Multi-view law judgment method based on relative consciousness theory and law BUG white box interpretation device

Abstract

The invention discloses a multi-view law judgment method based on a relative consciousness theory and a law BUG white box interpretation device, and relates to the technical field of law artificial intelligence and interpretable AI. The method comprises the following steps: respectively constructing semantic maps representing different cognitive coordinate systems for a defendant, a control party and a judge in a legal case based on a DIKWP model; by comparing and analyzing the multi-view maps, the defects and asymmetry of judgment reasoning in the aspects of evidence consistency, law application, reasoning paths, vertical bias and the like are detected; and quantifying the vertical deviation degree of the judge visual angle and identifying specific defect links. The device is a white box interpretation system for realizing the method, comprises a data processing unit, a graph rendering unit, a defect labeling unit, an interaction control unit and the like, and can visually present the multi-view reasoning chain, the detected BUG and the risk weight thereof. According to the method, the problem of prejudice possibly existing in single-view-angle legal reasoning is solved, and the justice of judicial referees and the transparency and interpretability of AI legal decisions are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of legal artificial intelligence and explainable AI technology, in particular, it relates to a multi-perspective legal judgment method based on relative consciousness theory, and a legal BUG white box explanation device supporting the method. The method and device aim to perform semantic inversion and comparison of legal reasoning from different subject positions, detect bias and asymmetry in the adjudication process, and output the analysis of the adjudication reasoning chain in a transparent and visualized manner, thereby assisting the trial personnel to calibrate the ruling and improving the explainability of AI legal decision-making. BACKGROUND

[0002] Judicial adjudication requires neutrality and fairness, but in practice, the bias of different subject perspectives may affect the adjudication conclusion. For example, the defendant's perspective may emphasize self-defense facts, and the prosecution's perspective may emphasize evidence of conviction, while the judge should remain neutral between the two. However, human judges inevitably have unconscious position tilting due to cognitive limitations and value orientation. How to discover and correct the perspective bias in the adjudication is very important to ensure the fairness of the trial. In addition, with the participation of artificial intelligence in legal decision-making, its potential bias (resulting from unbalanced training data, etc.) also needs to be examined from multiple perspectives to prevent the AI output conclusion from having a single perspective bias.

[0003] The "consciousness relativity" theory points out that consciousness is not absolutely objective and unified, and different cognitive subjects have their own understanding coordinate system, and the information asymmetry between them will cause cognitive bias. This patent uses "each consciousness has its own coordinate system, and there is always conversion loss or error between them" to illustrate this point. This means that in legal judgment, the "relative consciousness" of different participants (such as the prosecution and the defense) may lead to different interpretations of the same facts, and thus cause disputes. If the cognitive perspectives of multiple subjects can be compared and aligned, it will help to find the tilt in the adjudication reasoning. For example, in some cases in the United States, different judges hold "multi-level or opposite perspectives" on the same constitutional issue, and it is difficult to get a comprehensive conclusion by relying on single-path analysis only.

[0004] Meanwhile, the explainability of AI in the legal field has also attracted much attention. Many current legal AI systems (such as large models utilizing deep learning) are considered "black boxes," where users can only see the conclusions but not the internal reasoning process. This is unacceptable in judicial settings, as judges and lawyers need sufficient reasons to convince them of a conclusion. Therefore, academia and industry have proposed the need for "white-boxing" legal AI, hoping to visualize the basis and logic of AI decisions. Existing research has attempted to make legal reasoning paths machine-readable using structures such as knowledge graphs. In particular, white-box evaluation methods for artificial intelligence emphasize making each step of AI's decision traceable and explainable. In legal settings, this means we need a device or system that can fully present the reasoning chain of AI (or AI-assisted humans) adjudication, annotating any flaws and potential risks, making the trial process transparent and auditable.

[0005] In summary, on the one hand, we need a multi-perspective legal judgment mechanism to balance the cognitive biases of different subjects and achieve relative fairness in the adjudication process; on the other hand, we need a white-box interpretation device to graphically present the reasoning process, helping to understand and correct flaws in the judgment reasoning. Currently, there is no mature solution that combines these two aspects. Some systems can output legal knowledge graphs but do not support multi-perspective comparisons; others only provide simple rule explanations, lacking annotation of reasoning flaws and risk assessment. The multi-perspective legal judgment method and white-box interpretation device proposed in this invention are designed precisely to address these shortcomings. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-perspective legal judgment method based on the theory of relative consciousness, and a white-box explanation output device for legal bugs. By introducing a subject-object semantic inversion mechanism, this method constructs a perspective mapping of the defendant, the prosecution (prosecutor / plaintiff), and the judge on the DIKWP semantic graph. It reconstructs the case reasoning process from different cognitive coordinate systems, and then compares and analyzes these perspectives to discover potential positional biases and semantic asymmetries in the judgment. The accompanying white-box explanation device can output the above multi-perspective reasoning results in a visual graph format, clearly indicating the reasoning defects, repair paths, and their impact weights on the conclusion, thereby supporting manual review and intelligent correction.

