Method and system for generating a decision

Abstract

One or more embodiments of the specification provide a judgment generation method and system, which realizes the alignment of artificial intelligence reasoning path and legal argumentation logic by decomposing the trial thinking of a judge into a chain-verification trial process of "fact structuring-dispute focus identification-evidence analysis-law application-judgment generation". The input and output of each stage are objectified and structured, so that the entire judgment generation process is "visible", the decision basis is clear, the reasoning steps are traceable, and a controllable, reliable, rigorous, universal and transparent intelligent auxiliary trial mechanism is built.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a method and system for generating judgments. Background Technology

[0002] In judicial practice, judges must manually read case files, analyze evidence, apply the law, and draft judgments. This process is highly dependent on personal experience and intellectual labor, faces efficiency bottlenecks, and is easily influenced by subjective factors. To improve the quality and efficiency of trials, intelligent technologies have emerged, and their development has mainly gone through four stages: document template-based trials, element-based trials, knowledge graph applications, and direct generation of large models. However, these methods still have certain limitations and cannot meet the high-level requirements of judgment quality, logical rigor, and procedural interpretability in practice. Summary of the Invention

[0003] In view of the above, one or more embodiments of this specification provide the following technical solutions:

[0004] According to a first aspect of one or more embodiments of this specification, a method for generating a judgment is proposed, the method comprising:

[0005] The complaint, answer, and court transcript of a legal case are input into a fact generation model. The fact generation model is then used to perform structured processing of the facts of the legal case through entity recognition and relation extraction, generating disputed and undisputed facts related to the legal case and identifying the focus of the dispute.

[0006] The evidence identification model is used to conduct a three-fold assessment of the authenticity, legality, and relevance of structured evidence in the legal case, and to obtain the results of the three-fold assessment, as well as to identify contradictory evidence and determine the acceptance of the evidence.

[0007] The fact-finding model is used to infer and ascertain the points of contention based on the structured evidence, the results of the three-fold assessment, and the results of acceptance, so as to generate a conclusion on the points of contention.

[0008] The applicable legal provisions for the legal case are obtained by using a legal provision screening model based on the points of contention and the conclusions reached regarding those points of contention.

[0009] The content ascertained during the trial is generated using a content generation model based on the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three aspects (relevance, authenticity, and legality), and the results of the acceptance of the findings.

[0010] The content generation model is used to generate the court's opinion based on the content ascertained during the trial and the applicable legal provisions;

[0011] The content generation model is used to generate judgment items based on the content ascertained during the trial, the content held by the court, the applicable legal provisions, and the litigation claims.

[0012] The judgment for the legal case is generated based on the findings of the trial, the opinion of this court, and the judgment items.

[0013] Optionally, the process by which the fact generation model determines the focus of the dispute includes:

[0014] Calculate the semantic similarity between different disputed facts;

[0015] The disputed facts are clustered according to a preset semantic similarity threshold and the semantic similarity, so as to merge semantically related disputed facts into the same disputed cluster, resulting in one or more disputed clusters;

[0016] The focus of the dispute is determined based on one or more of the dispute clusters.

[0017] Optionally, the method further includes:

[0018] Based on the cause of action of the aforementioned legal cases, the legal provisions in the legal provisions database are filtered to obtain first-level candidate legal provisions;

[0019] The legal provision screening model is used to screen legal provisions based on the points of contention and the conclusions reached regarding those points of contention, thereby obtaining the applicable legal provisions for the legal case, including:

[0020] The primary candidate legal provisions, the points of contention, and the conclusion of the point of contention are input into the legal provision screening model. The legal provision screening model determines the legal relationship of the legal case based on the conclusion of the factual determination. It then selects secondary candidate legal provisions from the primary candidate legal provisions that match the legal relationship and determines the secondary candidate legal provisions required to adjudicate the points of contention as the applicable legal provisions.

[0021] Optionally, after generating the judgment content, the content generation model performs legal consistency verification, factual closure verification, and procedural compliance verification on the judgment content; and outputs the judgment content after the judgment content passes the legal consistency verification, the factual closure verification, and the procedural compliance verification.

