A legal semantic reasoning and judgment suggestion generation system based on the DIKWP semantic model

By constructing a judicial knowledge graph using the DIKWP semantic model and 'BUG' theory, and combining it with multi-role perspectives, the transparency and impartiality issues of existing legal AI systems are resolved. This enables the generation of transparent, rigorous, and impartial judgment suggestions, thereby improving the interpretability and reliability of the judicial support system.

CN123118310APending Publication Date: 2026-11-17HAINAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611556967.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2025-10-17
Filing Date
2026-10-09
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing legal AI systems lack transparency, logical integrity, and impartiality in reasoning within the judicial field. They exhibit inexplicable reasoning processes and potentially biased conclusions, failing to meet the judicial requirements for credible results, logical rigor, and multi-faceted review.

Method used

The DIKWP semantic model is used to construct a judicial knowledge graph. Combined with the 'BUG' theory, it automatically identifies loopholes in the reasoning chain and generates correction suggestions. A multi-role semantic perspective review mechanism is introduced to provide judgment generation path learning and output white-box reports to ensure the transparency and fairness of reasoning.

Benefits of technology

It achieves transparency, interpretability, logical rigor, and impartiality in legal reasoning, and the generated judgment recommendations meet the requirements of multi-faceted review, thereby improving the reliability and efficiency of the judicial support system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118310A_ABST
    Figure CN123118310A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of legal artificial intelligence and judicial assistance, and discloses a legal semantic reasoning and judgment-like suggestion generation system based on the DIKWP semantic model. The system maps case facts and evidence, laws and regulations and precedents, reasoning rules, and judicial intent into a judicial semantic graph. Through a reasoning verification module, it detects gaps in evidence, logical contradictions, and interpretative ambiguities. Based on different legal roles, it generates multi-perspective graphs and compares points of contention. Then, combining rules, precedents, and machine learning, it generates suggestions on fact-finding, legal application, and sentencing ranges, outputting a traceable white-box report. The system also compares the reasoning results from prosecutors, defendants or defense attorneys, and neutral judges, identifying disputed nodes in the evidence-elements-legal provisions-conclusion chain. It is applicable to courts, procuratorates, and legal intelligent assistance scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of legal artificial intelligence and judicial intelligent assistance, specifically to a system that utilizes semantic knowledge graphs for legal semantic reasoning and automatically generates judgment-like suggestions. This system provides technical solutions for legal case semantic understanding, reasoning judgment, and suggestion generation for court judges, prosecutors, and intelligent legal assistants. Background Technology

[0002] In recent years, "smart courts" and legal artificial intelligence technologies have developed rapidly. A large number of court judgments and legal regulations have been digitized, providing a rich data foundation for intelligent systems. Currently, there are practices combining large-scale legal models with knowledge graphs to assist judicial work. For example, a domestic research team has released a large-scale vertical legal model that integrates over 200 million court judgments and legal regulations. Combined with legal knowledge graphs and retrieval-enhanced generation technology, it can provide traceable and verifiable answers in various legal scenarios. In some pilot projects, after inputting complete case files into the system, it can automatically generate case analysis reports and related legal documents, helping to improve the efficiency of case review and sentencing analysis; a procuratorate used an intelligent sentencing prediction system to generate sentencing recommendations, with an adoption rate exceeding 99%. This demonstrates the enormous potential of AI in assisting in the drafting of court judgments and sentencing recommendations.

[0003] However, existing large-scale models or document generation systems still have significant shortcomings. On the one hand, many systems are "black box" models, meaning their internal reasoning processes are invisible to users. This can lead to inaccurate or even fabricated content. For example, in the United States, there have been cases where lawyers used ChatGPT to draft legal documents but cited non-existent precedents, and in Brazil, judges have made errors by directly using ChatGPT to generate judgments. These errors stem from the model's lack of rigorous verification of legal semantics and reasoning chains, resulting in outputs lacking reliable basis and explanation. The judicial field demands credible results, rigorous logic, and sufficient evidence, thus placing higher demands on the interpretability and rigor of AI. On the other hand, each case involves multiple parties, and different roles (judges, prosecutors, defense lawyers) may have different perspectives on the facts and application of the law. If the intelligent system fails to balance these multiple viewpoints, it may produce biased conclusions, which is detrimental to judicial fairness.

