derivation relationship as directed edges; for any two supporting sources supporting the same causal edge, tracing back along the evidence derivation relationship to obtain their source node sets in the evidence source graph; when a statement subject node or evidence collection behavior node exists in the intersection of the two source node sets, it is determined that the two have a common pollution source node, and the two are merged into one mutually independent supporting source; when the evidence derivation relationship of a certain supporting source in the evidence source graph cannot be determined, it is determined that the supporting source has a common pollution source node with other supporting sources.

[0015] Further, determining the corroboration weight of the causal edge includes: for each independent support source, determining the time and spatial deviations between the time point and location carried by the support source and the event units at both ends of the causal edge, and determining the spatiotemporal consistency degree that decreases as the time and spatial deviations increase; using the product of the modal reliability of the mode to which the support source belongs and the spatiotemporal consistency degree as the single-source corroboration degree; multiplying the differences obtained by subtracting each single-source corroboration degree, and using the result of subtracting the multiplication as the corroboration weight of the causal edge; before determining the corroboration weight, it further includes: for any two support sources supporting the same causal edge, determining the time difference between the times carried by the two sources. The cross-modal causal divergence is obtained by adding the compatibility, spatial inaccessibility between the locations they occupy, and logical repulsion between their action predicates. The logical repulsion is penalized by a preset value when the action predicates of the two constitute a mutually exclusive pair in a preset mutually exclusive predicate list, and is zero otherwise. When the cross-modal causal divergence exceeds the contradiction determination threshold, the modal reliability of the two is compared: if the difference between the modal reliability of the two is not less than the reliability discrimination threshold, the support source with higher modal reliability is retained, the other support source is marked as an item to be verified, and the confirmation weight of the causal edge is re-determined based on the retained support source; if the difference between the modal reliability of the two is less than the reliability discrimination threshold, the confirmation weight of the causal edge is reset to zero.

[0016] Furthermore, before matching the candidate element subgraph with the target element subgraph, the method further includes: using the product of the action predicate slot of the action node mapped according to a preset intensity table to obtain the action intensity, the legal interest sensitivity mapped according to the legal interest type of the passive object slot of the action node, and the spatiotemporal proximity between the action node and the event unit corresponding to the victim as the danger creation quantity; using the product of the confirmation weight of the causal edge and the time distance decay factor that decreases as the time interval between the event units at both ends of the causal edge increases as the edge power of the causal edge; taking the maximum value of the product of the edge power of each causal edge on each directed path from the action node to the result node in the fact causal graph; when the passive object slot of the result node falls into the element result range, the maximum value is taken as the danger realization degree, otherwise the danger realization degree is zero, and the element result range is the set of result types that the danger created by the action node may point to in the specification.

[0017] Furthermore, the method further includes: identifying event units that do not originate from the actor corresponding to the behavior node as exogenous intervention nodes on the directed path corresponding to the danger realization degree; determining the intervention effectiveness split ratio to be zero when there are no exogenous intervention nodes on the directed path; multiplying the differences obtained by subtracting the edge effectiveness of each exogenous intervention node from the incoming edge effectiveness on the directed path, using the result of subtracting the multiplication as the intervention effectiveness, and using the ratio of the intervention effectiveness to the sum of the intervention effectiveness and the danger realization degree as the intervention effectiveness split ratio; when the intervention... When the effectiveness split ratio is not greater than the blocking threshold, the candidate element subgraph is maintained; when the intervention effectiveness split ratio is greater than the blocking threshold but not greater than the sum of the blocking threshold and the tiered bandwidth, and the danger creation amount is not lower than the criminalization threshold, the directed path is removed and the candidate element subgraph is reconstructed into an attempted element subgraph; when the intervention effectiveness split ratio is greater than the sum of the blocking threshold and the tiered bandwidth, or when the intervention effectiveness split ratio is greater than the blocking threshold and the danger creation amount is lower than the criminalization threshold, the directed path is removed and the candidate element subgraph is reconstructed into a non-causal element subgraph.

[0018] Furthermore, the requirement template also includes requirement layer edges connecting requirement nodes. Obtaining candidate requirement subgraphs further includes: for each requirement layer edge, when both requirement nodes at its two ends are activated, and there exists a directed path in the fact-cause graph from the fact subgraph corresponding to the previous requirement node to the fact subgraph corresponding to the next requirement node, with the direction consistent with the requirement layer edge, the requirement layer edge is activated, and the maximum value of the product of the edge effectiveness of each causal edge on the directed path is taken as the implementation effectiveness of the requirement layer edge; otherwise, the requirement layer edge is not activated, and the inter-layer structure is recorded as missing.

[0019] Furthermore, the method also includes: taking the causal edges associated with the behavior nodes in the fact-cause graph as evidence edges, and classifying the evidence edges into the pre-behavior period, behavior period, and post-behavior period according to the chronological relationship between the timestamp intervals of the evidence edges and the timestamp intervals of the behavior nodes; for each evidence edge, finding the occurrence rate of the action predicate slot value of its endpoint event unit under the assumption of intent and the assumption of non-intent, and taking the natural logarithm of the ratio of the two, wherein the occurrence rate of the condition is obtained statistically based on the factual findings in the effective judgment documents; weighting and summing the natural logarithm with the corroboration weight of the evidence edge, and superimposing the prior logarithmic probability determined according to the case type to obtain the subjective cognitive intensity; when the subjective cognitive intensity is not less than the threshold for the establishment of the subjective element, the subjective element is determined to be established.

[0020] Furthermore, the weighted accumulation also includes a homology reduction factor, which decreases as the number of source nodes shared by the supporting sources of the evidence edge and other evidence edges in the same time period in terms of evidence derivation relationship increases; before determining the establishment of subjective requirements, it also includes: counting the number of time periods with evidence edges in the three time periods as evidence completeness, shrinking the subjective cognitive intensity to zero in a manner that the lower the evidence completeness, the greater the shrinkage, and comparing the shrunken subjective cognitive intensity with the threshold for the establishment of subjective requirements.

[0021] Furthermore, when the subjective element is met, the process also includes: obtaining the slot values ​​of the subjective cognitive graph in the three dimensions of object, means, and result by merging the evidence edges of each time period, and reading the slot values ​​of the actual legal interest infringement graph in the three dimensions; for each dimension, determining whether the slot value of the subjective cognitive graph falls into the synonym domain of the corresponding slot in the actual legal interest infringement graph, and determining that the dimension is misaligned if it does not fall into the synonym domain; when any dimension is misaligned, routing the matching process to the causal cognition error branch element subgraph corresponding to the misaligned dimension, wherein the object dimension, means dimension, and result dimension correspond to object error, attack error, and causal process error, respectively.

[0022] Further, the step of outputting a verification conclusion based on the matching results includes: using the product of the hierarchical matching value, the culpability form factor, and the key edge reinforcement coverage rate as the probative value score of the target element subgraph; the culpability form factor is determined based on whether the subjective element is established when the culpability form of the target element subgraph is intentional, and based on whether the duty of care is violated when it is negligent; the key edge reinforcement coverage rate is the proportion of causal edges with at least two independent supporting sources among the key causal edges on which the target element subgraph depends; for each preset alternative causal hypothesis subgraph, using the same method as determining the projection value, searching the factual subgraph in the factual causal graph for the factual subgraph that maximizes the projection value of the alternative causal hypothesis subgraph, and obtaining the explanatory value score of the alternative causal hypothesis subgraph; using the difference between the probative value score and the maximum value among the explanatory value scores as the exclusivity difference, and selecting an exclusivity threshold according to the nature of the case; when the exclusivity difference is not less than the exclusivity threshold, outputting a verification conclusion that the verification is passed, otherwise outputting a verification conclusion that the probative value is insufficient, and attaching a list of alternative causal hypothesis subgraphs that have not been excluded.

[0023] Furthermore, the alternative causal hypothesis subgraph includes alternative causal hypothesis subgraphs obtained through at least one of the following methods: formalized from the defense arguments presented in the defense opinion; retrieved from a defense type library corresponding to the case type; or formalized when there exists a causal path in the factual causal graph leading to the result involved in the target element subgraph without going through the actions of the accused.

[0024] Furthermore, before outputting the verification conclusion that the verification has passed, the process includes: classifying each of the alternative causal hypothesis subgraphs into three categories: no actor involved, actor involved but not caused by the act itself, and victim's own cause; when the number of alternative causal hypothesis subgraphs in any category is zero, the verification conclusion is downgraded to preliminary verification passed, and the alternative hypothesis set is marked as not closed and the missing categories are marked.

[0025] Furthermore, when determining each intermediate quantity, a traceability record for that intermediate quantity is generated simultaneously. The traceability record includes the input quantity identifier, the criterion name used, the parameter value, the threshold, and the measured value of that intermediate quantity.

[0026] Each of the aforementioned source tracing records is stored in the form of a directed acyclic graph; when the verification conclusion is output, the verification conclusion is recursively expanded into a derivation tree with the evidence materials as leaf nodes according to its input quantity, and it supports querying the intermediate quantities and verification conclusions affected by the specified evidence materials.

[0027] In another aspect, the present invention provides a legal constituent element subgraph matching and verification system based on a multimodal causal knowledge graph, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the legal constituent element subgraph matching and verification methods based on a multimodal causal knowledge graph as described above.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0029] 1. This invention decouples the case map into a physical fact layer and a legal requirement layer, and constructs the cross-layer mapping as the product of slot structure alignment and cross-modal corroboration weight. This ensures that the mapping value is always zero when the structure alignment is zero or the corroboration weight is all zero. The two types of false activations, namely, close descriptions but insufficient evidence and sufficient evidence but structural inconsistencies, are structurally eliminated. Furthermore, since the structure alignment is obtained by looking up a table in the synonym domain, the corroboration weight is calculated by a deterministic criterion, and the normative correlation coefficient is statically configured, the mapping result of the same piece of evidence data under the same normative template is uniquely determined and can be recalculated bit by bit. This eliminates the normative mapping drift and illusion caused by fluctuations in text surface similarity in a single-layer flat map.

[0030] 2. This invention determines the direction of directed edges by using a conjunctive hard judgment based on two criteria: temporal preconditions and spatiotemporal reachability. It eliminates loops solely based on minimum temporal margin, so that the causal direction no longer depends on the semantic model's judgment and drifts with the extracted wording. Furthermore, the acyclicity of the first layer is guaranteed by deterministic rules, providing a premise for the path effectiveness calculation of the topologically ordered scan. By replacing the transitive closure with the intervention effectiveness split ratio as the criterion for the continuation of the causal chain, the algorithm can still determine the interruption of the causal chain even when the path from the behavior node to the result node is connected. Moreover, the split ratio takes a normalized competitive form, where the stronger the main causal effectiveness, the smaller the share of the same intervention effectiveness. This allows for the classification of intervention factors of different intensities, and the subgraph of the elements is reconstructed into incomplete or non-causal types according to the values ​​of the split ratio and the amount of danger created, avoiding the situation where the results caused by the intervention factors are all attributed to the original actor.

[0031] 3. This invention multiplies the threshold values ​​of the two levels of illegality exclusion and liability exclusion with the continuous matching value of the appropriateness level in a multiplicative manner, so that when any exclusion is established, the level matching value is always equal to zero and is completely unrelated to the appropriateness score. The veto is transformed from a normative expression into an algebraic identity, with no compensation path. At the same time, the appropriateness level retains a continuous form to support the ranking among multiple candidate normative templates, achieving both optimizability and non-overstepping. Furthermore, the subjective intensity is determined by weighted summation of the log-likelihood ratios of evidence edges in three time periods: before the act, during the act, and after the act. This ensures that the determination of subjective elements does not depend on the defendant's admission, and can still provide auditable conclusions line by line even when the defendant remains silent or recants, avoiding the direct substitution of objectively occurring results for subjective cognition. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0033] Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention.

[0034] Figure 2 This is a comparison chart of the misjudgment rates of closely adjacent events during the same time period, provided in Embodiment 1 of the present invention.

[0035] Figure 3 This is a comparison diagram of the convergence of causal effect propagation provided in Embodiment 1 of the present invention.

[0036] Figure 4 This is a comparison diagram of two independent source counting methods provided in Embodiment 1 of the present invention.

[0037] Figure 5 This is a diagram illustrating the suppression effect of activation rates of three types of causal edges provided in Embodiment 1 of the present invention.

[0038] Figure 6 This is a comparison chart of cumulative scores for the same fact to be proven, provided in Embodiment 1 of the present invention.

[0039] Figure 7 This is a comparison diagram of the multiplicative zeroing characteristic and the weighted summation scheme provided in Embodiment 1 of the present invention.

[0040] Figure 8 This is a comparison chart of the increase in access counts provided in Embodiment 1 of the present invention.

[0041] Figure 9 This is a comparison chart of the decay characteristics of the degree of danger realization along the causal chain length provided in Embodiment 1 of the present invention.

[0042] Figure 10 This is a comparison diagram of causal interruption identification provided in Embodiment 1 of the present invention.

[0043] Figure 11 This is a diagram illustrating the impact of the thoroughness of counterfactual truncation on the limit contribution rate provided in Embodiment 1 of the present invention.