[0022] Optionally, the original evidence in the legal case is multimodal evidence; the method further includes:

[0023] The original multimodal evidence is input into the multimodal model, which performs semantic extraction on the multimodal evidence to obtain the semantic information corresponding to each multimodal evidence. The semantic information of each multimodal evidence is then semantically aligned based on a preset format to output the structured evidence.

[0024] Optionally, when the evidence identification model performs the three-fold identification of authenticity, legality, and relevance of the structured evidence in the legal case, it also performs cross-modal consistency verification on the structured evidence to identify temporal, logical, and / or monetary contradictions between different modal evidence, obtain consistency verification results, and combine the consistency verification results to perform the three-fold identification.

[0025] Optionally, the fact generation model is also used to generate a corresponding focus tree for each point of contention, wherein the root node of the focus tree is the point of contention, and the child nodes are the disputed facts corresponding to the point of contention;

[0026] The process by which the fact-finding model generates the focus determination conclusion of the disputed focus includes: after reasoning and identifying the disputed focus, generating the focus determination conclusion based on the focus tree.

[0027] Optionally, the method further includes:

[0028] The results of the acceptance, the conclusions on the determination of the focus, and one or more of the applicable legal provisions are displayed through a visual interface.

[0029] If the presented acceptance results are updated, the fact-finding model is reused to infer and ascertain the points of contention based on the structured evidence, the three-fold assessment results, and the updated acceptance results, so as to generate a new conclusion on the points of contention.

[0030] If the focus determination conclusion is updated, the legal provision screening model is reused to determine the new applicable legal provision for the legal case based on the updated focus determination conclusion.

[0031] If any input to the content generation model is updated, the updated data is re-inputted into the content generation model to regenerate the judgment.

[0032] Optionally, the process by which the fact-finding model infers and ascertains the points of contention based on the structured evidence, the results of the three-fold verification, and the acceptance results, to generate a conclusion on the points of contention, includes:

[0033] Based on the structured evidence, the results of the three-fold assessment, and the acceptance results, determine whether the conclusion on the focus of the assessment can be inferred;

[0034] In the absence of a conclusion on the identified focus, multiple investigation plans should be developed for the disputed focus.

[0035] Based on the structured evidence, the results of the three-dimensional identification, and the results of acceptance, each investigation plan is inferred to obtain the corresponding investigation plan results;

[0036] The focus identification conclusion is determined based on the results of the investigations corresponding to the multiple investigation plans.

[0037] Optionally, the method further includes:

[0038] The investigation plan, the investigation results, and the focus identification conclusion are displayed in the form of a mind map through a visual interface. The focus of the dispute is the main node of the mind map, the investigation plan is a child node of the main node, the investigation results are child nodes of the corresponding investigation plan, and the focus identification conclusion is a summary node of the multiple investigation results.

[0039] Optionally, the fact generation model is further used to determine the dependency relationship between the first and second points of contention when there are multiple points of contention; the fact determination model is further used to, under the premise that the first point of contention is the second point of contention, first perform reasoning and investigation on the first point of contention to obtain the first point of contention determination conclusion, and determine the second point of contention determination conclusion based on the first point of contention determination conclusion.

[0040] According to a second aspect of one or more embodiments of this specification, a judgment generation system is provided, the system comprising:

[0041] The fact generation module is used to input the complaint, answer, and court transcript of a legal case into the fact generation model. The fact generation model uses entity recognition and relation extraction to perform structured processing on the facts of the legal case, generate the disputed and undisputed facts involved in the legal case, and determine the focus of the dispute.

[0042] The evidence identification module is used to use the evidence identification model to identify the authenticity, legality, and relevance of the structured evidence in the legal case, obtain the results of the three-fold identification, and identify contradictory evidence and determine the acceptance result.

[0043] The fact-finding module is used to use a fact-finding model to reason and ascertain the points of contention based on the structured evidence, the results of the three-fold verification, and the results of acceptance, so as to generate a conclusion on the point of contention.

[0044] The legal provision screening module is used to screen legal provisions based on the points of contention and the conclusions of the point of contention using a legal provision screening model, so as to obtain the applicable legal provisions for the legal case.