[0004] To address these issues, scholars have proposed several exploratory approaches. One such approach is the DIKWP semantic model, which adds a "Purpose" layer to the classic "Data-Information-Knowledge-Wisdom (DIKW)" cognitive hierarchy. This model emphasizes introducing subjective intention to guide the objective cognitive process, forming a five-layer semantic structure: Data (D), Information (I), Knowledge (K), Wisdom (W), and Purpose (P). Research indicates that applying the DIKWP model to the judicial field helps ensure the semantic consistency of legal concepts and improves the transparency and fairness of the adjudication process. For example, by clarifying the semantic mapping between case data, legal knowledge, and adjudicative intent, the semantic transparency of judicial decisions and the cognitive transparency of participants can be enhanced. Furthermore, related research has also proposed ideas such as the "BUG" theory of consciousness and cognitive relativity. Among these theories, the "BUG" theory posits that flaws or imperfections in the cognitive process trigger conscious intervention, prompting the system to deviate from its conventional workflow for reflection and correction. Cognitive relativity emphasizes that knowledge and perception differ relative to the observer's framework; different cognitive agents may possess their own perspectives, and their "conscious experience" depends on their own knowledge base and objectives. These theories suggest that in legal reasoning, enabling AI to detect "bugs" (such as gaps in evidence or logical breaks) in the reasoning chain and trigger self-correction, while simultaneously examining differences in conclusions from the perspectives of different roles, would help improve the reliability and fairness of AI-assisted judicial recommendations.

[0005] In summary, there is an urgent need for a new legal intelligent reasoning system that can integrate the aforementioned semantic models and theories: it should construct a judicial knowledge graph using the DIKWP semantic model to enhance semantic understanding and reasoning transparency; it should automatically identify and repair problems in the reasoning chain using "BUG" theory; and it should introduce a multi-subject perspective review mechanism based on cognitive relativity, ultimately generating judgment-like suggestions and their white-box interpretations. This invention addresses this need by proposing a legal semantic reasoning and judgment-like suggestion generation system to improve the rigor, interpretability, and fairness of judicial AI assistance. Summary of the Invention

[0006] The purpose of this invention is to provide a legal semantic reasoning and judgment suggestion generation system based on the DIKWP semantic model. By introducing a multi-level semantic knowledge graph and intelligent reasoning mechanism, it overcomes the shortcomings of existing judicial artificial intelligence systems in terms of reasoning transparency, logical integrity, and fairness. Specific objectives include:

[0007] Constructing a semantic tensor graph of judicial knowledge: Using the DIKWP model, semantic modeling is performed on case facts, legal knowledge, judgment logic, and purpose, forming a multi-dimensional semantic tensor graph containing D / I layer (case data and facts), K layer (legal provisions and case library), W layer (legal reasoning logic), and P layer (sentencing or judgment purpose), to comprehensively represent the ontological semantic structure of the case.

[0008] Automatically identify loopholes in the reasoning chain and generate correction suggestions: Combine the "BUG theory" to build a reasoning verification module, check the reasoning chain in the above graph, identify reasoning breaks, evidence gaps, logical contradictions or ambiguities, and automatically provide suggestions for supplementing evidence, correcting logic or clarifying explanations to ensure the integrity and consistency of the judge's reasoning.

[0009] Introducing a multi-role semantic perspective review mechanism: Based on the idea of ​​"relative consciousness / cognitive relativity", a role perspective reconstruction module is designed to reconstruct the DIKWP graph structure from different legal subjects (such as the prosecutor's perspective and the defendant's perspective). By comparing the differences in conclusions or reasoning under different perspectives, potential biases or asymmetries can be identified, ensuring that the judicial recommendations take into account the reasonable demands of both the prosecution and the defense, and achieve balance and fairness.

[0010] The module enables learning of judgment generation paths. It leverages a vast database of existing judgments and cases to deduce suggested judgment paths from the DIKWP semantic graph through machine learning or rule-based reasoning. This module can generate complete suggested paths, including fact-finding, legal application, and sentencing rationale, based on the case graph, by matching applicable legal provisions, referencing sentencing standards, or similar cases.

[0011] Output-type white-box judgment reports: These reports generate white-box reports similar to human judgments, clearly demonstrating the reasoning process from the facts of the case to the application of law, and the basis for each step. For example, the report lists: the established facts and corresponding evidence, the cited legal provisions and their reasons for application, the chain of reasoning (including key nodes and their semantic relationships), system-identified "bugs" and corresponding corrections, and the final recommended judgment or sentencing range. The entire reporting process is transparent and traceable, facilitating human verification and trust.