[0044] Figure 12 This diagram illustrates the effect of evidence completeness correction on subjective cognitive strength, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0046] It should be noted that, in this application, multi-source evidence data may originate from case files transferred by investigative authorities, evidence materials submitted by litigants, expert opinions issued by judicial appraisal institutions, audio and video data retrieved from public video surveillance systems, data mirrors exported from electronic data forensics platforms, or other legitimate sources of evidence from which factual information of the case can be obtained. Multi-source evidence data must include at least two different modalities of evidence materials; optionally, multi-source evidence data may also include three-dimensional scan data of physical evidence, structured exported data of communication records, tabular data of financial transactions, or composite evidence units obtained by merging multiple basic modalities. In this application, legal elements refer to the set of elements and their interrelationships stipulated in legal norms that serve as prerequisites for the occurrence of specific legal consequences. These can be elements of a crime organized according to the three-tiered theory or four-element theory in criminal law, elements of tort liability or breach of contract in civil law, or elements of administrative illegality in administrative law. In this application, subgraph matching of constituent elements refers to the process of determining whether there is a correspondence between a factual causal graph constructed from multi-source evidence data and a constituent element template defined by legal constituent elements, which satisfies a pre-set criterion. The output of subgraph matching is not the legal conclusion itself, but a verification report for adjudicators to review, which carries a complete judgment trajectory and evidence tracing chain.

[0047] Example 1

[0048] This embodiment discloses a method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs. Figure 1 The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:

[0049] Step 1: Obtain multi-source evidence data of the case to be verified, extract event unit sets from the multi-source evidence data, and perform identity merging on event units describing the same event.

[0050] Among them, cases awaiting verification refer to specific cases where subgraph matching verification is required to determine whether the elements constituting the crime are established. These can typically be a criminal case that has completed investigation and evidence collection and entered the review, prosecution, or trial stage, or a civil case where evidence has been presented. For example, a case awaiting verification could be an intentional injury case with surveillance video, witness testimony, and forensic opinions as the main evidence, or a contract fraud case with electronic data, communication records, and written contracts as the main evidence.

[0051] An event unit refers to the smallest factual unit extracted from original evidence that can independently describe the occurrence of a particular event. "Smallest" means that the event described by this unit cannot be meaningfully divided in time and space; "independent" means that the unit itself carries all the attributes necessary to determine its causal relationship with other events.

[0052] In this embodiment, each event unit carries six fixed attributes: a modal identifier indicating its source modality, a timestamp interval represented by a closed interval, three-dimensional spatial coordinates, an action predicate slot describing the action type, an object slot describing the agent and the receiver, and modal reliability characterizing the legal evidentiary force and compliance of the evidence collection procedure of the modality.

[0053] In this context, a timestamp interval refers to a closed interval defined by the earliest possible start point and the latest possible end point, used to represent the possible position of the event on the timeline. It's understandable that the reason for insisting on using intervals rather than specific moments to represent time information is that the temporal precision of multimodal evidence varies by several orders of magnitude: surveillance video can be accurate to the frame, with its interval width approaching zero; while witness testimony often only provides information around 3 PM, with its interval width reaching the hour level. If all were converted to specific moments (e.g., taking the midpoint of the interval), low-precision evidence would be disguised as high-precision evidence, and subsequent temporal prioritization judgments would give a definite chronological order without any basis. Using intervals to represent time allows the precision differences to be explicitly included in subsequent criteria without being smoothed out; this is the first structural defense against normative mapping drift in this method.

[0054] Spatial location coordinates refer to the three-dimensional coordinates of the location where the event occurred within the unified spatial coordinate system of the case. It should be noted that three-dimensional coordinates are used here instead of two-dimensional planar coordinates to accommodate vertical differences such as floors and steps; in indoor cases, the spatial and temporal accessibility of adjacent locations on different floors with the same planar coordinates is completely different from that of adjacent locations on the same floor. If compressed into two dimensions, the former would be misjudged as spatial overlap.

[0055] Action predicate slots refer to semantic slots that describe the type of action involved in the event. Their values ​​are taken from a closed factual vocabulary list, such as hitting, carrying a weapon, calling for help, driving away, transferring money, deleting, etc. It is understandable that the closed nature of the factual vocabulary list is a prerequisite for subsequent synonym lookup. If any natural language expression is allowed to enter the slot, the synonym attribution judgment will degenerate into semantic similarity calculation. The drift that this method aims to eliminate is thus reintroduced.

[0056] Modal reliability refers to a scalar value characterizing the probative value of a particular modality of evidence. Its value is determined jointly by the legal probative value sequence of the evidence type and the compliance of the evidence collection procedure. For example, in this embodiment, authenticated physical evidence can be rated 0.95, interrogation transcripts with synchronized audio and video recordings can be rated 0.85, witness testimonies without synchronized audio and video recordings can be rated 0.60, and electronic data of unknown origin can be rated 0.30.

[0057] It should be noted that modal reliability is the only scalar in this method with a priori properties, and its value is set once across the entire case and does not change with the content of the case. The reason for emphasizing this point is that if this parameter were allowed to be adjusted according to the content of the case, the algorithm would have a knob that could be arbitrarily adjusted to change the conclusion, and its output would no longer be auditable, while auditability is a prerequisite for the existence of judicial support methods.

[0058] In this embodiment, when extracting event units from multi-source evidence data, corresponding front-end processing is invoked for different modalities: for video, action segmentation and target association processing are invoked; for audio, speech transcription and speaker separation processing are invoked; for graphic materials, layout parsing and concluding sentence location processing are invoked; for text materials, sentence-level event extraction processing is invoked; and for physical evidence and electronic data, structured reading of their formal requirements and content elements is invoked. It should be noted that the aforementioned front-end processing is not the inventive point of this method; any existing implementation capable of outputting the above six attributes can be used. The only requirement of this method is that its output format is strictly unified to the above six tuples, thereby making heterogeneous modalities comparable and connectable within the same formal space.

[0059] In this embodiment, to avoid the same event being repeatedly recorded in different modalities and generating false multiple event chains at the fact layer, after generating the event unit set and before constructing the causal edge, the method of identity merging of event units is also included. The specific implementation method is as follows:

[0060] For any two event units, calculate the intersection-union ratio of their timestamp intervals and the Euclidean distance between their spatial locations;

[0061] For each of the two, the action predicate slot, the agent slot, and the patient slot are looked up in a table to determine whether they fall into the same synonym domain.

[0062] When the temporal intersection-union ratio is not lower than the temporal identity threshold, the spatial Euclidean distance does not exceed the spatial identity distance threshold, and all three slots fall into the same synonym domain, it is determined that the two describe the same event.

[0063] Two event units that are determined to be the same event are merged into the same event unit node. The slot value and spatial position of the one with higher modal reliability are taken as the merged value. The intersection of the timestamp intervals of the two is taken as the merged timestamp interval. The sum of the modal identifiers of the two is taken as the merged modal identifier. The support sources of the two are merged into the same support source set.

[0064] For example, in this embodiment, the event unit and the fact space can be represented by the following formula:

[0065]

[0066]

[0067] The event unit identity merging operator can be represented by the following equation:

[0068]

[0069]

[0070]

[0071]

[0072] in, For the first A fragment of original evidence that can independently describe a particular event; Extract operators for event units; Modal identifier; , These represent the earliest possible start point and the latest possible end point of the event in terms of time. Spatial location coordinates; To assign a value to the action predicate slot; , These are the values ​​for the slots of the action object and the received object, respectively. Modal reliability; For event unit set, It is the total number of its elements; The set of possible values ​​for the modal identifier; The intersection-union ratio of the two time intervals; The threshold for time identity is 0.5, for example. The spatial identity distance threshold is 2 meters, for example, in an indoor scene; This is a synonymous field operator, and its definition is given in step 4.

[0073] Understandably, the logic behind merging the data by taking the intersection of the timestamp intervals rather than the union is that since both materials describe the same event, that event must fall within both intervals simultaneously. Taking the intersection refines rather than loses the temporal information. This is precisely the direct benefit of multimodal verification in the temporal dimension. Once high-precision modalities (such as surveillance video) and low-precision modalities (such as witness testimonies) are merged, the temporal location of the event is tightened to the level of the high-precision modality, while retaining the independent source identity provided by the low-precision modality.

[0074] It should be noted that the merged modal identifier is the sum of the two, which means that the merge provides the node with the support of two modalities, providing the correct input for the independent modal counting of the subsequent reinforcement rule. If the identity merging step is ignored, the blows captured in the video, the blows stated by the witness, and the blow injuries confirmed by the expert opinion will each become independent nodes, resulting in a false three-way causal chain of blows at the factual level. At the same time, the counting of the reinforcement rule will also be wrong, as it will mistakenly treat the three modalities of the same event as the three isolated pieces of evidence for each event, thus preventing the key facts that should pass the reinforcement test from passing.

[0075] Step 2: Perform joint verification of the temporal precondition criterion and the spatiotemporal reachability criterion on any two event units in the event unit set, connect the event units that pass the verification into causal candidate edges, and perform loop resolution on the resulting directed graph to generate a causal directed acyclic graph of the physical fact layer.

[0076] The physical fact layer refers to the lower-level topology consisting solely of the actual life facts that occurred in the case, where nodes are event units and edges represent the causal relationships between event units. It is called the physical fact layer to distinguish it from the upper-level legal elements layer. No normative concepts appear in this layer; terms such as subjective knowledge, illegal possession intent, and abuse of power are strictly prohibited as nodes in this layer.

[0077] Causal candidate edges refer to directed edges that have passed the joint verification of temporal preconditions and spatiotemporal reachability, but have not yet been verified by the reinforcement rule. They are called candidates because whether they can enter the subsequent scoring depends on the edge activation result in step 3.

[0078] The principle of temporal precedence means that the event that is the cause must not be later than the event that is the effect, and the two events must not be almost exactly contemporaneous. The latter restriction is necessary: ​​if the time intervals of the two events almost completely overlap, it is impossible to distinguish the order, and the direction obtained by forcibly orienting them is not recalculated.

[0079] Spatiotemporal reachability refers to the requirement that the spatial distance between two events must be within their available time window and within the maximum displacement that can be covered by the scene's maximum displacement rate. In essence, this criterion means whether it is possible for a person to traverse this distance within the available time between two events. If not, then there cannot be a direct causal relationship between the two events, regardless of how similar their semantic expressions may be.

[0080] In this embodiment, to prevent the direction of the causal edge from depending on the semantic model's judgment, thereby eliminating the drift of direction due to extracted wording, the specific implementation of causal orientation and acyclicity is as follows:

[0081] For any pair of event units, read the upper and lower bounds of their timestamp intervals. If the lower bound of the former is not later than the lower bound of the latter, and the ratio of the overlap duration of the two intervals to the length of the shorter interval does not exceed the overlap tolerance, the timing precedence is determined to be valid, and the direction is from the former to the latter.

[0082] Calculate the Euclidean distance between the spatial positions of the two event units, and calculate the maximum displacement that can be covered by the maximum displacement rate of the scene within the available time window from the earliest starting point of the former to the latest ending point of the latter. When the Euclidean distance does not exceed the sum of the maximum displacement and the position error tolerance, the spatiotemporal reachability is determined to be valid.

[0083] A causal candidate edge pointing from the former to the latter is established only if both of the above conditions are met; otherwise, the event unit pair is discarded.

[0084] Perform a depth-first search on the established directed graph. For each cycle detected, delete the edge with the smallest temporal pre-margin in the cycle. Repeat this process until there are no cycles in the graph.

[0085] For example, in this embodiment, the joint criterion for causal orientation can be calculated using the following formula:

[0086]

[0087]

[0088]

[0089] in, For the reason point to directional indication quantity; The overlap duration between the two time intervals; The length of the shorter of the two intervals; To prevent extremely small constants from being divided by zero, we take, for example, a value of 100%. ; For the overlap tolerance, an example value of 0.2 is used; The maximum displacement rate of the scene; For self The earliest possible starting point is Take the positive part of the available time window between the latest possible endpoints; For spatial position error tolerance, 1.5 meters is taken as an example; Let be the set of causal candidate edges obtained by orientation.

[0090] The maximum displacement rate of the scene is determined according to the scene type of the case: for example, 3 meters per second for indoor walking scenes and 30 meters per second for outdoor scenes involving vehicles. In this embodiment, the scene type can be automatically determined by the on-site spatial scale and the involvement of vehicles recorded in the case file. If the maximum straight-line distance recorded in the on-site investigation record does not exceed 50 meters and there is no record of vehicles, it is judged as an indoor walking scene; otherwise, it is judged as an outdoor scene involving vehicles.

[0091] It is understandable that the available time window is positive because when the latest endpoint of the latter is earlier than the earliest starting point of the former (i.e., the time sequence is completely reversed), the time window is negative. After taking the positive part, it becomes zero, so the displacement budget is zero. All cases except the same point are unreachable. This makes the event pair with reversed time sequence intercepted for the second time at the reachability criterion, forming redundant protection.

[0092] It should be noted that this criterion is a substantial improvement over the previous method of dividing the distance by the time difference by the rate limit. The original method resulted in the denominator approaching zero when the centers of the two events' time intervals coincided, causing the ratio to jump to an extreme value. This would invariably classify adjacent events occurring at the same time as unreachable, which is extremely common in surveillance scenarios, thus losing a large number of real causal edges at the factual level. By using the displacement budget method, not only is the division by zero eliminated, but two events occurring sequentially at the same location (with a zero Euclidean distance) are guaranteed to be reachable, which aligns with the normality of case facts. Figure 2 This diagram shows a comparison of the misjudgment rates of the two spatiotemporal reachability representations in this embodiment for adjacent events occurring simultaneously; from Figure 2 As can be seen, at the level with the smallest center interval of the event time interval, the ratio-based reachability criterion will misclassify up to 71.5% of truly reachable events as unreachable, while the displacement-based criterion of this solution has a zero misclassification rate at the same level. In monitoring scenarios, closely adjacent events at the same time are no longer systematically excluded from the fact layer, and the recall of causal edges at the fact layer is substantially guaranteed.