[0045] The content generation module is used to generate, using a content generation model, the ascertained facts of the trial based on the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three elements (relevance, authenticity, and legality), and the results of acceptance; to generate the court's opinion based on the ascertained facts of the trial and the applicable legal provisions using the content generation model; to generate judgment items based on the ascertained facts of the trial, the court's opinion, the applicable legal provisions, and the claims; and to generate a judgment for the legal case based on the ascertained facts of the trial, the court's opinion, and the judgment items.

[0046] According to a third aspect of one or more embodiments of this specification, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described above by executing the executable instructions.

[0047] According to a fourth aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method described above.

[0048] According to a fifth aspect of one or more embodiments of this specification, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method as described above.

[0049] As can be seen from the above embodiments, by adopting the technical solution provided in this specification, the judge's trial thinking is decomposed into a chain-based verifiable trial process of "fact structuring - identification of points of contention - evidence analysis - application of law - judgment generation," thus aligning the artificial intelligence reasoning path with the legal argumentation logic. The inputs and outputs of each stage are visualized and structured, making the entire judgment generation process "visible," with clear decision-making basis and traceable reasoning steps, thus constructing a controllable, reliable, rigorous, universal, and transparent intelligent trial mechanism. Attached Figure Description

[0050] Figure 1 This is a flowchart of a judgment generation method provided in an exemplary embodiment.

[0051] Figure 2 This is a flowchart of a process for determining the focus of a dispute, provided in an exemplary embodiment.

[0052] Figure 3 This is a flowchart of a focus determination conclusion generation process provided in an exemplary embodiment.

[0053] Figure 4 This is a schematic diagram of a focus determination conclusion display page provided in an exemplary embodiment.

[0054] Figure 5 This is a flowchart of an applicable legal provision screening process provided in an exemplary embodiment.

[0055] Figure 6 This is a schematic diagram of the structure of a device provided in an exemplary embodiment.

[0056] Figure 7 This is a block diagram of a judgment generation system provided in an exemplary embodiment. Detailed Implementation

[0057] In judicial practice, judges must manually read case files, analyze evidence, apply the law, and draft judgments. This process is highly dependent on personal experience and intellectual labor, faces efficiency bottlenecks, and is easily influenced by subjective factors. To improve the quality and efficiency of trials, intelligent technologies have emerged, and their development has mainly gone through four stages: document template-based trials, element-based trials, knowledge graph applications, and direct generation of large models. However, these methods still have certain limitations and cannot meet the high-level requirements of judgment quality, logical rigor, and procedural interpretability in practice.

[0058] The document template stage uses a pre-set judgment template to generate a judgment through keyword replacement. While technically simple, it lacks intelligence and is only suitable for simple cases with highly similar facts, lacking targeted analysis and reasoning for specific cases.

[0059] The element-based trial stage can predefine key facts and legal elements for specific causes of action (such as road traffic disputes and labor disputes), and extract elements from structured materials using rule matching algorithms and fill them into the judgment template. This method achieves a degree of automation, but it has significant limitations: its applicability is strictly limited to the pre-defined element system and causes of action, making it unsuitable for complex, novel, or non-standardized cases; the element identification method is rigid, lacking semantic understanding and adaptive capabilities; and it requires the separate development and maintenance of templates for each cause of action, resulting in poor system scalability and high maintenance costs.

[0060] In the application phase of knowledge graphs, a specialized legal knowledge graph is constructed, modeling legal concepts, case elements, and legal relationship relationships as a graph structure. Graph algorithms are then used for reasoning to assist in generating judgments. Compared to element-based trials, it possesses a certain degree of logical reasoning capability. However, this technical approach has drawbacks: the construction of the knowledge graph heavily relies on manual annotation and knowledge injection by domain experts, resulting in a long construction cycle and extremely high initial and subsequent maintenance costs; its reasoning ability is entirely limited by the pre-set rules and paths within the graph, making it difficult to handle novel cases or complex logical relationships not covered by the graph, and resulting in insufficient system flexibility and creativity.