[0012] To achieve the above objectives, this invention provides the following technical solution: a legal semantic reasoning and judgment suggestion generation system based on the DIKWP semantic model, comprising a semantic modeling module, a reasoning verification module, a multi-perspective reconstruction module, a path generation module, and a report generation module connected in sequence. The functions of each module are as follows:

[0013] Semantic Modeling Module: This module parses and maps input case materials (such as indictments, lists of evidence, legal texts, etc.) into the DIKWP semantic space, constructing a semantic tensor graph of judicial knowledge. Specifically:

[0014] The Data / Information (D / I) layer extracts the basic factual elements and evidentiary data of the case, such as the time of the incident, the people involved, the course of events, and physical evidence, and formalizes them into structured data nodes and related information descriptions.

[0015] The knowledge layer (K) imports relevant legal provisions, judicial interpretations, and similar case judgment rules retrieved, which are incorporated into the graph as knowledge nodes. These knowledge nodes are linked to fact nodes through semantic relationships (e.g., a certain fact corresponds to the fulfillment of a certain legal requirement).

[0016] The Wisdom Layer (W) represents the logical nodes in the legal reasoning process, including the deductive relationship from factual basis to legal conclusion, the argumentation steps, and logical rules (such as causal, inductive, and analogical reasoning chains). W-layer nodes connect the facts of the D / I layer and the legal principles of the K layer through logical relationships, forming a chain of reasoning.

[0017] The purpose layer (P) embodies the judicial intent and value objectives, including sentencing purposes (punishment, education, rehabilitation, etc.) or judicial goals (fairness, efficiency, etc.). P-layer nodes are connected to W-layer logical nodes and are used to evaluate the degree to which different reasoning paths satisfy the expected judicial objectives.

[0018] Figure 1 The diagram illustrates the knowledge graph structure of the aforementioned DIKWP five-layer semantic model in legal cases: the lower layer contains case factual data and evidence, the middle layer contains legal knowledge and existing cases, and the upper layer contains reasoning logic and the purpose of the judgment. The nodes and links in the diagram together constitute a semantic tensor graph, used to support subsequent reasoning and analysis.

[0019] Inference Validation Module: This module implements a link validation mechanism based on "BUG" theory, automatically reviewing the inference process in the semantic graph. It traverses each inference step in layer W, checking whether its preconditions and supporting evidence are sufficient.

[0020] If a reasoning step is found to lack necessary supporting evidence (e.g., the legal requirements do not correspond to any factual evidence nodes), it is determined that there is an evidence gap bug.

[0021] If a contradiction is found between the conclusion and the premise (e.g., the legal provisions corresponding to two knowledge nodes conflict with each other, or the witness testimonies contradict each other), then it is determined that there is a logical contradiction bug.

[0022] If the application of a legal concept is ambiguous or uncertain (e.g., the legal clause contains vague wording that may have multiple interpretations), it is considered an ambiguous interpretation bug.

[0023] For each identified bug, the module generates corresponding correction suggestions: for gaps in evidence, it suggests supplementing relevant evidence or conducting factual investigations; for logical contradictions, it suggests introducing additional arguments to eliminate the contradiction or choosing the more credible side; for ambiguous clauses, it suggests referring to authoritative interpretations or similar precedents for clarification. When a bug is detected, the system essentially simulates the "conscious intervention" of a human judge—actively identifying problems and reflecting on and correcting them. The above verification and suggestion process is as follows: Figure 2 The process described is as follows: input inference chain, output list of issues and modification suggestions.

[0024] Multi-Perspective Reconstruction Module: Based on the principles of cognitive relativity, this module re-examines the case from the perspectives of different subjects. In terms of implementation, the positions of different roles can be simulated by adjusting the target node in the P-layer.

[0025] From the prosecutor's perspective: The P layer is set to be biased towards the goal of "achieving conviction and punishment given evidence", and the W layer reasoning is guided to focus more on the path to establishing the charges, such as strengthening the weight of the nodes of evidence pointing to the defendant's guilt.

[0026] From the defendant's (defense) perspective: The P layer is set to prioritize the purpose of "protecting the defendant's rights when there is doubt about the evidence," guiding the W layer logic to consider more possible paths to innocence or lesser guilt, such as emphasizing insufficient evidence or circumstances favorable to the defendant.

[0027] From the perspective of a neutral judge: In contrast, layer P aims for impartial judgment, without bias towards either side, and conducts reasoning according to conventional weights.

[0028] Based on the different settings mentioned above, the module generates three corresponding DIKWP semantic maps (or correspondingly adjusted sub-maps). It then compares the reasoning results of each map with key intermediate conclusions: if there are significant differences between the prosecutor's and defendant's perspectives on a key fact or application of law, it indicates that the disputed point may be biased or uncertain and requires closer examination. Based on this, the system can adjust the reasoning from a neutral perspective, strengthening the argumentation on the disputed point, ensuring that the final recommendation takes into account the concerns of both parties and avoids one-sided injustice. It points out that cognitive relativity emphasizes the importance of the differences and alignment of perspectives among different cognitive subjects; this module aims to achieve the conversion and alignment between the cognitive frameworks of different legal roles to ensure the relative objectivity and fairness of judicial conclusions.