[0093] It's understandable that the directional indicator is taken as the product of two indicator functions rather than their weighted sum because, under the conjunctive criterion, if any condition is not met, the entire function immediately becomes zero, leaving no compensation path. If a weighted sum were used instead, the highly reliable temporal precession could compensate for obvious spatiotemporal inaccessibility to some extent, thus allowing a physically impossible causal edge to pass. Once such an edge enters the graph, it will continuously pollute the attribution conclusion in subsequent path calculations, and because it is completely legal in form, it is extremely difficult to detect.

[0094] For example, in this embodiment, the loop resolution operator can be represented by the following formula:

[0095]

[0096]

[0097] in, For loop elimination operators; The resulting causal directed acyclic graph represents the physical fact layer after resolution. This is the set of loops detected by depth-first search; It is one of the loops; This represents the timing margin of the side, i.e., how much earlier the cause precedes the effect.

[0098] It is understandable that using temporal margin as the basis for deletion has clear rationale: the edge with the smallest margin is closest to the critical state where the order is indistinguishable, and is therefore the least reliable in terms of orientation in a loop. Deleting it causes the least damage to the graph structure, and the deletion result is entirely determined by the margin, a deterministic quantity, without any randomness. Therefore, the same evidence data will always produce the same fact layer graph in two runs.

[0099] It is important to note that acyclicity is a prerequisite for subsequent calculations. If causal loops exist in the graph (e.g., strike → fall → strike), the path product of the danger realization degree in subsequent step 5 will propagate infinitely along the loop, rendering the intervention effectiveness split ratio meaningless. Furthermore, the recalculation by topological order in subsequent step 6 will be impossible because graphs containing loops do not have a topological order. Therefore, the fact layer needs to be a directed graph with recalculated directions and an acyclic structure to lay the foundation for path calculations on it. Figure 3 This diagram shows a comparison of the convergence of causal force propagation before and after loop resolution in this embodiment; from Figure 3 As can be seen from the above parameter settings, the residual of the effectiveness propagation on the fact layer containing the loop always stays around 0.18 and does not converge to zero, while the residual of the fact layer after loop resolution drops to the level of machine accuracy after the fifth round. The difference between the two is about eleven orders of magnitude.

[0100] For each causal candidate edge in the layered causal directed acyclic graph, the number of independent modalities supporting the edge is counted, and it is detected whether there is a common pollution source set in the evidence source graph for each supporting source. Causal candidate edges that meet the reinforcement conditions are assigned cross-modal corroboration weights and set to the active state. At the same time, cross-modal evidence contradictions are detected and resolved, and an activated fact causal graph is generated.

[0101] The corroboration rule refers to the rule in judicial proof that certain key facts cannot be established based on evidence from a single source. In essence, this means that the establishment of a key causal relationship requires support from two or more independent sources of evidence.

[0102] Supporting sources refer to specific evidence that provides factual basis for a causal candidate, such as a surveillance video file, an interrogation record, or an expert opinion.

[0103] The set of public pollution sources refers to the statement subject node or evidence collection behavior node that is common to two supporting sources in the evidence source diagram.

[0104] An evidence source graph is a directed graph consisting of nodes representing evidence carriers, evidence collection actions, and the subject of a statement, with edges representing derived relationships. For example, a video clip may derive from a specific retrieval action, which in turn derives from a surveillance camera; a written record may derive from a specific interrogation, which in turn derives from a specific subject of a statement.

[0105] In this embodiment, the evidence source diagram can be constructed by sequentially reading the derived relationships from the evidence catalog in the case file, the procedural records of evidence collection, and the formal requirements of the evidence materials. It should be noted that when the case file is incomplete or the source relationship of a supporting source cannot be determined, it is conservatively treated as having a common source of pollution, meaning its independence is not recognized. This is because the normative intent of the corroboration rule is to prevent a single piece of evidence from being used to determine a case. If evidence with an unclear source is presumed to be independent, the protective effect of this rule is rendered ineffective; it is better to temporarily exclude genuinely independent evidence than to allow substantially similar evidence to masquerade as an independent source.

[0106] In this embodiment, to identify pseudo-multimodalities, i.e., supporting sources that superficially belong to different modes but are actually from the same source, the specific implementation method for detecting the common pollution source set is as follows:

[0107] Construct a diagram of the sources of evidence in the case;

[0108] For any two sources supporting the same causal candidate edge, calculate their source set in the evidence source graph by tracing upwards along the derived edge.

[0109] When there is a statement subject node or the same evidence collection behavior node in the intersection of two source sets, it is determined that the two have a common pollution source set.

[0110] Supporting sources of common pollution sources are combined into one when counting the number of independent modes.

[0111] For example, in this embodiment, the evidence source map, pollution equivalence relation, and effective count of independent modes can be obtained by the following steps.

[0112] First, define the evidence source map and the pollution equivalence relations on it:

[0113]

[0114]

[0115]

[0116] Secondly, by taking the quotient of this equivalence relation, we obtain the effective count of independent modes:

[0117]

[0118] in, Image showing the source of the evidence; , , These are respectively the set of evidence carrier nodes, the set of evidence collection behavior nodes, and the set of statement subject nodes; For sets derived from edges; Let be the set of source sources obtained by tracing back along the derived edges in the evidence source graph; This indicates that there is a pollution equivalence relationship between the two supporting sources; To support causal edges The set of supporting sources; This indicates that the quotient is taken according to the pollution equivalence relationship, that is, sources with the same pollution source are combined and counted as one; This is the effective count of independent modalities, i.e., the number of unrelated support sources.

[0119] It should be noted that the definition of pollution equivalence includes both the subject of the statement and the evidence-gathering behavior, rather than just the subject of the statement, because they correspond to two different sources of correlation: sharing the subject of the statement leads to subjective bias, as two statements by the same person are inevitably consistent; sharing the evidence-gathering behavior leads to procedural correlation, as oversights in the same inspection will simultaneously affect all its outputs. Both types of correlation are sufficient to negate independence in terms of proof theory, hence their inclusion. If only the subject of the statement is considered, the combination of photographs and inspection records from the same inspection, which is extremely common in practice, would be mistakenly considered as two independent modalities. Figure 4A comparison diagram of two independent source counting methods under pseudo-multimodal material addition in this embodiment is shown; from Figure 4 As can be seen, when the same source material is continuously added, the number of independent sources obtained by counting according to the modal identifier increases linearly to 9 with the number of added copies, which has long exceeded the reinforcement lower limit. However, the number of independent sources obtained by folding the source of evidence map is always 1, which is always below the reinforcement lower limit. The source set intersection criterion makes the same source material count as only one independent source no matter how many copies are added, thus blocking the circumvention path of artificially increasing the modal number by stacking the output of the same subject or the same evidence collection behavior.

[0120] In addition, after completing the independence determination, in order to distinguish between the two cases of barely passing the reinforcement test and the high consistency verification of multiple sources in the subsequent normative projection, it is also necessary to assign a quantitative verification weight to the edge that passes the reinforcement test.

[0121] Among them, cross-modal confirmation weight refers to a scalar that characterizes the confirmation strength obtained by a certain causal candidate edge. It is obtained by aggregating the modal reliability and spatiotemporal consistency of each independent support source, and takes the establishment of the reinforcement rule as a multiplicative premise.

[0122] Spatiotemporal consistency refers to the degree of agreement between the time and location recorded in a supporting source and the spatiotemporal attributes of the event units at both ends of the causal edge. It is understandable that even if a testimony points to the same event, if the time point it describes differs greatly from the video recording, its corroborating effect on that edge should be reduced. Although it does not constitute a contradiction, its identification is not accurate enough.

[0123] For example, in this embodiment, the cross-modal verification weights can be calculated using the following formula:

[0124]

[0125]

[0126]

[0127] in, To verify weights across modalities; To support the equivalence classes of sources under the pollution equivalence relation, multiplying them means that each independent source contributes only once; Modal reliability of this source; The spatiotemporal consistency between the source and the facts stated on the edge; , These are the deviations between the time point, location, and spatiotemporal attributes of the event units at both ends of the edge, as recorded by the source. , As a spatiotemporal tolerance scale, 300 seconds and 5 meters are used as examples respectively; To reinforce the hard gates of the rules; To activate the edge set.

[0128] Understandably, the reason the aggregation formula uses the form of multiplying the complements of each term by one, rather than an arithmetic mean, is that its design intention is to make the mathematical form isomorphic to the normative intent of the reinforcement rule. This form implies that multiple independent sources can independently prove the edge, and any one of them proves it. Therefore, adding an independent source only increases the weight, and the increase decreases with the number of existing sources. In contrast, the arithmetic mean can result in adding a source with slightly lower reliability, which is completely contrary to the normative intent. Adding more corroborating evidence reduces the strength of the finding, which is unacceptable from a judicial perspective.

[0129] It should be noted that the multiplicative indicator of the reinforcement rule, the hard gate, is the core of this step. Its irreparable nature leads to two deterministic consequences: 1) The corroboration weight of an isolated evidence edge is always zero, so this edge will not be included in the matching score of any level thereafter, no matter how crucial its content or how certain its expression; 2) The source set intersection criterion automatically folds pseudo-multimodal evidence, blocking the path of artificially increasing the number of modalities by adding materials from the same source. If this gate is omitted, a victim's statement can support a key causal edge, the lower limit of the entire proof system will be pulled down to the level of isolated evidence, and the applicability of the solution in judicial scenarios will cease to exist. Figure 5 The diagram illustrates the suppression effect of the reinforcement rule hard gate on the activation rates of three types of causal edges in this embodiment; from Figure 5 As can be seen, without the application of the hard gate, the activation rates of isolated evidence edges and pseudo-multimodal edges are both 100%. After the application of the hard gate, the activation rates of the two edges drop to zero, while the activation rate of true multi-source edges remains at 100%. In other words, the hard gate precisely intercepts the two types of unqualified support, isolated evidence and pseudo-multimodal edges, in an uncompensated multi-fold manner, while not inhibiting the causal edges that truly obtain multi-source independent verification.

[0130] Furthermore, if two contradictory multimodal pieces of evidence are scored concurrently, it would be equivalent to awarding points to each side of the contradiction, which is logically absurd. Therefore, before activating an edge, it is necessary to detect and forcibly resolve cross-modal contradictions.

[0131] Among them, a pair of contradictory evidence nodes refers to two sources of support that support the same causal candidate edge but are incompatible in time, unreachable in space, or whose action predicates are logically mutually exclusive.

[0132] Cross-modal causal divergence is a scalar used to measure the degree of incompatibility between two supporting sources. It consists of the sum of three terms: temporal incompatibility, spatial inaccessibility, and logical repulsion.

[0133] In this embodiment, to avoid the double-buffing of contradictory evidence, the specific implementation method for contradiction detection and resolution is as follows:

[0134] For any two sources of support supporting the same causal candidate edge, calculate their cross-modal causal divergence.

[0135] When the divergence exceeds the contradiction determination threshold, the two supporting sources are marked as a pair of contradictory evidence nodes, and the activation states of all causal candidate edges supported by them are frozen.

[0136] Compare the modal reliability of the two contradictory parties: when the difference between the two is not less than the reliability discrimination threshold, retain the one with higher reliability, transfer the one with lower reliability to the item to be verified, and recalculate the cross-modal verification weight with the retained set of supporting sources.

[0137] When the difference between the two is less than the reliability discrimination threshold, it is determined that they are indistinguishable, and the cross-modal verification weight of the causal candidate edge is directly set to zero so that the edge does not participate in the matching and scoring of any subsequent layer.

[0138] For example, in this embodiment, the cross-modal causal divergence and contradiction resolution law can be expressed by the following formula:

[0139]

[0140]

[0141]

[0142] in, For cross-modal causal divergence; This represents the minimum distance between two time intervals; the value is zero when the two intervals intersect. The center of the interval; This indicates that two action predicates form a mutually exclusive pair in a pre-defined list of mutually exclusive predicates, such as being present and not being present, armed and unarmed, driving away and staying in place; The penalty value is a fixed value for logically contradictory conditions, exemplarily set to 1.0; The threshold for determining contradiction is set to 1.0, meaning that simple logical contradiction is sufficient to trigger resolution. The reliability discrimination threshold is set to 0.15 for example. The mode with lower reliability is the one that is contradictory to the middle. Figure 6 This embodiment shows a comparison chart of the cumulative scores for the same fact to be proven before and after the resolution of contradictory evidence; from Figure 6 As can be seen, after the cross-modal causal divergence exceeds the contradiction judgment threshold, the cumulative score of the scheme without resolution for the same fact to be proved is still 1.63, that is, each side of the contradiction adds points, while the optimal retention method reduces it to 0.85, and the whole side zeroing method reduces it to zero.

[0143] Step 4: Perform canonical causal projection on the activated fact causal graph using the preset statutory requirement layer requirement template. The canonical causal projection is a bilinear product of slot structure alignment and cross-modal confirmation weight. Apply inter-layer consistency constraints to the requirement layer edges to generate candidate requirement subgraphs.

[0144] The statutory elements layer refers to the upper-level topology composed of the normative concept nodes defined by the constituent elements template and the normative relationships between them. Element nodes include objectively harmful acts, constituent element results, subjective knowledge, and legitimate defense; element layer edges include the attribution relationship between act and result.