[0061] The large-scale model's direct generation stage allows case materials to be directly input into the model, resulting in a complete judgment. However, this approach also has limitations: Firstly, the reasoning process is uncontrollable and lacks transparency, making it impossible to ensure the model strictly adheres to the trial logic of "fact finding—legal application." Secondly, the inherent "illusion" phenomenon of the model can easily lead to generated content that does not match the facts of the case, posing a risk of fact-finding errors. Furthermore, it is difficult to systematically guarantee that the judgment conforms to rigorous reasoning paradigms such as legal syllogisms; the legal logic chain is weak, failing to guarantee the persuasiveness and authority of the judgment.

[0062] To address the aforementioned issues, this specification provides a judgment generation scheme. By breaking down the judge's judicial thinking into a chain-like, verifiable process of "fact structuring - identification of points of contention - evidence analysis - application of law - judgment generation," it aligns the AI ​​reasoning path with legal argumentation logic. The inputs and outputs of each stage are visualized and structured, making the entire judgment generation process "visible," with clear decision-making basis and traceable reasoning steps. This constructs a controllable, reliable, rigorous, universal, and transparent intelligent auxiliary judicial mechanism.

[0063] The judgment generation scheme provided in this manual can be applied to electronic devices, such as PCs, mobile phones, tablets, laptops, and PDAs (Personal Digital Assistants).

[0064] The judgment generation scheme provided in this manual can also be applied to either a Client-Server (CS) architecture or a Browser-Server (BS) architecture. Taking the CS architecture as an example, the client can upload files such as the complaint, answer, evidence, and court transcripts of a legal case. The server then generates a chain of judgments based on this data. Intermediate results obtained during the chain generation process, such as the results of the determination of the admissibility of evidence, the results of the acceptance of evidence, and the conclusions on the key points of contention, as well as the generated complete judgment, can be returned to the client for the user to browse / adjust. This manual does not impose any special limitations on this.

[0065] Figure 1 This is a flowchart of a judgment generation method provided in an exemplary embodiment.

[0066] Please refer to Figure 1 The method for generating the judgment may include the following steps:

[0067] Step 101: Input the complaint, answer, and court transcript of the legal case into the fact generation model. Use the fact generation model to perform structured processing on the facts of the legal case through entity recognition and relation extraction to generate the disputed and undisputed facts involved in the legal case and determine the focus of the dispute.

[0068] In some embodiments, the fact generation model may be a large language model, such as any published large language model. This specification does not impose any special restrictions on its base model, model size, etc.

[0069] In some embodiments, the complaint, answer, and court transcript of a legal case requiring judgment can be input into the fact generation model. The fact generation model can identify basic information about the legal case from the complaint, answer, and court transcript through entity recognition, such as party information, claims, and facts stated by each party. Next, the fact generation model can perform deep semantic analysis and relationship extraction to identify key entities and relationships between entities in the legal case, such as time, place, people, amounts, and actions (e.g., A pays B amount Y at time X). The identified entities and relationships can be structured and organized according to a preset template to generate a fact list for the legal case. Furthermore, through semantic analysis, disputed facts and undisputed facts can be distinguished.

[0070] In some embodiments, the fact generation model may also identify the focus of the dispute in the legal case through techniques such as clustering. The focus of the dispute is usually expressed in the form of legal points of contention, such as "whether the loan agreement is valid" or "whether liability for breach of contract exists".

[0071] Step 102: Using the evidence identification model, the structured evidence in the legal case is assessed for authenticity, legality, and relevance to obtain the results of the three-factor assessment, and contradictory evidence is identified and the results of acceptance are determined.

[0072] In some embodiments, the evidence identification model may also be a large language model, which may be the same large language model as the aforementioned fact generation model, or it may be a different large language model. This specification does not impose any special restrictions on this.

[0073] In some embodiments, the structured evidence of the legal case can be input into the evidence identification model, which then performs a three-fold assessment of the structured evidence for authenticity, legality, and relevance to obtain the assessment results. It can also identify contradictory evidence and determine the acceptance of contradictory evidence.

[0074] Structured evidence can be obtained by structuring the original multimodal evidence of the legal case using a multimodal model, which will be described in detail in subsequent embodiments.

[0075] Step 103: Using the fact-finding model, the disputed issues are inferred and identified based on the structured evidence, the results of the three-fold determination, and the results of acceptance, so as to generate a conclusion on the determination of the disputed issues.