[0029] Path generation module: After the above verification and multi-perspective balancing, the system uses a decision-based path generation learning mechanism to produce specific decision suggestions. This module can employ machine learning models or rule / case-based reasoning methods.

[0030] Utilizing a large database of effective court judgments as training or matching materials, the model learns reasoning patterns from case elements to judgment outcomes. For example, the model learns how facts are presented, legal provisions are cited, arguments are presented, and sentencing is determined in a judgment. The reported legal model provides high-quality answers through precise "rule-fact" mapping; this module similarly achieves precise matching and deductive output between rules (legal provisions) and facts.

[0031] In practical execution, the module extracts key information from the semantic graph (verified facts, legal points, and objectives from a balanced perspective) and generates a judgment path through the model: including which legal provisions to apply to which facts, whether the crime or legal liability is established, and the sentencing range or outcome determined by referring to sentencing standards.

[0032] The path generation process also considers case-based reasoning: for example, by retrieving past cases similar to the current case through the knowledge layer, the reasoning behind those judgments is incorporated into the reasoning for this case, ensuring that the suggestions have a basis for reference and persuasiveness (similar cases are judged similarly). This can improve the consistency between the suggested results and judicial practice, making the suggestions output by AI more "close to the thinking of human judges".

[0033] Report Generation Module: Finally, the system integrates the above results to generate a white-box report similar to a judgment. The report format references human judicial documents, but adds a transparent explanation of the machine reasoning process.

[0034] The factual section lists the key facts of the case after semantic extraction and verification, indicating the sources of the evidence upon which the facts are based, making the factual findings clear and unambiguous. For example: "The defendant committed act X on a certain date, as evidenced by witness testimony Y and physical evidence Z."

[0035] The section on applicable law lists the applicable legal provisions and their interpretations, explaining how each provision corresponds to the aforementioned facts. It demonstrates that clearly defining the correspondence between evidence and legal requirements reduces errors in subjective interpretation. This report will show that each factual point satisfies a specific legal requirement, citing relevant legal provisions and providing necessary explanations (such as citing authoritative cases or rebuttals to clarify ambiguous clauses).

[0036] Reasoning Chain Display: This section uses visualization or tables to show the reasoning path from facts to conclusions. For example: "Fact A + Legal Clause B → Requirements Met → Conclusion C is Derived," displaying the W-layer logical deduction process. If the verification module previously found and corrected bugs, this section also includes a note explaining this (e.g., "Due to insufficient evidence, the lowest value has been assumed" or "Due to differences between the two opinions, the interpretation more favorable to the defendant has been adopted"), to demonstrate the rigor of the reasoning process and the adjustments made.

[0037] Conclusions and Recommendations: Provide the final judgment-like recommendations, including the wording of the judgment (such as sentencing recommendations or civil liability determinations). Additionally, as this is a "white-box" report, you may add explanations of the differences between this conclusion and different perspectives (if any) and the sources of confidence in making these recommendations (e.g., reference to the key points of a guiding case).

[0038] The overall output of the report is transparent and traceable, allowing users (judges or prosecutors) to easily verify the basis for each step. This white-box approach enhances users' trust in AI-assisted decision-making and serves as a powerful tool for trial assistance or case-handling reference.

[0039] Compared with the prior art, the present invention has the following significant innovations and advantages:

[0040] Deep semantic modeling: The DIKWP five-layer model is introduced into the legal field to achieve a full-link semantic characterization from case facts to judgment objectives. This essential semantic modeling ensures the consistency and transparency of legal concepts, making up for the shortcomings of traditional legal AI that is based solely on text matching or shallow rules.

[0041] Automatic Reasoning Chain Verification: This invention innovatively applies "BUG" theory to the legal reasoning process, automatically scanning the reasoning chain to identify vulnerabilities and providing fix suggestions. It essentially incorporates an intelligent "quality control" step, ensuring the rigor and reliability of the output suggestions. Traditional systems often lack mechanisms to verify the correctness of their own reasoning, while this invention achieves AI self-reflection and improvement through BUG detection.

[0042] Multi-role perspective balancing: A semantic perspective mechanism is proposed that can simulate the positions of both the prosecution and the defense for reasoning, and adjust neutral reasoning based on differences. This method ensures that AI suggestions are unbiased and in line with the principle of judicial fairness, representing a pioneering exploration in judicial AI.