[0145] A requirement template is a formalized description of the constituent elements of a specific crime or type of liability. Its contents include each requirement node, the normative slot of each requirement node, the normative relationship between requirement nodes, the normative weight and activation threshold of each requirement, and the set of indispensable requirement nodes for that crime.

[0146] A normative slot refers to a position in the element template that awaits to be filled by facts. It typically includes a normative predicate slot, a normative subject slot, and a normative object slot. For example, in the element node of intentional injury, the normative predicate slot is "violent means," the normative subject slot is "the actor," and the normative object slot is "the physical health of others."

[0147] Synonymous domains refer to the collection of factual terms that, in legal doctrine, are sufficient to serve as subordinate facts of a particular normative slot. For example, the synonymous domain of the normative slot of violence includes: striking with a weapon, strangling with bare hands, throwing heavy objects, etc., but does not include verbal disputes; the synonymous domain of the normative slot of another person's physical health includes body parts such as the torso, head, and limbs, but does not include property.

[0148] In this embodiment, the synonym field table is organized as a two-level lookup table with standardized slots as keys and factual phrase aggregations as values. Its content is compiled from legal provisions, judicial interpretations, and guiding cases, and is statically configured, not changing with each case. The synonym field table is version-updated as laws are amended, and the version number used for each match is recorded. It should be noted that recording the version number is a necessary condition for auditability. The matching results for the same case may differ after legal amendments. If the version number is not recorded, historical matching conclusions cannot be verified.

[0149] Specifically, semantic similarity cannot serve as a mapping between the factual layer and the requirement layer because it only answers whether something is similar or not, while a canonical mapping answers whether something is qualified or not. The factual state must structurally fill the slots in the requirement template, and this filling must be supported by reinforced evidence. Both are indispensable; therefore, this method constructs the mapping as a bilinear product of the two, which is the key difference between this method and all other embedding retrieval schemes.

[0150] In this embodiment, to ensure that the mapping result of the same piece of evidence data under the same element template is uniquely determined, the specific implementation method of performing canonical causal projection is as follows:

[0151] Establish a one-to-one correspondence between the normative predicate slots, normative subject slots, and normative object slots in the requirement template and the action predicate slots, agent slots, and recipient slots in the fact state subgraph, respectively;

[0152] For each pair of corresponding slots, determine whether the value of the fact slot falls into the synonym domain of the specification slot according to the preset specification-fact synonym domain table.

[0153] The ratio of the number of slot pairs falling into the same synonym domain to the total number of slots in the component template is taken as the slot structure alignment degree.

[0154] Using the normative correlation coefficient of this requirement as the weight, the cross-modal corroboration weights of each activated causal edge participating in the projection are normalized and weighted to obtain the evidence corroboration terms.

[0155] The product of the slot structure alignment and the evidence corroboration item is taken as the standard projection value. The corresponding requirement node is activated only when the value is not less than the requirement activation threshold.

[0156] For example, in this embodiment, the alignment of the slot structure can be calculated using the following formula:

[0157]

[0158] Furthermore, the canonical causal projection operator and the activation criterion for the essential node can be expressed as follows:

[0159]

[0160]

[0161] in, For slot structure alignment; For the first A set of standard slots for each component node. This represents the total number of its slots; For the first One standard slot; For the factual subgraph participating in the projection, its slot values ​​are... It is obtained by merging the slots of each event unit in the subgraph according to topological order; This is a synonym field for the specified slot; To standardize projection values; To standardize causal projection operators; Causal pair requirements The standard relevance coefficient is taken from the component template and belongs to the static configuration quantity; , These are the slot hit indicator vector and the edge verification weight vector, respectively. This is the normalized normalized correlation matrix. Activate the indicator value for the key node; The activation threshold for each requirement is set according to the strictness of the requirement's specifications. For example, the threshold is 0.55 for requirements involving objective harmful behavior and 0.70 for requirements involving subjective knowledge.

[0162] It should be noted that the reason why corroborating evidence must be normalized is that the upper bound of the unnormalized summation increases linearly with the number of participating edges, while the activation threshold is a fixed constant. This would mean that the more evidence there is, the easier it is to exceed the threshold, which is completely contrary to the design intention that the activation of an element should depend on the quality of corroboration rather than the number of evidence. Only when the normalized projected value falls between zero and one can the threshold be comparable across cases.

[0163] Understandably, the normative projection value is a strictly bilinear form, and its two properties provide two structural guarantees: 1) Multiplicative nullification: The structural alignment is zero, meaning the fact has not been filled into the slot of the requirement; or the corroboration weights are all zero, meaning all supporting edges are isolated evidence or have been resolved by contradictions; when either of these occurs, the projection value is always zero, and the requirement node can never be activated. Therefore, the two false activation situations of similar statements but insufficient evidence and sufficient evidence but structural inconsistency are structurally excluded, rather than relying on thresholds to roughly intercept them. 2) Unique and recalcible results: The structural alignment is obtained by looking up the synonym domain, the corroboration weight is given by the deterministic formula in step 3, and the normative correlation matrix is ​​statically configured. None of the three contain randomness or trainable parameters. Therefore, the projection value of the same evidence data under the same requirement template can be reproduced bit by bit, and can be directly used as the object of review and verification. This is the essence of transforming cross-layer mapping from retrieval to conditional judgment, and it is also the core mechanism of this method to eliminate semantic drift and illusion. Figure 7 A comparison diagram is shown between the multiplicative nulling property of the cross-layer mapping and the weighted summation scheme in this embodiment; from Figure 7 As can be seen, when the structural alignment is the same and the corroboration weight is reduced to zero, the mapping value of the multiplicative cross-layer mapping in this scheme is synchronously reduced to zero, while the mapping value of the equal-weighted summation scheme is still as high as 0.40. Moreover, it has already exceeded the requirement activation threshold when the corroboration weight is only 0.30. This proves that the multiplicative zeroing characteristic makes it impossible to activate the requirement node regardless of how closely the expression of isolated evidence or evidence that has been resolved by contradiction is close to the normative concept. The weighted summation structure naturally leaves a channel for such misactivation.

[0164] Furthermore, simply ensuring that each component node is correctly activated is not sufficient. If two valid component nodes can be connected by edges in the component layer, but there is no corresponding causal connection in the fact layer, then the two-layer decoupling will degenerate into two unrelated graphs. The component layer may appear to have a complete structure, but in reality, it is suspended.

[0165] Among them, the inter-layer consistency constraint refers to the constraint that requires each edge in the requirement layer to find a corresponding causal path with the same direction in the fact layer; otherwise, the edge in the requirement layer will not be activated.

[0166] In this embodiment, to ensure that the structure of the component layer is indeed supported by the fact layer, the specific implementation method for injecting inter-layer consistency constraints is as follows:

[0167] For each edge in the component layer, locate the fact state subgraph projected by the component nodes at both ends;

[0168] Retrieve directed paths from the previous subgraph to the next subgraph in the fact layer;

[0169] The requirement layer edge is activated only when there exists such a directed path and the direction of the path is consistent with the direction of the requirement layer edge, and the maximum effectiveness of the path is taken as the implementation effectiveness of the requirement layer edge.

[0170] If no such directed path exists, the edge of the required layer will not be activated, and the missing interlayer structure will be recorded in the source traceability record.

[0171] For example, in this embodiment, the inter-layer consistency constraint can be expressed as follows:

[0172]

[0173]

[0174] in, This is the activation indicator for the element layer edge; To ensure the effectiveness of the edge implementation at the component layer; , These are the key nodes. , The projected factual subgraph; It is the set of all directed paths in the fact layer from any node in the former to any node in the latter; For edge effectiveness, its definition is given in step 5.

[0175] It's important to note that this constraint makes the two-layer topology a true projection relationship rather than two parallel graphs: every structural relationship in the element layer must find a corresponding causal path in the fact layer; otherwise, the structural relationship is invalid. Without this constraint, the algorithm might output seemingly complete but unproven conclusions such as "the objectively harmful act is established," "the constitutive result is established," and "therefore, the attribution relationship between the two is established." This is precisely the part that requires the most proof in attribution judgment and is also the point of greatest contention between the prosecution and the defense.

[0176] Step 5: For each causal path connecting the behavior node and the result node in the candidate requirement subgraph, calculate the danger creation quantity on the actor side, the danger realization degree of the danger on the result node, and the intervention effectiveness diversion ratio of the exogenous intervention node on the path, and perform causal pruning and requirement subgraph reconstruction accordingly to generate the pruned requirement subgraph.

[0177] Among these, "danger creation" refers to whether the actor's conduct creates a danger that is not permitted by law. "Danger realization" refers to whether the danger is realized in the result of the elements constituting the crime. These two steps are the two basic judgments of the objective attribution theory of criminal law, corresponding to the composition of node attributes and the propagation of path effect on the diagram, respectively.

[0178] An exogenous intervention node refers to a causal node located on the causal path from the behavioral node to the result node, but which does not originate from the actor node in the evidence source diagram. Understandably, this definition provides a definite formal standard for exogenous intervention, independent of value judgments in individual cases. The medical institution's treatment behavior node, the victim's self-treatment node, and the third party's additional harm node are all exogenous as they do not originate from the actor; however, the actor's subsequent pursuit and attack node originates from the actor and is therefore not considered an exogenous intervention.

[0179] Intervention effectiveness diversion ratio refers to the share of causal effectiveness contributed by exogenous intervention nodes in the sum of intervention effectiveness and main causal effectiveness.

[0180] This is the key difference between this method and all transitive closure-based causal graphs. It's important to note that traditional methods determine causality on a graph by checking for the existence of a directed path between two nodes, i.e., calculating the transitive closure: if a path exists in the graph—harmful act - injury - hospitalization - death—then a causal relationship between the harmful act and the death is established. However, when the case file contains exogenous interventions such as hospital medication errors, the victim reopening the wound, or further harm by a third party, the legal causal chain is broken. The transitive closure is invisible to these interventions; it only concerns the connectivity of the path, not the allocation of effects along it. The consequence is that all results caused by the intervening factors are attributed to the initial actor, leading to cases that should be considered attempted or without causal relationship being deemed completed.

[0181] Specifically, this method asks a different question: what percentage of the total effectiveness of the pathway to the result is contributed by exogenous intervening nodes? If this percentage exceeds a blocking threshold, the causal chain is deemed interrupted, and the matching result is subsequently reconstructed.

[0182] In this embodiment, in order to make the creation of a calculable hazard that is not permitted by law, the specific calculation method for the hazard creation quantity is as follows:

[0183] Read the action pattern slot of the behavior node and map it to the behavior intensity index according to the preset intensity table;

[0184] The object targeted by the reading action is mapped to the sensitivity of legal interest according to the type of legal interest.

[0185] Calculate the spatiotemporal proximity between the behavior node and the victim node;

[0186] The product of the above three items is taken as the amount of hazard creation.

[0187] For example, in this embodiment, the amount of hazard created can be calculated using the following formula:

[0188]

[0189] in, For behavior nodes The amount of danger created; The intensity index is mapped from the action pattern slot according to the preset intensity table. For example, continuous strikes to vital points with a weapon are assigned 0.9, a single strike to a non-vital point with a weapon is assigned 0.5, and a single push is assigned 0.2. For example, to assess the sensitivity of the legal interests of the affected object, life is assigned 1.0, health 0.7, and property 0.4. , The spatial location and timestamp interval of the victim node; the index term as a whole represents the spatiotemporal proximity.

[0190] Understandably, the reason the danger creation factor is calculated using a product of three factors rather than a weighted sum is that all three conditions are indispensable: No matter how intense the act, if the targeted legal interest is not protected by the crime (the sensitivity of the legal interest approaches zero), or if the act is impossible to reach the victim in time and space (the temporal and spatial proximity approaches zero), it does not constitute a dangerous act that is not permitted by law. A weighted sum would allow a high-intensity act to cross the threshold for criminal liability solely based on its intensity, even if the act is far from the victim in time and space. In practice, this could lead to wielding a weapon in an empty field being considered a dangerous act against a specific victim.

[0191] After confirming that a hazard has been created, it is necessary to determine whether the hazard has been realized in the outcome.

[0192] Here, edge effectiveness refers to the causal transmission strength carried by a certain activated causal edge, which is equal to the product of the cross-modal confirmation weight of the edge and the causal time-distance decay factor.

[0193] The essential result range refers to the set of result types that a danger created by a certain behavior may point to in terms of norms.

[0194] In this embodiment, in order to make the danger a calculable quantity in the constitutive result, the specific calculation method for the degree of danger realization is as follows:

[0195] For each directed path from the behavior node to the result node, take the product of the edge effects of each active causal edge on it;

[0196] Take the maximum value of the product across all paths as the degree of realization of the danger;

[0197] The inspection result node determines whether the passive object falls within the synonymous domain of the legal interest created by the behavioral node, and only when it falls within this domain is the degree of realization included in the attribution of liability.

[0198] For example, in this embodiment, edge effectiveness and danger realization can be calculated using the following formula:

[0199]

[0200]

[0201] in, Serving the border; It is the causal time-distance decay factor; For example, a causal timescale is taken as 3600 seconds; For self-behavior nodes To the result node The degree of danger realized; This is the set of all directed paths between the two. For behavior The dangerous legal interests created within the synonymous domain; The passive object of the result node.