[0076] In some embodiments, after completing the "three-fold" assessment (legality, authenticity, and relevance) and the judgment of the admissibility of the structured evidence in the legal case (i.e., the admissibility result), a legal fact-finding based on the points of contention can be carried out. Specifically, a fact-finding model can be used to transform the evidence into a legal fact-finding conclusion with adjudicative significance (i.e., the conclusion on the point of contention) for each specific point of contention through legal logical reasoning.

[0077] In some embodiments, the fact-finding model may also be a large language model, which may be the same large language model as the aforementioned fact-generating model, the same large language model as the aforementioned evidence-finding model, or a large language model that is different from both the aforementioned fact-generating model and the aforementioned evidence-finding model. This specification does not impose any special restrictions on this.

[0078] In some embodiments, the structured evidence of the legal case, the results of the three-fold assessment, the results of acceptance, and the points of contention can be input into the evidence assessment model. The fact-finding model then uses the structured evidence, the results of the three-fold assessment, and the results of acceptance to reason and ascertain each point of contention, thereby generating a conclusion on the point of contention.

[0079] The conclusion regarding the focus of the dispute may include the established facts and the reasons for the determination. The established facts may be specific legal facts related to the focus of the dispute, as ascertained by the court based on accepted evidence. The reasons for the determination may be the evidence and reasoning upon which the determination is based, for example: "Based on… evidence, it is sufficient to establish…", etc.

[0080] Step 104: Using the legal provision screening model, legal provisions are screened based on the points of contention and the conclusions of the point of contention to obtain the applicable legal provisions for the legal case.

[0081] In some embodiments, when determining the applicable legal provisions for a legal case, the provisions in a pre-built legal provisions database can be initially screened based on the cause of action of the legal case to obtain first-level candidate provisions. Then, the first-level candidate provisions, the points of contention, and the conclusion of the point of contention determination are input into the legal provisions screening model, which continues to screen the provisions to obtain the applicable legal provisions for the legal case.

[0082] In some embodiments, the legal provision screening model may also be a large language model, which may be the same large language model as the aforementioned fact generation model, the same large language model as the aforementioned evidence determination model, or the same large language model as the aforementioned fact determination model. Alternatively, it may be a large language model that is different from the aforementioned fact generation model, the aforementioned evidence determination model, and the aforementioned fact determination model. This specification does not impose any special restrictions on this.

[0083] Step 105: Using a content generation model, generate the content to be ascertained during the trial based on the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three aspects, and the results of the acceptance of the findings.

[0084] Step 106: Using the content generation model, generate the content that the court believes based on the content ascertained during the trial and the applicable legal provisions.

[0085] Step 107: Using the content generation model, generate judgment content based on the content ascertained during the trial, the content held by the court, the applicable legal provisions, and the litigation claims.

[0086] Step 108: Generate the judgment for the legal case based on the findings of the trial, the opinion of the court, and the judgment items.

[0087] In some embodiments, the content generation model can be used to generate the trial findings, the court's opinion, and the judgment items in the judgment, and these contents can be combined into a complete judgment.

[0088] In some embodiments, the content generation model may also be a large language model, which may be the same large language model as the aforementioned fact generation model, the same large language model as the aforementioned evidence determination model, the same large language model as the aforementioned fact determination model, or the same large language model as the aforementioned legal provision screening model. Alternatively, it may be a large language model that is different from the aforementioned fact generation model, the aforementioned evidence determination model, the aforementioned fact determination model, and the aforementioned legal provision screening model. This specification does not impose any special restrictions on this.

[0089] In some embodiments, the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three aspects (relevance, admissibility, and legality), and the results of acceptance can be input into the content generation model first. The content generation model is then instructed to output the findings of the trial. Next, the findings of the trial and the applicable legal provisions are input into the content generation model, which is then instructed to generate the court's opinion. Then, the findings of the trial, the court's opinion, the applicable legal provisions, and the claims are input into the content generation model, which is then instructed to generate the judgment. The generated judgment can also be logically verified. If the verification is correct, the generated judgment can be output. Finally, the findings of the trial, the court's opinion, and the judgment can be combined according to a judgment template to generate a complete judgment for the legal case.