[0043] Judgment Path Learning Generation: This invention integrates machine learning and knowledge reasoning to learn judgment formation paths from a large number of existing judgments and combines them with the semantic graph of the current case to generate personalized inferences and outputs. Compared with simple template filling or sentencing prediction, the suggestions generated by this invention are more flexible and intelligent, supported by big data and able to be adjusted according to the semantics of individual cases, resulting in outputs that are more in line with actual case handling.

[0044] White-box interpretable output: This groundbreaking approach provides white-box reports in a judgment-like format, detailing the basis and process of AI decision-making. This interpretable AI output is of significant value in the judicial field, meeting judges' requirements for AI-assisted decisions to be "understandable and trustworthy." In contrast, many existing systems only provide conclusive results without explanation, while the report in this invention enhances the auditability and credibility of the results.

[0045] In summary, through the above technical solutions, this invention constructs a legal intelligent decision support system that combines deep semantic understanding, logical self-checking, multi-perspective balancing, and interpretable output. It is expected to greatly improve the intelligence level and reliability of judicial assistance, and has high practical value and patent innovation. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0047] Figure 2 This is a flowchart of the inference verification module in this invention.

[0048] Figure 3 This is a schematic diagram illustrating the comparison of semantic perspectives across multiple roles in this invention.

[0049] Figure 4 This is a schematic diagram of an example fragment of the class-based white-box report generated by the present invention. Detailed Implementation

[0050] The system modules and processes of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Example 1: Overall System Flow

[0051] See Figure 1 and Figure 2 This embodiment demonstrates the application process of the system of the present invention in a simple criminal case. Case details: Defendant A is accused of stealing property from a store.

[0052] Semantic Modeling: First, the case materials are input into the system. The facts of the case include: Defendant A entered a store late one night and took several items of unknown value; there is surveillance video as evidence. The legal basis involves the theft provisions of the Criminal Law, whose elements include "secretly stealing public or private property, where the amount is relatively large or the theft is repeated," etc. The system, through natural language processing and knowledge graph mapping, constructs the DIKWP semantic tensor graph:

[0053] The D / I layer contains fact nodes: "entering the store" and "taking the goods," as well as evidence nodes: "surveillance video proving A's behavior." The information layer supplements this with information where the value of the goods is not yet clear.

[0054] Layer K includes the provisions of the Criminal Law concerning theft, as well as judicial interpretations of the criteria for determining "relatively large amount" (such as monetary thresholds). Several precedents from similar theft cases were also retrieved as references, clearly defining the sentencing ranges corresponding to different values.

[0055] Layer W establishes logical relationships: for example, a node indicating "A committed theft" can be used to deduce "the objective elements of theft," and "the stolen property is worth X yuan" relates to "whether the amount is substantial." Currently, due to the unknown value, a node in Layer W is formed here as a pending node.

[0056] Layer P contains objective nodes such as "punish theft according to law" and sentencing objectives such as "balancing punishment and education." These objectives are associated with nodes in Layer W and are used to assess the severity of the offense. For example, if the value is not significant but the circumstances are egregious, appropriate punishment may be recommended for deterrent purposes.

[0057] Inference verification (bug detection): After constructing the graph, the system verifies the W-layer inference chain. Figure 2 For theft, one of the essential elements is that the value of the property reaches the statutory standard:

[0058] The system detected that the corresponding node "the value of the stolen property reaches a substantial amount" currently lacks clear evidence (the value is undetermined). This is marked as an evidence gap bug. The system suggests: "The value of the stolen property needs to be determined. For example, obtain the specific amount through appraisal or market price inquiry." If the appraisal result is insufficient to meet the "substantial" standard, the applicable law may be adjusted from a felony to a misdemeanor (e.g., attempted theft or a case involving a small amount).

[0059] Secondly, the system checks for logical inconsistencies. The evidence in this case is relatively simple (surveillance video), and there is no contradictory evidence; therefore, there is no such bug.

[0060] Regarding the ambiguity check, the wording of the legal provisions is relatively clear, and the term "relatively large amount" has judicial interpretation standards, so there is no ambiguity. However, the system prompts: "The value determination standard is based on a judicial interpretation of a certain year. Article X stipulates that the starting point for a relatively large amount of theft is Y yuan." It is recommended to clarify this in the report to prevent inconsistent understanding.

[0061] To address the evidence gap bug, the system not only suggested obtaining expert opinions but also consulted the knowledge base for similar case handling: for example, in a precedent where the value could not be precisely determined, the court used the minimum valuation method and prosecuted the offender for a lesser amount. The system uses this as a corrective measure: if expert opinions cannot be obtained, reasoning can continue based on the minimum value estimate (or the lesser amount scenario) to avoid interruption.