[0202] It should be noted that, since the cross-modal confirmation weight is strictly less than one and the causal time-distance decay factor is no greater than one, the edge power is strictly less than one, and the danger realization degree after path multiplication is also strictly less than one. This boundedness is a prerequisite for setting the thresholds of the subsequent split ratio and hierarchical matching value, so it is explicitly stated here; if the edge power can be greater than one, the longer the path, the greater the power, and the entire attribution calculation will lose its meaning.

[0203] In this embodiment, the solution for the danger realization degree does not require enumerating all paths; it can be completed in one scan using dynamic programming according to the topological order. The specific implementation method is as follows:

[0204] Perform a topological sort on the causal directed acyclic graph of the fact layer;

[0205] Starting from the behavior node, maintain a state variable for each node in topological order. This state variable represents the maximum path efficiency from the behavior node to that node. The behavior node's own state variable is initialized to one, and the state variables of the other nodes are initialized to zero.

[0206] Traverse each node in topological order. For each outgoing edge, update the state of the endpoint of the outgoing edge by the product of the state of the node and the edge strength of the outgoing edge, and take the larger of the updated value and the original value.

[0207] After the traversal is complete, the state of the resulting node is the maximum path efficiency.

[0208] Understandably, the transition equation for the above process is that the final state variable is the maximum value of the product of the states of all its predecessors and the corresponding incoming edges. Its correctness is guaranteed by the topological order, and when a node is traversed, the states of all its predecessors have been finally determined. This process visits each edge only once, so its time complexity is linearly related to the number of edges in the graph. This makes this method feasible even in complex cases with a large factual layer. In contrast, if the method of enumerating subgraphs is used to solve the problem, its complexity will increase exponentially with the number of nodes. Figure 8 This diagram shows a comparison of the increase in access counts for the two methods of solving the danger realization degree in this embodiment; from Figure 8 As can be seen, when the number of fact layer nodes increases to fourteen, the number of edge visits for solving the danger realization degree by dynamic programming according to topological order is 24, while the number of visits for solving by path enumeration is 24576, which is about 1024 times that of the former. Moreover, the gap between the two continues to widen on logarithmic coordinates as the number of nodes increases, making this method still feasible in complex cases with a large fact layer.

[0209] It should be noted that the criterion of the range of the result of the elements is indispensable. If this criterion is omitted, the property damage ultimately caused by the minor injury through a long chain will also be included in the degree of danger of the crime of assault. The path is connected, and the effect is positive, but the result is not within the normative range of the danger created by the act of assault. This criterion truly puts the limitation of realization in the result of the elements of the crime into practice, rather than just an unverifiable statement.

[0210] Understandably, the degree of danger realization is determined by multiplying the effects along the path rather than summing or averaging them. This allows the effectiveness of a long chain of causality to naturally diminish, with the effectiveness multiplied by a factor less than one at each intermediate stage. This is isomorphic to the legal intuition that the equivalence of causality decreases as the number of intermediate stages increases. At the same time, taking the maximum value among all paths means representing the degree of danger realization with the strongest causal path, which aligns with the practical practice of choosing the most equivalence path in liability determination. Figure 9 This diagram illustrates a comparison of the decay characteristics of the degree of danger realization along the causal chain length in this embodiment; from Figure 9 As can be seen, under the premise of the same unilateral efficacy, as the number of intermediate links in the causal chain increases to ten, the attribution efficacy obtained by multiplying the path efficacy decreases to 0.11, which has fallen below the lower limit of attribution efficacy, while the attribution efficacy obtained by summing and averaging is always 0.80, which does not decrease with the chain length at all.

[0211] Furthermore, once both the amount of hazard creation and the degree of hazard realization can be calculated, the intervention effectiveness diversion ratio can be constructed and causal pruning can be performed.

[0212] In this embodiment, in order to identify the interruption of the causal chain and provide graded handling, the specific implementation method of performing causal pruning and component subgraph reconstruction is as follows:

[0213] Identify all exogenous intervention nodes on the path where the danger realization degree reaches its maximum value;

[0214] When there are no exogenous intervention nodes on the path, the split ratio is directly determined to be zero, and the causal chain continues.

[0215] When there are exogenous intervention nodes, noise or aggregation is applied to their incoming edge effectiveness to obtain the intervention effectiveness.

[0216] The ratio of intervention effectiveness to intervention effectiveness plus risk realization is taken as the intervention effectiveness diversion ratio;

[0217] When the shunt ratio is not greater than the blocking threshold, the completed requirement subgraph remains unchanged;

[0218] When the diversion ratio exceeds the blocking threshold but does not exceed the sum of the blocking threshold and the graded bandwidth, and the amount of danger created is not lower than the minimum threshold for criminal liability, the causal path is cut off and the candidate element subgraph is reconstructed into an attempt-type element subgraph.

[0219] When the shunt ratio exceeds the sum of the blocking threshold and the tiered bandwidth, or when the amount of danger created is lower than the threshold for criminalization, the causal path is cut off and reconstructed into a non-causal element subgraph.

[0220] For example, in this embodiment, the set of exogenous intervention nodes, intervention efficacy, and shunt ratio can be expressed by the following formula:

[0221]

[0222]

[0223]

[0224] Furthermore, the causal pruning operator can be represented by the following equation:

[0225]

[0226]

[0227] in, For path The set of exogenous intervention nodes on; The path that maximizes the degree of danger. The effect of the incoming edge on the path of the intervening node; The noise or aggregate value representing the interventional effect; We will continue using the source operator of the evidence source diagram from step 3; For intervention effectiveness diversion ratio; Pruning operator for cause and effect; Subgraph of candidate requirements; The blocking threshold is set to 0.5, which means that the intervention effect accounts for more than half of the total. The bandwidth for the tiered classification between attempted and non-causal types is 0.3, for example. The lower limit for criminalizing the creation of danger is exemplarily set at 0.15; , These are the path deletion operator and the subgraph replacement operator, respectively. , These are the component sub-diagram templates for attempted and non-causal types, respectively.

[0228] It should be noted that the branch in the segmented definition where the shunting ratio is always zero when there are no exogenous intervening nodes is indispensable. If a uniform ratio formula is used, in cases without any intervening factors, the empty product equals one, making the numerator equal to one. Substituting this into the formula, the resulting shunting ratio is not less than one-half. Cases without any intervening factors are thus judged as causal blockages, a flaw that directly leads to a reversal of the conclusion. Similarly, the intervention effect is expressed using noise or aggregation rather than multiplication because multiplication results in more intervening nodes but a smaller numerator, contrary to the intuitive direction that more intervening factors lead to greater interference with the causal chain.

[0229] Understandably, the diversion ratio is a normalized competitive form that combines the intervention effect with the sum of the intervention effect and the main causal effect. The effect is that the stronger the main causal path effect, the smaller the share of the same intervention effect. Therefore, the diversion ratio will be completely different when encountering minor medical defects after serious injury and when encountering major medical accidents after minor injury, rather than being blocked in all cases or not blocked in all cases. This property enables this method to provide a classification judgment consistent with legal doctrine when dealing with high-frequency practical situations such as the victim's special physical condition, third-party negligent intervention, and improper treatment. Figure 10 This diagram shows a comparison of the intervention effectiveness shunting ratio criterion and the transitive closure criterion in causal interruption identification in this embodiment; from Figure 10As can be seen, under the setting of weak main causal power, the intervention power only needs to be 0.30 to make the shunting ratio cross the blocking threshold, while under the setting of strong main causal power, it needs to reach 0.80 to cross the threshold. The difference in the threshold crossing position between the two is significant. In contrast, the equivalent shunting ratio of the transitive closure criterion is always zero in the entire interval and never recognizes causal interruption.

[0230] It should be noted that the output of this step is not simply a matter of causality not being established, but rather a classification and reconstruction of the element subgraphs into attempted or no-causal type based on the values ​​of the diversion ratio and the amount of danger created. This classification is necessary because when the accused's actions have created sufficient danger but the result cannot be attributed to them, the correct legal doctrinal exit is attempted, not innocent. If no classification is used and the output is uniformly invalid, the auxiliary conclusions provided by this method in this situation will systematically favor the accused, thus constituting an error.

[0231] Step 6: Perform cascaded sequential subgraph matching on the pruned requirement subgraphs in the order of dueness, illegality, and culpability. Structural truncation counterfactual intervention is performed on the exclusion subgraphs of the illegality hierarchy. The gating values ​​of each exclusion subgraph are multiplied by the consecutive matching values ​​of the positive requirements in a multiplicative hard-gate manner to obtain the hierarchy matching value.

[0232] The three levels of assessment are as follows: The "legitimacy level" refers to determining whether the active elements of a crime (objective act, object of the act, harmful result, causal relationship, and subjective element) are complete. The "illegality level" refers to determining, assuming legitimacy is established, whether there are grounds for excluding illegality (legitimate defense, emergency avoidance, victim's consent, legal act, etc.). The "culpability level" refers to determining, assuming illegality is established, whether there are grounds for excluding liability (lack of capacity for responsibility, unavoidable mistake of understanding of illegality, lack of expectation of possibility, etc.). These three levels form a hierarchical progression; if a previous level is not established, subsequent levels do not need to be assessed.

[0233] It's important to note that the traditional flat weighted summation approach encounters a fundamental obstacle here. The judgment of legal elements is structurally cascading and contains hard veto points: even if both the objective act and the harmful result are met, the overall requirement is not met if there are grounds for excluding illegality such as legitimate defense, emergency avoidance, or victim's consent. Weighted summation, however, is a continuous, monotonous, and compensable scalar operation: even if the grounds for excluding illegality take the most unfavorable value, it only reduces its weight in the total score, which can be fully compensated by the high scores of the other items. The consequence is that hard legal prohibitions are smoothed out by probability, resulting in a 0.93% compliance rate, a 0.31% illegality rate, and a combined score of 0.78%, leading to a conclusion that is fundamentally untenable in legal dogma—a conclusion that is fundamentally flawed.

[0234] Specifically, this method adopts different mathematical forms for the three levels: the dueness level retains a continuously optimizable form because the same fact may simultaneously constitute the elements of several crimes, requiring sorting and selection; while the illegality and culpability levels adopt gate values ​​of zero or one, and intervene in the objective function in a multiplicative manner.

[0235] In this embodiment, in order to achieve both optimizability and non-boundary compliance within the same expression, the specific implementation of constructing the equivalent hierarchical continuous matching value is as follows:

[0236] For each node in the target element subgraph, the normative projection values ​​of the activated nodes are weighted and summed using their normative weights as coefficients.

[0237] For each element layer edge in the target element subgraph, the realization effectiveness of the edge is calculated with the structural term coefficient as the weight only when both of its two endpoints are activated and the inter-layer consistency constraint is met.

[0238] Normalization is performed using the sum of the normalized weights and the sum of the structural terms as the denominator;

[0239] For the set of essential nodes of the crime, take the product of their activation indicators as a local hard gate, and multiply it with the normalization result above.

[0240] For example, in this embodiment, the successive matching value of the dueness hierarchy can be calculated using the following formula:

[0241]

[0242]

[0243] in, For each consecutive matching value of the appropriate hierarchy; This is a diagram of the components required after pruning; For target component sub-diagram template; , These are the set of its component nodes and the set of its component layer edges, respectively. The standard weight of the essential node; The coefficient for the structure term is 0.4, for example. This is the set of essential elements of the crime, such as the aggravated result node for a crime with aggravated consequences, or the specific identity node for a crime with a specific identity.

[0244] Understandably, the introduction of structural terms allows the matching value to distinguish between two situations that are indistinguishable under flat scoring: each element is matched individually but is factually unrelated to the others, and each element is matched while the normative relationship between the elements is also realized by evidence. The former merely puts together several unrelated facts, such as someone did possess a weapon or a death did occur somewhere, but there is no causal relationship between the two; the latter is the true fulfillment of the elements of the crime.

[0245] It should be noted that the essential requirement local hard gate of the last term is the first hard constraint embedded in the continuum: once the essential requirement is not activated, the entire appropriateness matching value is always zero and is not compensated by the other high-scoring requirements; while the rest remains continuous, allowing multiple candidate crimes to still be ranked. In this way, optimizability and non-boundary inaccessibility are obtained simultaneously in the same expression, rather than having to choose between the two.

[0246] Furthermore, the determination of grounds for excluding unlawfulness is itself counterfactual in legal theory: the question is not how much harm the defensive act caused, but whether the defender would still have caused the same harm to legal interests if there had been no unlawful infringement.

[0247] Among them, structural truncation counterfactual intervention refers to a deterministic operation that sets the infringement intensity of the unlawful infringement node to zero on the element subgraph, cuts off all its incoming edges, and then recalculates the amount of legal interest damage according to topological order.

[0248] The limit contribution rate refers to the ratio of the difference between the original amount of damage to the legal interest and the amount of damage to the legal interest obtained from the counterfactual recalculation to the original amount of damage to the legal interest. In essence, it means: by what percentage the amount of damage decreased after the unlawful infringement was stopped; the more thorough the decrease, the more the damage was caused by the infringement itself, and the more the defensive action was within limits.

[0249] In this embodiment, to transform the defense limit from a vague value judgment into a recalculated causal contribution rate, the specific implementation method of structurally truncated counterfactual intervention is as follows:

[0250] Locate the nodes of illegal infringement in the sub-graph after pruning;

[0251] Set the attack intensity of the node to zero and cut off all its incoming edges to obtain a counterfactual truncation graph;

[0252] The amount of legal interest damage caused by the defender node is recalculated along the counterfactual truncation diagram in topological order;

[0253] The ratio of the difference between the original amount of legal interest damage and the recalculated amount of legal interest damage to the original amount of legal interest damage is taken as the limit contribution rate;

[0254] When the limit contribution rate is not less than the limit tolerance, the gate value of the rejection subgraph is determined to be one, that is, rejection is established; when the limit contribution rate is less than the limit tolerance, the gate value is determined to be zero, and the process switches to the over-defense branch.