[0007] Technical Solution Overview:

[0008] 1. Multi-perspective DIKWP graph construction: Based on the DIKWP (Data-Information-Knowledge-Intelligence-Intent) model, semantic graph perspectives are constructed for the main participants in the case:

[0009] 2. Defendant's Perspective Map (Defense Perspective): This constructs a semantic map of the facts and reasons for the case from the defendant's (or defense's) standpoint. For example, it emphasizes evidence favorable to the defendant, questionable aspects of unfavorable evidence, and legal exemptions. In the DIKWP five-layer model, the defendant's perspective might highlight the goal of "avoiding conviction / mitigation" at the intent level.

[0010] 3. Prosecution Perspective Map (Accusation Perspective): From the perspective of the prosecutor or plaintiff, this constructs a semantic map focusing on conviction or accountability. For example, it strengthens the chain of evidence pointing to the defendant's guilt or emphasizes the punitive spirit of the law. Its intent layer reflects the goal of "achieving punishment / maintaining order."

[0011] 4. Judge's Perspective Map (Neutral Judgment Perspective): A semantic map formed by synthesizing the above information from the judge's perspective. The judge's perspective strives for balance, but in practice, it may lean towards one side due to cognitive biases. This invention distinguishes between the ideal neutral perspective and the actual judgment perspective: the ideal neutral perspective refers to the reasoning structure that should be formed from a completely fair and objective standpoint, while the actual judgment perspective is extracted from the judgment document and reflects the judge's actual reasoning in the case.

[0012] Each of the above semantic graphs is a mesh structure containing nodes at each layer of DIKWP and their relationships. During construction, structured information from the case (e.g., factual information, defense arguments, prosecution arguments, court findings, etc.) can be used to populate different perspective graphs. To achieve semantic inversion between subject and object, the system introduces a semantic mirroring relationship between the defendant's and prosecution's perspectives: for example, some arguments alleged by the prosecution are points that need to be refuted from the defendant's perspective, and vice versa. Through this mirroring mapping, it is ensured that graph elements from the two opposing perspectives can be compared one-to-one.

[0013] 1. Multi-perspective reasoning and comparative analysis: The three semantic maps above are automatically compared and analyzed, focusing on the following aspects:

[0014] 2. Consistency between Evidence and Facts: Examining the differences in the descriptions of the same facts from the defendant's and prosecution's perspectives. For example, the same evidence might be given high credibility from the prosecution's perspective, but weakened or denied from the defendant's perspective. This invention quantifies this inconsistency by comparing the differences in attribute weights for corresponding nodes using a semantic matching algorithm.

[0015] 3. Differences in Applicable Law: Compare the legal rules cited by both the prosecution and the defense with the rules ultimately applied by the judge. For example, if the prosecution emphasizes a strict clause while the defendant cites an exemption clause, the judge will choose one. If the judge ignores the exemption grounds raised by the defendant, this will become apparent in a multi-perspective comparison—this node is significant in the defendant's perspective but missing in the reasoning chain of the judge's perspective, indicating that the judgment may be biased.

[0016] 4. Length and Branches of Reasoning Paths: The defendant's perspective may offer alternative explanatory paths (such as the hypothesis of innocence), the prosecution's perspective offers a path of guilt, and the judge's perspective actually only employs one of them. This invention extracts these parallel reasoning chains and examines whether the judge responds to or excludes the other path. If a reasonable reasoning chain from another perspective is ignored, the judgment indicates a perspective asymmetry, which is a type of judicial bug.

[0017] 5. Position Bias Measurement: Based on the above comparison results, the "deviation" between the judge's perspective and the respective perspectives of the prosecution and defense is calculated. We can use a semantic coverage index: for example, how many key nodes in the defendant's perspective map are mentioned / considered in the judge's reasoning; and how many nodes from the prosecution's perspective are considered. Judges' positional bias is expressed as a percentage or score. If it is found that the judge's view almost completely covers the prosecution's perspective while rarely involving the defendant's perspective, a significant bias exists. This method marks it as a "positional bias bug."