[0062] After verification, the system updates the graph: a hypothetical node "assume a value of approximately Y yuan (based on the lowest estimate)" is added to layer W to temporarily fill the evidence gap, and will be updated again when new evidence becomes available.

[0063] Multi-view reconstruction: After map correction, the system enables the multi-role perspective module ( Figure 3First, a prosecutor's perspective graph is generated: the goal of layer P is set as "pursuing liability for theft," and the system strengthens the prosecution's strong points—surveillance evidence directly identifying the defendant's actions, assuming the value is at least a certain amount, thus influencing the reasoning chain to support the theft charge. Then, a defendant's perspective graph is generated: the goal of layer P is "mitigating / exempting the defendant from liability," and the system weakens the influence of some evidence—such as considering that although surveillance footage shows the behavior, it cannot prove subjective intent to steal, or arguing that the value is low—influencing the reasoning chain to support the charge of non-acquittal or characterization as a significantly minor offense. Finally, the two perspectives are compared:

[0064] From the prosecutor's perspective, the graph supports the conclusion of "completed theft" (assuming the value meets the standard).

[0065] From the defendant's perspective, the map may lead to the conclusion that "the value does not meet the standard for conviction, or there is a lack of criminal intent" (e.g., claiming it was taken by mistake).

[0066] Difference Analysis: The system identifies the value judgment as a key point of contention—one side presumes sufficient guilt, while the other argues against it. Another potential point of contention is subjective intent—the defendant or defense may claim no intent to steal (e.g., believing the item to be theirs). These become points of contention that require balancing.

[0067] The system adjusts the neutral judge's perspective graph accordingly: adding detailed argumentation nodes to the points of contention at layer W. On one hand, it cites the prosecutor's perspective: surveillance footage shows the behavior was covert and without permission, presuming intent to steal; on the other hand, it also incorporates the defense's perspective: if the defendant has no prior criminal record and the value of the stolen item is low, the subjective malice may be minimal. Regarding the value issue, the neutral perspective temporarily adopts the minimum estimated value Y yuan suggested by the verification module, noting that this value does not meet the "relatively large" standard. If Y yuan < the relatively large standard, the neutral perspective graph will conclude that "the theft did not reach a relatively large amount, and is only a common theft" (possibly corresponding to a minor offense / public order case). In this way, through multi-perspective comparison, the system ensures that all controversial issues are considered, balancing the positions of the prosecution and the defense.

[0068] Path Generation: Following the above processing, the system begins generating suggested paths similar to those for judgments. Referring to the judgment generation path learning module, the system invokes a pre-trained legal reasoning model or rule engine:

[0069] The model extracts key information from a complete picture from a neutral perspective: nature of the behavior (secret theft), evidence (video), value (Y yuan), legal provisions (theft clause in the Criminal Law, judicial interpretation of Article Z), etc.

[0070] Based on the training knowledge, the model determines the applicable law: Since Y is below the "relatively large amount," it is recommended to apply the "small amount" circumstances category for theft, or to impose administrative penalties (if no crime is constituted). If Y is small but the act is clear, the model may recommend conviction for theft with a lighter sentence, or, if strictly adhering to the standard, may not be considered a crime (depending on the gap between Y and the standard).

[0071] The model references similar cases: for example, it retrieves a similar case (the defendant stole a small number of items, and the court ruled that he was exempt from criminal punishment but was given an administrative fine). The processing logic of this case is incorporated into the path: that is, "the facts are clear but the harm is not significant, so it is dealt with leniently."

[0072] Ultimately, the path module generates a sequence of decision nodes: Facts established → Theft is determined, but the amount does not meet the criminal threshold → It is recommended not to treat it as a criminal offense or to impose a lighter sentence. If the system still tends to classify it as criminal theft (depending on Y and legal provisions), the path will include sentencing steps, such as "Based on Article X of the Criminal Law, it is recommended to sentence him to n months of detention, suspended sentence," etc.

[0073] In addition, the approach will include legal basis: such as citing a judicial interpretation to explain why the amount did not reach a substantial level, citing criminal law provisions to explain the basis for punishment, and citing relevant cases to explain the reasons for discretion. These will be presented in detail in the report.

[0074] Report Generation: Finally, the system converts the above reasoning path into a white-box report similar to a decision. Figure 4 A sample report is shown below:

[0075] First section: List the case name (e.g., "a theft case") and the nature of the report (intelligent judgment suggestion report), noting that the report is not a formal judgment and is for reference only.