[0255] For example, in this embodiment, the structurally truncated counterfactual intervention, the amount of damage to legal interests, and the contribution rate to the limit can be expressed by the following formula:

[0256]

[0257]

[0258]

[0259]

[0260] in, This is a counterfactual truncation diagram; For structure truncation operators; This is a node for illegal infringement; For the defense of human nodes; The amount of damage to legally protected interests; This is the set of result nodes in the graph; The strongest path from the defender node to the result node; To the limit of contribution rate; For the sake of limitation, an example value of 0.6 is used; For the first A threshold value for a reason to preclude illegal activities.

[0261] Understandably, the structural truncation operator not only sets the attack strength to zero but also cuts off all incoming edges of the node, as this is a necessary condition for thorough intervention. If only the strength is set to zero while the incoming edges are retained, the attack strength may be indirectly reconstructed through propagation from upstream nodes during topological recalculation. For example, an upstream armed approaching node may still inject power into the attacking node, rendering the intervention ineffective. The contribution rate will be systematically underestimated, causing what should be legitimate self-defense to be judged as excessive self-defense. This is an easily overlooked but seriously problematic implementation detail. Figure 11 This diagram illustrates the impact of the thoroughness of counterfactual truncation on the limit contribution rate in this embodiment; from Figure 11As can be seen, when the upstream injection intensity of the infringing node gradually increases, the limit contribution rate obtained by setting it to zero and cutting off the incoming edge is always 0.78, which is always above the limit tolerance. However, the rate of the rate of the rate of the rate of the infringing node that is only set to zero and the incoming edge is retained decreases to 0.32 as the upstream injection intensity increases. It falls below the limit tolerance when the upstream injection intensity is about 0.4. This proves the decisive role of cutting off the incoming edge in the thoroughness of the intervention. If this operation is omitted, the intensity of the infringing node will be indirectly reconstructed through the upstream node, and the legitimate defense that should be established will be systematically misjudged as excessive defense.

[0262] It should be noted that this method uses deterministic structural truncation rather than probabilistic intervention estimation, which is not only sufficient but also more preferable in judicial assistance scenarios: probabilistic forms require the introduction of unobserved confounding assumptions and rely on identifiability conditions, and their conclusions cannot explain to the adjudicator how the number was calculated; while each step of structural truncation is an explicit operation on the graph, and the recalculation result can be reproduced step by step by any third party on the same piece of evidence data, which is truly verifiable.

[0263] Furthermore, once the gating values ​​for each blocking event have been determined, the hierarchical matching value can be synthesized.

[0264] In this embodiment, the method for determining the threshold value of the grounds for exemption from liability is as follows: For lack of liability capacity, the threshold value is determined by the projection value of the elements corresponding to the assessment opinions on age and mental state at the time of the act; if the projection value exceeds the activation threshold, the threshold value is determined to be one. For unavoidable mistake of awareness of illegality, the threshold value is determined by the accessibility of the actor to relevant regulatory information; if all accessible evidence is missing and the actor has fulfilled the obligation to conduct reasonable inquiries, the threshold value is determined to be one. For lack of expectation possibility, the threshold value is determined by whether the space of selectable actions at the time of the act is empty; if there are no reachable nodes pointing to legal alternative actions in the factual layer, the threshold value is determined to be one.

[0265] For example, in this embodiment, the multiplicative hard-gate synthesis of the hierarchical matching value can be represented by the following formula:

[0266]

[0267] in, For hierarchical matching values; A collection of grounds for excluding illegal acts; A collection of reasons for excluding responsibility; This is the gate value for the responsible level.

[0268] It should be noted that the multiplicative structure of this formula provides a property that the weighted summation structure cannot provide in principle: when any rejection gate value is one, the hierarchical matching value is always equal to zero and is completely independent of the due diligence score. In the weighted summation, the extreme value of any term can be compensated by the remaining terms; while in the product, a zero factor makes the entire product zero, and there is no compensation path. Thus, the veto has been transformed from a normative statement into an algebraic identity, which is a faithful implementation of the hierarchical structure of legal dogma at the mathematical level.

[0269] Furthermore, cascading is essentially short-circuiting. If each of the three levels calculates separately and then multiplies, the algorithm is still formally performing parallel scoring, which wastes computing power and loses the interpretability crucial to judicial scenarios regarding which level was rejected and why.

[0270] In this embodiment, to ensure true sequentiality and preserve interpretability of the cascade, the specific implementation of hierarchical short-circuit evaluation and source tracing is as follows:

[0271] Calculate the value in the order of appropriateness, illegality, and culpability;

[0272] When the gating composite value of any level reaches zero, the calculation of subsequent levels is immediately terminated;

[0273] Push the identifier of this layer, the name of the criterion that causes it to be zeroed, the measured value of the criterion, and the threshold value of the criterion onto the traceability stack;

[0274] When outputting, the top content of the trace stack is used as the reason for rejection, and the remaining content is used as the complete judgment trajectory.

[0275] For example, in this embodiment, the evaluation of hierarchical short-circuit can be expressed as follows:

[0276]

[0277]

[0278] in, We take one, two, and three in sequence to represent the three levels of due process, illegality, and responsibility; This is the gating synthesis value for this level; The evaluation function for the hierarchical level.

[0279] Understandably, short-circuit evaluation not only saves computational power, but more importantly, it inherently provides a conclusion on why the statement is invalid. Traditional solutions output a comprehensive score, and the judge has no way of knowing whether the low score is due to insufficient objective behavior or the existence of a precondition. In contrast, short-circuit evaluation outputs a conclusion that can be directly challenged: the illegality level is rejected by the self-defense subgraph, the contribution rate is limited to 0.72, and the tolerance for exceeding the limit is 0.6.

[0280] Step 7: Accumulate the log-likelihood ratios corresponding to the evidence edges for pre-behavioral preparation, behavioral action form, and post-behavioral treatment to obtain the subjective cognitive intensity. Perform a resonance verification of the consistency between subjective and objective causality in the three dimensions of object, means, and result. If necessary, route the matching process to the subgraph of the causal cognition error branch.

[0281] It should be noted that the most difficult element in the legal constitution is the subjective element, including intent, negligence, intent to illegally possess, and malicious collusion. These mental states cannot be directly observed. Existing solutions either rely on self-admission keywords in the transcript, such as "I wanted to kill him," or treat the subjective element as a general point in the overall scoring. The former becomes ineffective when the accused recants or remains silent, while the latter equates the objective outcome directly with the subjective intent when the objective result deviates from the actor's perception.

[0282] Among these, evidence edges related to pre-action preparation refer to causal edges where the timestamp interval is earlier than the action node and the action predicate falls within the synonym domain of preparation, such as tool purchase, route reconnaissance, information retrieval, and victim's daily routine investigation. Evidence edges related to the form of action during the action refer to causal edges where the timestamp interval coincides with the action node and describes the specific form of the action, such as the location of the blow, the number of blows, the intensity of the blow, and the duration. Evidence edges related to post-action handling refer to causal edges where the timestamp interval is later than the action node, such as trace erasure, transfer of property, escape route, and whether rescue was provided.

[0283] Subjective cognitive intensity refers to a scalar measure, obtained by summing the log-likelihood ratios of the three types of objective evidence mentioned above, which represents the degree to which the actor's subjective state of mind tends to be intentional.

[0284] Specifically, this method does not train any classifier, but instead adopts the analytical structure of summing the log-likelihood ratios of evidence: each type of objective evidence edge has different conditional occurrence rates under the two competing hypotheses of intent and innocence, and the logarithm of their ratios is the contribution of that evidence edge to the subjective determination, which can be directly added together.

[0285] In this embodiment, the lookup values ​​for condition occurrence rates are obtained by statistically analyzing the frequency of each action predicate in intentional and negligent cases in the fact-finding sections of effective judgments. These values ​​are categorized by case type and updated offline periodically. It should be noted that this offline updating does not affect the certainty of a single match; under the same version of the lookup value, the same piece of evidence will necessarily yield the same subjective cognitive intensity. Each match also records the lookup version number used.

[0286] In this embodiment, in order to ensure that the determination of subjective elements can still yield a traceable conclusion even when the accused remains silent, the specific implementation method for performing subjective causal inference is as follows:

[0287] Based on the chronological relationship between the timestamp intervals of each evidence edge and the timestamp intervals of the action node, the evidence edges are categorized into three time periods: before the action, during the action, and after the action.

[0288] For each piece of evidence, find the condition occurrence rate of the predicate under the intentional and unintentional assumptions according to its action predicate, and take the natural logarithm of the ratio between the two.

[0289] The logarithmic values ​​above are weighted and accumulated using the cross-modal verification weight and the homology reduction factor of that edge as weights;

[0290] The intensity of subjective cognition is obtained by superimposing prior log odds based on case type;

[0291] The number of time periods in which evidence edges actually exist in the three time periods is used as the evidence completeness, and the subjective cognitive intensity is subjected to zero contraction according to the completeness.

[0292] When the corrected subjective cognitive intensity is not less than the threshold for the establishment of subjective elements, the subjective elements are determined to be established.

[0293] For example, in this embodiment, the sum of the log-likelihood ratios over the three time periods and the intensity of subjective cognition can be calculated using the following formula:

[0294]

[0295]

[0296]

[0297] Therefore, the criteria for the correction of the completeness of evidence and the establishment of the subjective element can be expressed as follows:

[0298]

[0299]

[0300] in, Subjective cognitive intensity; , , These are the cumulative log-likelihood ratios for the three time periods; , , These are the evidence edge sets for three different time periods; , These represent the conditional occurrence rates of the action predicate under the intentional and unintentional assumptions, respectively. To verify weights across modalities; It is the homology reduction factor; It is the union of the source sets of other supporting evidence from the same period; Prior probabilities determined by case type; The completeness of evidence is rated as one, two, or three. The contraction index is taken as 1.5 for example; The corrected subjective cognitive intensity; The subjective element establishes the indicator quantity; As a threshold for the establishment of subjective requirements, we take the natural logarithm of nineteen, which is approximately 2.94, corresponding to a posterior probability of nineteen to one.

[0301] It should be noted that the introduction of the source reduction factor is to eliminate the conditional independence assumption implicit in directly adding the log-likelihood ratios. Directly adding the likelihood ratios of the three time-period evidence edges is equivalent to assuming that each evidence edge is conditionally independent under a given subjective assumption. However, the pre-existing evidence of purchasing the murder weapon and searching for the crime method is clearly highly correlated, often stemming from the same premeditation, the same equipment, and the same time period. Directly adding them would double-count the same psychological fact, leading to a systematic overestimation of the intensity of subjective cognition. The reduction factor is harmonicly attenuated based on the number of overlapping sources, ensuring that the total contribution of the highly correlated evidence combination is close to the strongest one, rather than the sum of all the others. This mechanism shares the same set of source set operations with the public pollution source criterion in step 3; both are applications of the same principle in two different locations.

[0302] Understandably, using the log-likelihood ratio instead of weighted scoring has three advantages. First, the contributions of evidence from different time periods are both additive and divisible, allowing the impact of individual pieces of evidence to be audited and examined separately. The defense can argue that the conditional occurrence rate of a particular piece of evidence is inappropriate without challenging the entire model. Second, there is a strict analytical correspondence between the intensity of subjective cognition and the posterior probability—that is, the posterior probability equals the exponent of that intensity. Therefore, the threshold has a clear meaning of a standard of proof, rather than a predetermined score line. Third, all inputs are objective evidence, allowing conclusions to be drawn even when the accused remains silent or recants, without relying on admissions.

[0303] It should be noted that the evidence completeness correction is to prevent the determination of subjective intent based on isolated evidence. For example, if the case file only contains the escape route after the act, but there is no evidence of preparation before the act, and the form of the actions during the act is unclear, then if the likelihood ratio of the escape claim is large enough, the uncorrected subjective cognitive strength can still exceed the threshold. This essentially determines subjective intent based on indirect evidence from a single period, which contradicts the spirit of reinforcement established in step 3. The contraction factor adopts a power form of the completeness ratio, so that the contraction is significant when there is evidence for only one period, moderate when there is evidence for two periods, and no contraction when all three periods are complete. This is consistent with the practical experience that subjective intent can only be inferred when evidence from three periods corroborates each other, while avoiding a direct zeroing out with a rigid gate, since extremely strong evidence from a single period is indeed more probative than no evidence at all. Figure 12This diagram illustrates the effect of evidence completeness correction on subjective cognitive strength in this embodiment; from Figure 12 As can be seen, in the case of evidence from only a single time period, the uncorrected subjective cognitive strength is 3.60, which has exceeded the threshold for establishing the subjective element and constitutes a factual isolated evidence. After completeness correction, it drops to 0.69 and no longer exceeds the threshold. However, in the case of evidence from all three time periods, the corrected strength is still 3.60, and the determination is not affected. This proves that completeness correction accurately suppresses the threshold-crossing ability of evidence from a single time period in a power-law contraction manner, while not compromising the completeness of the three time periods that corroborate each other.

[0304] Furthermore, subjective cognitive strength alone is insufficient. If there is a discrepancy between subjective cognition and the objectively occurring harm, directly substituting the objective result for the subjective intent will result in a systemic error in attribution of responsibility.