[0018] 6. Semantic Symmetry Check: According to the theory of relativistic consciousness, an ideal, fair judgment should semantically consider all parties equally. This invention detects injustice by identifying semantic symmetry deficiencies. For example, if a judgment details facts favorable to the plaintiff while downplaying facts favorable to the defendant, this asymmetry will be prominently displayed in a two-view graph comparison (imbalance in node density and weight). Based on this, the system identifies semantic symmetry deficiencies and pinpoints where the asymmetry exists (fact finding, legal analysis, or reasoning).

[0019] 7. Defect Identification and Correction Path: Based on the results of multi-perspective comparison, this invention can identify several defective links in the judge's reasoning process:

[0020] 8. For example, if the judge finds that a key defense raised by the defendant has been omitted from the reasoning chain, this omission is marked as a defective link, and the suggested correction path is "to add a commentary on the defense".

[0021] 9. For example, if it is found that a judge has over-relyed on a certain piece of evidence without considering the controversy over its reliability, the defect is "one-sided evidence evaluation", and the corrective path is to "add a reliability analysis of the evidence and respond to objections".

[0022] 10. For the overall tendency of bias, it is pointed out as a high-level defect, and the correction path is "to incorporate the other party's perspective and balance considerations into the judgment analysis." The system provides correction suggestions for different defective aspects based on a preset strategy library (this part can work in conjunction with the error correction suggestion module in Patent 2).

[0023] 11. Legal Bug White-Box Explanation Device Architecture: To implement the above method, a dedicated device is needed for presentation and interaction. This device includes:

[0024] 12. Data Processing Unit: Responsible for extracting multi-perspective information from the judgment and related materials and constructing a semantic graph data structure. It can be considered the software implementation of the aforementioned multi-perspective construction steps.

[0025] 13. Semantic Graph Rendering Unit: Converts semantic graphs into visual graphics. This device uses color and shape encoding: for example, different colors distinguish nodes from the perspectives of the defendant, prosecution, and judge; dashed and solid lines or different arrows represent reasoning relationships from different perspectives. Multiple perspective graphs can be displayed overlappingly or side-by-side for comparison. For example, on the white-box interface, the representation of the same fact node in three perspectives will be presented in an associative manner for easy comparison.

[0026] 14. Defect Annotation Unit: Detected defects are marked on the graph with special markers (such as flashing red circles). For example, a missing node in the defendant's viewpoint can be marked in red at the corresponding node in the defendant's viewpoint, with a line pointing to the corresponding blank space in the judge's view, indicating "This viewpoint was not considered." For overall bias, intuitive symbols such as a tilted scale icon can be displayed on the interface.

[0027] 15. Interactive Control Unit: Allows users to select different viewpoints for displaying and hiding the graph, zoom in on a portion of the inference chain, and click to view detailed node information, among other interactive operations. Users can also click on defect markers to display detailed explanations and correction suggestions.

[0028] 16. Risk Weight Calculation Unit: Each marked defect is assigned a risk weight value to quantify its potential impact on the final judgment. The weight can be calculated based on the degree to which the defect relates to core issues, the degree of deviation from the judgment's objective, etc. For example, omitting a key defense argument might result in a high risk weight, indicating its potential to affect the fundamental fairness of the judgment; while omitting minor details would have a lower weight. The device displays the weight on the interface using color variations or warning level icons; for example, a red exclamation mark indicates a high-risk defect.

[0029] 17. Report Output Unit: Generates a report from the above multi-perspective analysis results and map snapshots for archiving or further review. The report includes a summary of key biases, a list of deficiencies, and recommendations.

[0030] 18. Workflow: When a judgment (or AI-generated judicial opinion) is input into the system, the device first acquires single-view semantic data, then generates multi-view graphs for analysis and comparison. Next, the defect labeling unit and risk calculation unit locate and assess bugs, and finally, the data is presented to the user through a visual interface. Users can intuitively "see" how the judgment reasoning process unfolds and converges from different perspectives on the device. With the help of this white-box interpretation tool, judges can clearly understand whether their judgment reasoning is balanced and whether there are any omissions; judicial supervision authorities can also use this to review whether the judgment is fair. Attached Figure Description

[0031] Figure 1 This is a diagram illustrating a multi-perspective legal judgment method.