[0076] The factual section lists the established facts one by one, such as "Defendant A entered a store late at night on XX / XX / XXXX and took several items without paying; the above facts are supported by screenshots from surveillance video and are sufficient to establish them." Regarding the value, it states: "The preliminary estimated value of the stolen property is approximately Y yuan" and "It does not reach the threshold of Z yuan for a substantial amount in theft cases." This clearly presents the correspondence between evidence and facts, enhancing semantic transparency.

[0077] Legal Analysis Section: First, state the main legal provisions considered, such as "According to Article XX, Paragraph YY of the Criminal Law of the People's Republic of China (theft provision) and the Supreme People's Court's Interpretation on Several Issues Concerning the Application of Law in Handling Theft Criminal Cases...". Then, explain in sections:

[0078] Qualitative Analysis: This section explains whether the defendant's actions constitute theft. It states that "the defendant's actions meet the objective elements of theft—secretly stealing another's property; and subjectively, having the intent to illegally possess it (inferred from the course of the act)." However, it acknowledges that "because the value of the property involved is only Y yuan, which does not reach the 'relatively large amount' required by criminal law, it does not meet the legal standards for conviction of theft." Judicial interpretations are cited to explain the value standard.

[0079] Circumstances and Purpose: It is explained that the defendant had no history of violence or prior criminal record, and the circumstances were relatively minor; in the spirit of combining punishment with education (P-level purpose), it is not appropriate to punish him as a serious crime.

[0080] In response to the controversy from multiple perspectives, the report specifically noted: "This conclusion comprehensively considers the views of both the prosecution and the defense. The sufficiency of evidence argued by the prosecution and the insufficient value of the defendant's defense have both been taken into consideration. Based on legal standards, the system tends to adopt the latter view, thereby ensuring the fairness of the ruling." This demonstrates how the system balances different perspectives.

[0081] Conclusion Section: Provides a suggested course of action. For example: "In summary, this system recommends that A's actions not be treated as a criminal offense. However, given that his actions constitute a violation of public security regulations, it is recommended that he be transferred to the public security organs for processing in accordance with Article XX of the 'Public Security Administration Punishment Law' (which may result in XX days of detention or a fine). If the judicial authorities still determine that criminal liability should be pursued, it is recommended that he be exempted from criminal punishment or given probation if convicted of theft." The conclusion section should list possible options with reasons to allow the decision-maker to make a discretionary decision based on the actual circumstances.

[0082] White-box explanation: At the end of the report, the system includes a white-box explanation listing key nodes and paths in the reasoning chain (textual description or diagram). For example: "[Fact Node] The defendant secretly took the goods → [Legal Node] Meets the elements of theft → [However] The condition of a substantial amount is not met → [Reasoning Conclusion] Not a theft crime." It also notes previous bugs identified and their handling: "Note: Due to the lack of accurate value of the goods, this system assumes a minimum value of Y yuan for analysis; this assumption tends to favor the defendant. If subsequent evidence shows a higher value, the conclusion can be reassessed." This explanation clarifies the assumptions and adjustments made during the AI ​​reasoning process, ensuring transparency and traceability.

[0083] The report concludes with a list of cited legal provisions and case studies, serving as a footnote for judges to review and verify. These citations demonstrate the verifiability of the recommendations.

[0084] As can be seen from the above embodiments, when processing cases, the system of the present invention can decompose the complex legal reasoning process into clear semantic levels, and ensure the completeness and fairness of the reasoning through automatic verification and multi-perspective analysis, ultimately generating a detailed recommendation report for decision-making reference. In practice, judges or prosecutors can quickly understand the key points of the case and AI suggestions based on the report, and can verify or adjust each step listed in the report, thereby significantly improving the efficiency and quality of case handling. Example 2: Integrated Application

[0085] This embodiment illustrates the effect of integrating the system of this invention into a court's intelligent case-handling assistance platform. A certain intermediate people's court introduced this system as an intelligent judgment drafting assistant, applicable to assisting in the trial of criminal cases: - In a pilot program with real cases, the presiding judge imported the electronic materials from the case file into the system. The system automatically outputs a white-box report similar to a judgment, reducing manual document drafting time by an average of approximately 70%. Judges reported that the legal basis and reasoning process listed in the report were clear and complete, requiring only minor modifications to form a formal judgment. Simultaneously, the sentencing recommendations proposed by the system highly aligned with the court's sentencing standards, and in some complex cases, the system's alerts to evidentiary flaws (such as incomplete chains of evidence) helped judges promptly remedy omissions. - In terms of application by the procuratorate, a certain local procuratorate used the system to review arrest and indictment opinions and assist in drafting. The system clearly points out cases with insufficient evidence in the report and suggests supplementary investigation directions, enabling prosecutors to issue supplementary investigation outlines accordingly, thus improving the quality of case handling from the source. In cases involving guilty pleas and acceptance of punishment, the sentencing range recommendations provided by the system are consistent with local sentencing guidelines. Prosecutors refer to the system's recommendations to propose sentencing negotiation plans, and the rate of adoption by the court exceeds 95%, demonstrating the system's excellent usability.