[0305] Among them, the consistency resonance verification of subjective and objective causality refers to the operation of comparing the subjective cognitive map obtained by back-inferring the subjective cognitive intensity with the actual legal interest infringement map in terms of object, means, and result to see if their slot values ​​fall into the same synonym domain.

[0306] Misalignment refers to the indicator value of a slot in the subjective cognitive map on a certain dimension that does not fall into the synonym domain of the corresponding slot in the objective infringement map. A value of zero indicates that the dimension is consistent, and a value of one indicates that the dimension is misaligned.

[0307] Error in causal perception refers to a situation where the causal process perceived by the actor is inconsistent with the actual causal process. Based on the dimension of the error, it can be divided into: object-dimensional error (e.g., mistaking A for B and committing harm); means-dimensional error (e.g., intending to strike A but actually striking B); and result-dimensional error (e.g., intending to drown but actually causing B to fall off a bridge and die).

[0308] In this embodiment, to avoid directly and crudely equating objective results with subjective intent, the specific implementation method for performing the subjective-objective consistency resonance verification is as follows:

[0309] Based on the evidence edges corresponding to the intensity of subjective cognition at each time period, the slot values ​​of the subjective cognition map in the three dimensions of object, means, and result are merged.

[0310] The slot values ​​of the same three dimensions are read from the actual occurrence of the legal interest infringement map;

[0311] For each dimension, determine whether the former falls into the synonym domain of the latter, and take its complement as the misalignment degree of that dimension;

[0312] When the misalignment in all three dimensions does not exceed the misalignment tolerance, the resonance is deemed to have passed, and the matching process continues along the original component subgraph.

[0313] When the misalignment degree of any dimension exceeds the misalignment tolerance, the dimension with the largest misalignment degree is selected, and the matching process is routed to the causal recognition error branch component subgraph corresponding to that dimension.

[0314] For example, in this embodiment, the subjective-objective consistency resonance verification and cognitive error diversion can be expressed by the following formula:

[0315]

[0316]

[0317] in, For dimension The degree of misalignment on; The slot value in this dimension of the subjective cognitive map represents the object, means, or result perceived by the actor. The values ​​of the objective legal interest infringement map on the same dimension; To recognize the error-diverting operator; As a misalignment tolerance, it is set to zero for example, meaning that a misalignment in any dimension will trigger a traffic split. The dimension with the greatest misalignment; This is a template for the sub-diagram of causal cognitive error elements in the corresponding dimension.

[0318] It should be noted that the reason why the resonance check reuses the synonymous domain operator in step 4 instead of setting up a separate set of criteria is to ensure that the criteria for determining conformity are completely consistent with the criteria for determining specification mapping. If different criteria are used in the two places, it is possible that a certain fact is sufficient to fill a specification slot, but in the resonance check, it is judged as misaligned with that slot, resulting in a self-contradictory outcome. Such contradictions are extremely difficult to detect in practice because the two are located at different points in the process.

[0319] Understandably, this step establishes a mandatory diversion exit: once subjective perception and objective harm do not fall into the same synonym domain in either dimension, the matching process must not continue along the original requirement subgraph and must be routed to the error of perception branch. This structurally prevents the systemic attribution error of directly substituting objectively occurring results for subjective intent, an error that is unavoidable in a flat scoring scheme because a high score for objective results itself contributes to the score for subjective elements.

[0320] Step 8: Calculate in parallel the probative scores of the pruned subgraph and the target subgraph, as well as the explanatory scores of each preset alternative causal hypothesis subgraph. Compare the difference between the two with the exclusion threshold, and perform a closure check on the alternative hypothesis set to output the verification conclusion.

[0321] It should be noted that a high matching score does not equate to meeting the standard of proof, because the standard is exclusive. The judicial standard of proof requires that criminal cases be beyond a reasonable doubt and civil cases be of high probability. However, traditional methods only score the prosecution's target subgraph and never calculate the explanatory power of the defense's alternative causal hypothesis in parallel. Therefore, there is no correspondence between the output score and the standard of proof: a case supported by only a single piece of evidence and with unexcluded reasonable alternative explanations can still receive a high score of 0.9.

[0322] Among them, the alternative causal hypothesis subgraph refers to the element subgraph corresponding to other causal hypotheses that can explain the same set of objective results, in addition to the causal relationship claimed by the prosecution, such as accidental events, suicide, third-party crimes, and the natural development of pre-existing diseases.

[0323] The probative score refers to the score of the target element subgraph in terms of hierarchical matching, culpability form, and key edge reinforcement.

[0324] The explanatory power score refers to the maximum matching score that a given alternative causal hypothesis subgraph can obtain under the same set of projection operators and edge effects.

[0325] The exclusion threshold refers to the lower limit that the difference between the probative value score and the maximum value among all explanatory values ​​must reach. It is divided into two levels according to the nature of the case: criminal and civil, corresponding to the standards of proof of beyond reasonable doubt and high probability, respectively.

[0326] In this embodiment, the alternative causal hypothesis subgraph is obtained from the union of three paths: first, it is formalized from the defense arguments explicitly stated in the defense opinion; second, it is retrieved from a library of common defense types corresponding to the case type; and third, it is automatically generated from a strong causal path in the factual layer that is not used by the target subgraph. That is, when there is a strong causal path in the factual layer leading to the same result node but not passing through the defendant's behavior node, it is automatically formalized into an alternative hypothesis. It should be noted that the third path is particularly important: it enables the algorithm to discover alternative interpretations that the defense has not yet raised but that are supported by the evidence itself, thereby avoiding the validity of the exclusive conclusion being constrained by the quality of the defense.

[0327] In this embodiment, to change the output condition from an absolute score to a relative advantage, the specific implementation method for performing exclusive verification is as follows:

[0328] The probative score of the target subgraph is calculated by multiplying the hierarchical matching value, the value of the culpability form selector, and the key edge reinforcement coverage.

[0329] For each alternative causal hypothesis, the same canonical causal projection operator and the same set of edge efficiencies are reused to search the fact subgraph that maximizes the projection value on the fact layer, thus obtaining the explanatory power score of the hypothesis.

[0330] The difference between the maximum score of the objective subgraph's proof power and the maximum score of all explanatory power is taken as the exclusivity difference.

[0331] Select an exclusion threshold based on the nature of the case. When the exclusion difference is not less than the threshold, output a subgraph matching verification report with edge-by-edge tracing information; otherwise, output a report of insufficient probative value, along with a list of alternative hypotheses that have not been excluded.

[0332] For example, in this embodiment, the proof strength score of the target subgraph can be calculated using the following formula:

[0333]

[0334] Furthermore, the explanatory power of alternative hypotheses and the criterion for exclusive output can be expressed as follows:

[0335]

[0336]

[0337] in, Score the proof power of the target subgraph; For hierarchical matching values; the part in parentheses is the sin form selector; The amount of care required for a negligent act is determined by the projection value of whether the actor had the possibility of foreseeing and avoiding the act at the time of the act. The set of key causal edges on which the target element subgraph depends, i.e., those with a normative correlation coefficient higher than the criticality lower limit; the last term represents the critical edge reinforcement coverage rate. For the first The explanatory power score of each alternative hypothesis; , These are the slot set and hypothesis subgraph for the hypothesis, respectively. To replace the total number of hypotheses; This represents the exclusivity difference. For example, the exclusion threshold is 0.35 for criminal cases and 0.10 for civil cases.

[0338] It should be noted that the introduction of the culpability form selector is necessary. If the intensity of intent and subjective cognition were always used as the multiplicative factor, the probative value of negligent crimes and aggravated offenses would always be zero. Negligent crimes do not require intent and subjective cognition; their subjective element is the violation of the duty of care. The culpability form comes from the static attributes of the target element template, so the branches of this selector are determined before matching begins, without introducing runtime uncertainty.

[0339] Understandably, the explanatory power of alternative hypotheses must reuse the same projection operator and the same set of edge effects, only changing the target template, in order to ensure that both the prosecution and the defense are evaluated under the same metric. If the prosecution's score and the defense's score are calculated using different methods, then the difference between the two is meaningless, and the exclusivity metric becomes invalid.

[0340] It's important to note that the essence of this step is changing the output condition from an absolute score to relative advantage: even if the probative value score of the target subgraph is as high as 0.9, if there is an alternative hypothesis with an explanatory power of 0.8 that has not been excluded, the difference is only 0.1, which is still insufficient to pass under the criminal standard. The system will output "insufficient probative value" along with a list of the excluded hypotheses. Conversely, if the target subgraph score is moderate and the explanatory power of all alternative hypotheses is very low, it can pass. This establishes a direct correspondence between the algorithm output and the normative connotation of "beyond a reasonable doubt" with an adjustable threshold for the first time; and since criminal and civil cases use two threshold levels, the same method can cover both standards of proof without the need for separate modeling for the two types of cases.

[0341] Furthermore, the validity of an exclusivity measure depends on whether the set of alternative hypotheses is exhaustive. If only two easily excluded hypotheses are listed, even a large difference will not eliminate reasonable doubt, because genuine reasonable doubt may not have been included in the comparison at all.

[0342] Among them, the closed-loop check refers to the examination of whether the set of pre-set alternative hypotheses covers three mutually exclusive categories: no actor involved, actor involved but not caused by the act, and victim's own reasons.

[0343] In this embodiment, to prevent false exclusivity caused by a non-closed set of hypotheses, the specific implementation of the closure check is as follows:

[0344] The pre-defined substitution hypothesis is categorized into the three mutually exclusive categories mentioned above;

[0345] Count the number of hypotheses actually involved in the calculation for each category;

[0346] When the number of hypotheses in any category is zero, the output verification report will indicate the categories where the alternative hypothesis set is not closed or missing, and the conclusion level of this output will be downgraded to preliminary verification passed.

[0347] For example, in this embodiment, the closure indicator and the conclusion level can be represented by the following formula:

[0348]

[0349]

[0350] in, It is a closed-ended indicator; , , These represent three categories in turn: no perpetrator involved (accidental events, natural events, victim suicide), perpetrator involved but not caused by the act itself (caused by intervening factors, caused by other crimes), and victim's own reasons (special physical condition, natural progression of pre-existing diseases, self-inflicted risk). This is at the conclusion level.

[0351] Understandably, the closure check does not change the exclusivity difference itself, but rather the level of the conclusion. The intention of this design is that the algorithm should not reject output due to an incomplete set of hypotheses, but must let the user know the extent of its exclusivity conclusion's applicability—this is precisely the essence of auditability. If all outputs are rejected, this method will be completely unusable in cases where defense arguments have not yet been presented or the defense type library is incomplete; if all outputs are accepted as valid, its exclusivity conclusion risks being overestimated.

[0352] Step 9: Simultaneously construct the edge-by-edge tracing record during the calculation process of the above steps, and output a structured judicial verification report carrying the complete judgment trajectory and evidence tracing chain.

[0353] Among them, the edge-by-edge traceability record refers to the record that is written synchronously when each intermediate quantity is calculated, which includes its input quantity identifier, the name of the criterion used, the parameter value and the threshold.

[0354] The source closure refers to the union of the source closure of a certain intermediate quantity with the source closure of all its input quantities, that is, the complete source chain that is recursively expanded to the leaf node (event unit and original evidence material).

[0355] In this embodiment, the source tracing record is organized hierarchically as event unit → causal edge → corroboration weight → element projection → hierarchical gating → exclusivity difference, and stored in the form of a directed acyclic graph. During output, it is recursively expanded into a derivation tree based on the intermediate values ​​queried, and supports reverse lookups based on which conclusions a particular piece of evidence affects. It should be noted that reverse lookup capability is indispensable in practice: the defense often needs to argue for the exclusion of a single piece of evidence. In this case, it is essential to be able to ascertain which conclusions are affected and to what extent after excluding the evidence; otherwise, it is impossible to determine whether the exclusion claim has substantial significance for the final conclusion.

[0356] For example, in this embodiment, the structure of the source tracing record and the source tracing closure can be shown as follows:

[0357]

[0358]

[0359] in, For any intermediate quantity; The criterion or operator name for generating this quantity; Its input set, Its number; The set of thresholds used for this criterion. Its number; This is the measured value of this quantity; For tracing the closure.

[0360] It should be noted that the recursive definition of source closure ensures that any output conclusion can be fully expanded into a derivation tree with the original evidence materials as leaves. In judicial scenarios, this is not an optional additional function, but a necessary condition. A verification report that cannot explain in detail which evidence, which criteria, and which thresholds the conclusion is based on lacks verifiability in court and its probative value is close to zero.

[0361] After completing the above nine steps, this method ultimately outputs a structured judicial verification report. The report includes: the activation status and normative projection values ​​of each element node; the determination results of causal pruning and the corresponding intervention effectiveness diversion ratio; the judgment trajectory of the three levels, including the identifier of the rejected level, the criterion name, the measured value, and the threshold; the subjective cognitive strength, the completeness of evidence, and the results of subjective-objective resonance verification; the exclusivity difference, the explanatory power score of all alternative hypotheses, and the list of hypotheses not excluded; the conclusion level; and a complete tracing chain from the conclusion back to the specific evidence materials, along with the version number of the synonym table and the condition occurrence rate lookup table used in this matching.

[0362] It should be noted that this report is intended as supplementary material for adjudicators' review, not as a legal conclusion itself. Every value in the report can be independently recalculated by a third party using the same set of evidence data, and every threshold can be individually questioned and adjusted; this is the fundamental difference between this method and the traditional approach of providing a composite score and claiming the elements are met. This report can be directly used for downstream tasks such as self-examination of the evidence system during the prosecution stage, analysis of points of contention during the trial stage, and comparison of constituent elements in similar case searches.