[0032] Figure 2 A schematic diagram of the human-machine interface for a white-box explanation device for legal bugs. Detailed Implementation

[0033] Example 1: Multi-Perspective Analysis of a Criminal Case. In a criminal appeal case, the court of first instance found the defendant guilty but imposed a heavy sentence. The defendant's grounds for appeal included several arguments that were not adopted by the court of first instance. We input the first instance judgment and appeal materials into the system of this invention. The system constructed: Defendant's Perspective Map: Emphasizing factors of innocence or mitigation, such as questionable key evidence, no prior criminal record, and insufficient motive, highlighting the goal of "avoiding wrongful convictions" at the intent level. Prosecution's Perspective Map: Emphasizing the complete chain of evidence for guilt and the defendant's clear and malicious motive, with the intent level being "combating crime and upholding justice." Judge's Perspective Map: Based on the actual reasoning in the first instance judgment, emphasizing the evidence for conviction and the legal basis, and downplaying the defendant's arguments.

[0034] Comparative analysis revealed that the defendant's defense that "key physical evidence may have been contaminated" was a crucial point in the defendant's perspective map, but completely absent from the judge's perspective map—the device marked this point with a red circle, indicating "this defense point was not evaluated in the judgment," with a risk weight assessment of medium (because even if this defense were considered, a conviction might still be established, but procedural fairness is involved). Another finding was that the judge's discussion of motives closely resembled the prosecution's perspective, while the mitigating factor of the defendant's lack of prior criminal record was only briefly mentioned. The device showed through semantic coverage that the judge's viewpoint covered 90% of the prosecution's perspective but only 40% of the defendant's perspective, indicating a clear bias (warned by a tilted scale icon, indicating high weight). In response to these deficiencies, the system added suggestions: "The judgment should include an explanation of the credibility of the physical evidence to avoid procedural flaws affecting credibility," and "Balance the evaluation of the defendant's favorable circumstances, such as the lack of prior criminal record, and their impact on sentencing to reflect sentencing fairness." Based on these suggestions, the judge adjusted the reasoning in the second instance judgment, supplemented the response to the defense's arguments, and appropriately reduced the sentence. After revision and re-examination using the device, the position coverage index tended to be balanced (the defendant's perspective coverage increased to 70%), and the red defect markers disappeared or turned green (indicating that the problem has been resolved), indicating that the judgment tended to be more coordinated and fair under multiple perspectives.

[0035] Example 2: White-box Analysis of AI Legal Decision-Making. A certain intelligent legal consultation system (LLM-driven) provides mediation suggestions for a contract dispute. To verify the reliability of the AI ​​suggestions, we conducted a white-box analysis using the device of this invention. The system breaks down the AI's suggestions into three semantic components: the perspective of Party A (equivalent to the defendant), the perspective of Party B (equivalent to the plaintiff), and the AI's own neutral suggestion perspective. The results showed that the AI's "neutral suggestion" actually leaned more towards protecting Party B's interests, with its arguments almost entirely based on information provided by Party B, while giving insufficient consideration to Party A's claims. On the device interface, the perspective maps of Party B and the AI ​​almost overlapped, and many nodes in Party A's perspective map (such as different interpretations of a certain clause in the contract) were not reflected in the AI ​​suggestions. The device marked these omissions with multiple red circles. The risk assessment indicated that this one-sided bias might lead to a mediation that was unacceptable to both parties. Based on this, we asked the AI ​​to readjust its suggestions based on new evidence provided by Party A, so that it covered the key points of both parties. After adjustment, the device checked again, and all the multi-perspective elements were reflected, indicating that the suggestions had become more balanced and fair.

[0036] Example 3: Judicial Trial Decision Support System. A court applied the method and device of this invention to assist in the collegial panel discussion of difficult cases. Panel members can simultaneously view the reasoning graphs of different opinions through a white-box interface: the differences between the perspectives of judges who support a guilty verdict and those who disagree are directly displayed on the graph. For example, if a panel member believes there is insufficient evidence for an acquittal, many key points (issues of evidence reliability) are missing from the majority viewpoint's graph. Through this device, the panel notices these points, is forced to confront the objections, and discuss supplementary evidence or accept the possibility of an acquittal. Ultimately, the court decides to conduct further investigations to obtain crucial evidence, eliminate reasonable doubt, and then convict. In this process, the device of this invention plays a "cognitive calibration" role, making explicit and structured the different cognitive coordinate systems implicit in the judge's mind, thereby improving the quality of collective decision-making.