[0086] Through the aforementioned integrated pilot projects, it is foreseeable that the system of this invention has broad application prospects: it not only improves the efficiency of judicial case handling but also enhances the transparency and standardization of the adjudication process. The system's reasoning verification and multi-perspective analysis functions are particularly important in cases of complexity and conflicting evidence, enabling the discovery of details that might be overlooked manually. In conclusion, the legal semantic reasoning and judgment suggestion generation system of this invention demonstrates outstanding performance in both technological advancement and judicial adaptability, possessing strong value for widespread application.

Claims

1. A legal semantic reasoning and judgment suggestion generation system based on the DIKWP semantic model, characterized in that, Includes a processor, a memory, and a program stored in the memory and running on the processor, the program being configured to implement: a semantic modeling module that maps case material parsing to the DIKWP semantic space, and constructs a judicial semantic tensor graph containing D / I layer (case facts and evidence), K layer (laws and regulations and similar cases), W layer (reasoning rules and links), and P layer (judgment purpose / intention); The reasoning verification module traverses the W-layer reasoning chain to detect three types of "BUGs": evidence gaps, logical contradictions, and interpretation ambiguities. It outputs correction suggestions and updates the graph. The multi-perspective reconstruction module generates graphs from the perspectives of the prosecutor, defendant / defense, and neutral judge based on the purpose weights of the P-layer, compares the differences in conclusions, and identifies points of contention. The path generation module, after verification and reconstruction, combines rule / case reasoning and machine learning to generate judgment-like suggested paths that include fact-finding, legal application, and sentencing ranges. The report generation module outputs a white-box report, presenting a mapping of evidence, elements, legal provisions, reasoning, and conclusion, as well as fixed bugs and a visualized reasoning chain.

2. The system according to claim 1, characterized in that, The edge bands in the judicial semantic tensor graph have relation types and weights. The relation types include at least fact-element, element-legal provision, causality, induction, and analogy.

3. The system according to claim 1, characterized in that, The reasoning and verification module automatically locates constituent elements that are not supported by evidence nodes in response to evidence gaps and generates supplementary evidence collection or alternative evaluation guidelines.

4. The system according to claim 1, characterized in that, The reasoning verification module addresses logical contradiction bugs by performing consistency analysis on conflicting evidence and conflicting legal provisions, and providing priority or credibility selection strategies.

5. The system according to claim 1, characterized in that, The reasoning and verification module addresses interpretation ambiguity bugs by introducing authoritative explanations, guiding cases, or normative opinions from the institute to resolve ambiguities and solidify them into graph constraints.

6. The system according to claim 1, characterized in that, The multi-view reconstruction module generates three sets of graphs by adjusting the objective weight of the P layer and the importance of the nodes in the W layer. When the difference in key conclusions exceeds a preset threshold, the corresponding node is marked as a point of contention and a weighted review is triggered.

7. The system according to claim 1, characterized in that, When performing case similarity retrieval, the path generation module comprehensively considers the matching degree of cause of action, constituent elements, alignment of factual elements, and similarity of sentencing elements, and only introduces the key points of the judgment when the similarity reaches a threshold.

8. The system according to claim 1, characterized in that, The white-box report should include at least the following: a list of factual points and sources of evidence, applicable laws and reasons, a step-by-step reasoning chain, explanations of points of contention and fixed bugs, a recommended main text and sentencing / treatment range, and a list of reference cases.

9. The system according to claim 1, characterized in that, It also includes a user feedback collection unit, which is used to write back the modification opinions of the presiding judge / prosecutor to the graph parameters and path generation strategy to achieve continuous learning.

10. A method for generating class-decision suggestions based on the DIKWP semantic model, characterized in that, include: S1: Analyze case materials and construct a judicial semantic tensor graph containing D / I, K, W, and P layers; S2: Perform bug detection and correction on the W-layer inference chain and update the graph; S3: Reconstruct the three-view graph based on the objective weights of different P-layers and compare the differences, marking the points of contention; S4: From a neutral perspective, combine rule / case reasoning and machine learning to generate decision paths; S5: Outputs white-box reports and visualizes the reasoning chain.