[0363] For example, the values ​​of each exemplary parameter in this embodiment can be summarized as follows: modal reliability is 0.30 to 0.95 depending on the type of evidence; temporal identity threshold is 0.5; spatial identity distance threshold is 2 meters in indoor scenarios; overlap tolerance is 0.2; maximum displacement rate of the scene is 3 meters per second in indoor walking scenarios and 30 meters per second in outdoor scenarios with vehicles; spatial position error tolerance is 1.5 meters; spatiotemporal tolerance scale is 300 seconds and 5 meters respectively; reliability resolution threshold is 0.15; contradiction judgment threshold is 1.0; logical repulsion penalty value is 1. 0; Activation threshold for elements is set at 0.55 to 0.70 depending on the type of element; Behavioral intensity index is set at 0.2 to 0.9 depending on the action form; Sensitivity to legal interests is set at 0.4 to 1.0 depending on the type of legal interest; Causal time interval scale is set at 3600 seconds; Blocking threshold is set at 0.5; Tiered bandwidth is set at 0.3; Lower limit for criminal liability based on the amount of danger created is set at 0.15; Structural coefficient is set at 0.4; Tolerance limit is set at 0.6; Threshold for establishing subjective elements is set at the natural logarithm of 19; Contraction index is set at 1.5; Misalignment tolerance is set at zero; Exclusion threshold is set at 0.35 in criminal cases and 0.10 in civil cases. It should be noted that the above values ​​are all exemplary and can be adjusted accordingly based on the type of case, regional judicial practice, and applicable standard of proof; The adjusted values ​​are also recorded in the source record to ensure the recalculation of the conclusion.

[0364] Example 2

[0365] This embodiment discloses a subgraph matching and verification system for legal constituent elements based on a multimodal causal knowledge graph.

[0366] Specifically, the legal elements subgraph matching and verification system based on multimodal causal knowledge graph can be integrated into an electronic device, such as a terminal or server. The terminal can be a workstation, a case assistance terminal, a laptop, or a personal computer; the server can be a single server, a server cluster consisting of multiple servers, a private computing node deployed within a law enforcement intranet, or a distributed computing cluster or judicial data service platform with graph database support capabilities. When the electronic device is running, it can implement the legal elements subgraph matching and verification method based on multimodal causal knowledge graph in Embodiment 1 of this application.

[0367] In this embodiment, the legal constituent element subgraph matching and verification system based on multimodal causal knowledge graph can also be integrated into multiple electronic devices. For example, the system can be integrated into multiple servers, and the multiple servers can work together to implement the legal constituent element subgraph matching and verification method based on multimodal causal knowledge graph in Embodiment 1 of this application.

[0368] In this embodiment, considering the confidentiality requirements of judicial data, all modules of the system can also be deployed in a physically isolated stand-alone environment. Since all the criteria of this method are deterministic algebraic expressions and graph theory algorithms, and do not depend on any external model services that require network access, its offline performance is consistent with that in the network environment.

[0369] In this embodiment, the server can also be implemented in the form of a terminal.

[0370] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs, characterized in that, The method for matching and verifying the subgraphs of legal constituent elements includes: Acquire multi-source evidence data of the case to be verified, and extract multiple event units from the multi-source evidence data. The event units include time information, spatial information, and semantic slots. When two event units satisfy the temporal precondition and spatiotemporal reachability condition, a causal edge is established from the preceding event unit to the following event unit to obtain a fact causal graph. Based on whether the supporting sources of the causal edge are of the same origin in the evidence derivation relationship, the number of mutually independent supporting sources is determined, and the corroboration weight of the causal edge is determined accordingly. The corroboration weight of a causal edge with fewer than two mutually independent supporting sources is zero. Determine the slot alignment between the canonical slot of the component node in the component template and the semantic slot of the corresponding fact subgraph in the fact causal graph. Determine the projection value of the component node based on the product of the slot alignment and the confirmation weight of the causal edge in the fact subgraph. Activate the component node when the projection value is not less than the activation threshold to obtain a candidate component subgraph. The candidate component subgraph is matched with the target component subgraph, and a verification conclusion is output based on the matching result.

2. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, The multi-source evidence data includes evidence materials in at least two modalities among video, audio, images, text, physical evidence, and electronic data; The time information is a timestamp interval defined by the earliest start point and the latest end point; the spatial information is spatial location coordinates under a unified spatial coordinate system; and the semantic slots include action predicate slots, agent slots, and recipient slots whose values ​​are taken from the fact vocabulary table. The event unit also includes a modal identifier and a modal reliability, the modal reliability being determined based on the legal probative value of the evidence type and the compliance of the evidence collection procedure; In the event unit, the event unit corresponding to the behavior performed by the actor is the behavior node, and the event unit corresponding to the result of the infringement of legal interests is the result node; The specification slots include: specification predicate slots, specification subject slots, and specification object slots; Determining the projection value of the element node includes: The normative predicate slot, the normative subject slot, and the normative object slot are respectively associated with the action predicate slot, the agent slot, and the recipient slot of the fact subgraph; For each pair of corresponding slots, determine whether the slot value of the fact subgraph falls into the synonym domain of the normative slot according to the preset synonym domain table. The synonym domain is a collection of fact terms that are sufficient to serve as subordinate facts of the normative slot in the normative sense. The ratio of the number of slot pairs falling into the synonym domain to the total number of standard slots in the element node is used as the slot alignment degree. Using the normative correlation coefficient of the essential node to each causal edge as the weight, the corroboration weights of each causal edge in the fact subgraph are weighted and summed, and then normalized using the sum of the normative correlation coefficients to obtain the evidence corroboration item. The product of the slot alignment and the evidence verification item is used as the projection value; the activation threshold is set according to the type of the requirement node.

3. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 2, characterized in that, The synonym field table is organized with standardized slots as keys and factual phrases as values. The content of the synonym field table is determined based on at least one of legal provisions, judicial interpretations, and guiding cases and is updated according to the version. Each time the projection value is determined, the version number of the synonym field table used is recorded; Before obtaining the aforementioned fact-cause graph, the following is also included: For any two event units, determine the intersection-union ratio of their timestamp intervals and the Euclidean distance between their spatial coordinates, and determine whether their action predicate slots, action object slots, and action object slots fall into the same synonym domain according to the synonym domain table. When the intersection-union ratio is not lower than the temporal identity threshold, the Euclidean distance does not exceed the spatial identity distance threshold, and all three slots fall into the same synonym domain, the two are merged into one event unit. The merged event unit takes the slot value and spatial coordinate of the one with higher modal reliability, takes the intersection of the timestamp intervals of the two, takes the union of the modal identifiers of the two, and merges the corresponding evidence materials of the two.

4. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, The establishment of a causal edge pointing from a preceding event unit to a subsequent event unit includes: The two event units are respectively designated as the first event unit and the second event unit. When the earliest start point of the timestamp interval of the first event unit is not later than the earliest start point of the timestamp interval of the second event unit, Furthermore, if the ratio of the overlap duration of the two timestamp intervals to the length of the shorter interval does not exceed the overlap tolerance, it is determined that the timing prerequisite condition is met, and the first event unit is the preceding event unit. The duration between the earliest start point of the first event unit and the latest end point of the second event unit is used as the available time window. When the available time window is negative, it is set to zero. The maximum displacement rate of the scene is the sum of the displacement that can be covered within the available time window and the position error tolerance as the displacement budget. When the distance between the two spatial position coordinates does not exceed the displacement budget, it is determined that the spatiotemporal reachability condition is met. The maximum displacement rate of the scene is determined according to the scene type of the case scene. The method further includes: A depth-first search is performed on the fact-cause graph. For each loop detected, the causal edge with the smallest temporal margin in the loop is deleted until the fact-cause graph contains no loops. The temporal margin is the difference between the earliest starting point of the endpoint and the starting point of the causal edge.

5. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, Determining the number of mutually independent support sources includes: Construct an evidence source graph for the case to be verified, wherein the evidence source graph has nodes for evidence carrier, evidence collection behavior and statement subject, and directed edges for the evidence derivation relationship; For any two sources of support supporting the same causal edge, trace back along the evidence derivation relationship to obtain the set of source nodes of the two in the evidence source graph; When there is a statement subject node or evidence collection behavior node in the intersection of two source node sets, it is determined that there is a common pollution source node between the two, and the two are merged into a single independent source of support. When the evidence derivation relationship of a certain supporting source in the evidence source diagram cannot be determined, it is determined that the supporting source and other supporting sources have a common pollution source node.

6. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, Determining the corroboration weights of the causal edges includes: For each independent support source, determine the time deviation and spatial deviation between the time point and location carried by the support source and the event units at both ends of the causal edge, and determine the spatiotemporal consistency that decreases as the time deviation and spatial deviation increase; The product of the modal reliability of the mode to which the support source belongs and the spatiotemporal consistency is used as the single-source corroboration. Multiply the differences obtained by subtracting each of the single-source corroboration degrees from 1, and use the result of subtracting 1 from the product as the corroboration weight of the causal edge. Before determining the corroboration weights, the process also includes: For any two support sources supporting the same causal edge, the temporal incompatibility between the times they carry, the spatial unreachability between the locations they carry, and the logical repulsion between their action predicates are added together to obtain the cross-modal causal divergence. The logical repulsion is given a preset penalty value when the action predicates of the two constitute a mutual exclusion pair in a preset mutual exclusion predicate table, and is zero otherwise. When the cross-modal causal divergence exceeds the contradiction determination threshold, the modal reliability of the two is compared: if the difference between the two modal reliability is not less than the reliability resolution threshold, the support source with higher modal reliability is retained, the other support source is marked as an item to be verified, and the confirmation weight of the causal edge is re-determined based on the retained support source; if the difference between the two modal reliability is less than the reliability resolution threshold, the confirmation weight of the causal edge is reset to zero.

7. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to any one of claims 1 to 6, characterized in that, The method further includes: On the directed path corresponding to the degree of danger realization, event units that do not originate from the actor corresponding to the behavior node are identified as exogenous intervention nodes; When there is no exogenous intervention node on the directed path, the intervention effectiveness split ratio is determined to be zero; when there is an exogenous intervention node, the difference obtained by subtracting the edge effectiveness of each exogenous intervention node from the incoming edge on the directed path is multiplied together, and the result of subtracting the multiplication is taken as the intervention effectiveness, and the ratio of the intervention effectiveness to the sum of the intervention effectiveness and the danger realization degree is taken as the intervention effectiveness split ratio. When the intervention effectiveness shunt ratio is not greater than the blocking threshold, the candidate element subgraph is maintained; When the intervention effectiveness diversion ratio is greater than the blocking threshold and not greater than the sum of the blocking threshold and the graded bandwidth, and the danger creation amount is not lower than the criminalization threshold, the directed path is cut off and the candidate element subgraph is reconstructed into an attempted element subgraph. When the intervention effectiveness shunting ratio is greater than the sum of the blocking threshold and the tiered bandwidth, or when the intervention effectiveness shunting ratio is greater than the blocking threshold and the danger creation amount is lower than the criminalization threshold, the directed path is cut off and the candidate element subgraph is reconstructed into a non-causal element subgraph.

8. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, The method further includes: taking the causal edges associated with the behavior node in the fact-cause graph as evidence edges, and classifying the evidence edges into the pre-behavior period, behavior period, and post-behavior period according to the chronological relationship between the timestamp intervals of the evidence edges and the timestamp intervals of the behavior node; For each piece of evidence, the occurrence rate of the action predicate slot value of its endpoint event unit under the assumption of intent and the assumption of non-intent is retrieved, and the natural logarithm of the ratio of the two is taken. The occurrence rate of the condition is obtained from the factual findings in the effective judgment documents. The natural logarithm is weighted and accumulated using the corroboration weight of the evidence edge, and then superimposed with the prior logarithmic probability determined according to the case type to obtain the subjective cognitive intensity. When the intensity of subjective cognition is not less than the threshold for the establishment of subjective requirements, the subjective requirements are determined to be established.

9. The method for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs according to claim 1, characterized in that, The step of outputting verification conclusions based on matching results includes: The product of the hierarchical matching value, the culpability form factor, and the key edge reinforcement coverage rate is used as the probative score of the target element subgraph; the culpability form factor is determined based on whether the subjective element is established when the culpability form of the target element subgraph is intentional, and based on whether the duty of care is violated when it is negligent; the key edge reinforcement coverage rate is the proportion of causal edges with at least two independent support sources among the key causal edges on which the target element subgraph depends; For each preset alternative causal hypothesis subgraph, the same method as determining the projection value is used to search the fact causal graph for the fact subgraph that maximizes the projection value of the alternative causal hypothesis subgraph, and the explanatory power score of the alternative causal hypothesis subgraph is obtained. The difference between the probative value score and the maximum value among the explanatory value scores is used as the exclusivity difference. An exclusivity threshold is selected according to the nature of the case. When the exclusivity difference is not less than the exclusivity threshold, the verification conclusion that the verification is successful is output. Otherwise, the verification conclusion that the probative value is insufficient is output, along with a list of alternative causal hypotheses that have not been excluded.

10. A system for matching and verifying subgraphs of legal constituent elements based on multimodal causal knowledge graphs, characterized in that, The legal constituent element subgraph matching and verification system includes: processor; The memory stores a computer program that, when executed by a processor, implements the legal constituent element subgraph matching and verification method as described in any one of claims 1 to 9.

Images