[0037] In summary, the multi-perspective legal judgment method provided by this invention enables the judicial decision-making process to proactively examine its own performance under different viewpoints, identify biases, and correct them in a timely manner. The accompanying white-box interpretation device transparently presents the internal logic of legal reasoning, achieving interpretability and oversight of legal AI decision-making. This innovation, applying the theory of relative consciousness to legal reasoning and supplementing it with visualization tools, significantly enhances the transparency and credibility of legal AI and the judicial system. In terms of target markets, this invention can be used in intelligent auxiliary systems for court judgments, internal judgment quality inspection platforms for public security, procuratorial, and judicial organs, and platforms for compliance review of AI-generated legal content, possessing broad prospects for licensed applications and social value.

Claims

1. A multi-perspective legal judgment method based on the theory of relative consciousness, characterized in that, Includes the following steps: S1. Construction of Multi-Perspective DIKWP Semantic Graphs: Based on the DIKWP model, independent semantic graphs are constructed for the defendant, prosecution, and judge in the case. The semantic graphs contain nodes at five levels: data, information, knowledge, wisdom, and intent, as well as the reasoning relationships between nodes. S2. Multi-Perspective Reasoning Comparative Analysis: The defendant's perspective graph, prosecution's perspective graph, and judge's perspective graph constructed in step S1 are automatically compared to identify differences in factual descriptions, legal application, and reasoning paths among the parties. S3. Identification of Position Bias and Defects: Based on the comparison results of step S2, the degree of position deviation of the judge's perspective relative to other perspectives is calculated, and specific defective links in the reasoning process are located. S4. Output Analysis Results: Generate an analysis report that includes multi-perspective comparison results, position deviation, defective links, and correction suggestions.

2. The method according to claim 1, characterized in that, The construction of a semantic graph for the judge subject in step S1 includes: constructing an actual judge's perspective graph based on the actual content of the judgment document; constructing an ideal neutral perspective graph based on a fair and objective stance; the comparative analysis in steps S2 and S3 includes comparing the actual judge's perspective graph with the ideal neutral perspective graph.

3. The method according to claim 1, characterized in that, The multi-perspective reasoning comparison analysis described in step S2 specifically includes at least one of the following: consistency check of evidence and facts: using a semantic matching algorithm, compare the differences in attribute weights of the same evidence or fact node in the defendant's perspective graph and the prosecution's perspective graph; difference analysis of applicable law: compare the consistency of the legal rule nodes cited by both the prosecution and the defense with the legal rule nodes ultimately applied by the judge, and detect whether the judge's perspective has ignored the key legal basis put forward by one party; comparison of reasoning paths: extract parallel reasoning chains in the defendant's perspective and the prosecution's perspective, and check whether the judge's perspective reasoning path provides reasons for exclusion or responses to paths that were not adopted.

4. The method according to claim 1, characterized in that, The positional bias mentioned in step S3 is measured by the semantic coverage index, which is the proportion of key nodes in the defendant's perspective map or the prosecution's perspective map that are mentioned or considered in the judge's perspective map.

5. The method according to claim 1, characterized in that, The types of defects mentioned in step S3 include positional bias bugs and semantic symmetry deficiency bugs; wherein, the positional bias bug is judged by the severe imbalance in the semantic coverage of the prosecution and defense perspectives from the judge's perspective; the semantic symmetry deficiency bug is located by comparing the imbalance in the density, weight or detail of corresponding nodes in the dual-perspective graph.

6. A white-box apparatus for interpreting legal bugs to implement the method according to any one of claims 1 to 5, characterized in that, include: The data processing unit is used to extract information from legal documents and construct a data structure for a multi-perspective DIKWP semantic graph. The graph rendering unit is used to convert the multi-view semantic graph into a visual graphic and distinguish the view nodes and relationships of different subjects with different colors or shapes. The defect annotation unit is used to specially mark the defective links identified in step S3 on the visualization map; the interactive control unit provides a user interface that supports users to select, zoom, and click to view details on the visualization map; The report output unit is used to generate and output analysis reports.

7. The apparatus according to claim 6, characterized in that, Also includes: The risk weight calculation unit is used to assign a quantified risk weight value to each marked defect to assess the degree of impact of the defect on the judgment conclusion. The risk weights are presented on the interface through the graph rendering unit in the form of color depth or warning icons.

8. The apparatus according to claim 6, characterized in that, The graph rendering unit supports visualizing semantic graphs from the perspectives of the defendant, prosecutor, and judge in a side-by-side or overlapping manner.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images