[0090] As can be seen from the above description, by adopting the technical solution provided in this specification, and decomposing the judge's trial thinking into a chain-like verifiable trial process of "fact structuring - identification of points of contention - evidence analysis - application of law - judgment generation," the alignment of artificial intelligence reasoning paths with legal argumentation logic is achieved. The inputs and outputs of each stage are visualized and structured, making the entire judgment generation process "visible," with clear decision-making basis and traceable reasoning steps, thus constructing a controllable, reliable, rigorous, universal, and transparent intelligent auxiliary trial mechanism.

[0091] The following sections, using specific examples, will detail the implementation process of this manual from five aspects: fact extraction and determination of points of contention, evidence analysis, fact determination, selection of applicable legal provisions, and generation of judgments.

[0092] I. Fact Extraction and Determination of Points of Contention

[0093] In some embodiments, before inputting the complaint, answer, and court transcript of a legal case into the fact generation model, the complaint, answer, and court transcript can be preprocessed. For example, the complaint, answer, and court transcript can be input into a multimodal model for preprocessing. Specifically, the multimodal model can remove redundant information from the above documents and perform format normalization, word segmentation, etc.

[0094] In some embodiments, pre-processed complaints, answers, and court transcripts are input into the fact generation model, which extracts basic information about the legal case from these documents, such as party information, claims, and facts stated by each party. The fact generation model then performs deep semantic analysis and relationship extraction to identify key entities and relationships between them, such as time, place, task, amount, and action (e.g., A pays B amount Y at time X). The identified entities and relationships are then structured according to a preset template to generate a fact list for the legal case. Furthermore, semantic analysis can further distinguish between disputed and undisputed facts.

[0095] Each fact entry in the fact list may include: fact ID, fact content, fact dispute, and source (complaint, answer, or court transcript). For example, an undisputed fact entry may include: fact ID-F002, fact content-the plaintiff and defendant are friends and have a long-term economic relationship, and source-court transcript.

[0096] In some embodiments, the fact generation model can output a list of disputed facts consisting of extracted disputed facts and a list of undisputed facts consisting of extracted undisputed facts. These lists can then be displayed through a visual interface for judges to browse. Of course, judges can also adjust the fact lists. If a judge believes that a certain extracted fact is inaccurate, they can modify and adjust it, and the modified fact list can be saved. Furthermore, the modified fact list can be fed back to the fact generation model for learning and optimization.

[0097] In some embodiments, the fact generation model may also perform semantic analysis on disputed facts to identify the core points of contention in a legal case, i.e. the focus of the dispute, and the disputed facts associated with the focus of the dispute.

[0098] Please refer to Figure 2 The process by which a fact generation model identifies the focus of a dispute may include the following steps:

[0099] Step 201: Calculate the semantic similarity between different disputed facts.

[0100] In some embodiments, the disputed facts may be semantically encoded and their similarity calculated in a unified vector space, such as cosine similarity or Euclidean distance, and this specification does not impose any special limitations on this.

[0101] Step 202: Cluster the disputed facts according to the preset semantic similarity threshold and the semantic similarity, so as to merge semantically related disputed facts into the same disputed cluster, and obtain one or more disputed clusters.

[0102] In some embodiments, based on a preset semantic similarity threshold and the semantic similarity between different disputed facts calculated above, each disputed fact can be clustered, and semantically related disputed facts can be merged into the same disputed cluster to obtain one or more disputed clusters.

[0103] The semantic similarity threshold can be preset, for example, 85%, 80%, etc.

[0104] Step 203: Determine the focus of the dispute based on the one or more dispute clusters.

[0105] In some embodiments, through the aforementioned clustering, a dispute cluster can be obtained, that is, a dispute focus, and the disputed facts included in the dispute focus are the disputed facts in the dispute cluster.

[0106] In some embodiments, multiple dispute clusters can be obtained through the aforementioned clustering, and each dispute cluster can correspond to a dispute focus, that is, multiple dispute focuses are obtained, and the disputed facts included in each dispute focus are the disputed facts in the corresponding dispute cluster.

[0107] For example, suppose the legal case is as follows: Plaintiff Zhang San sues Defendant Li Si, demanding the return of a loan of 15,000 yuan. Plaintiff Zhang San's claims to this court are: an order for the defendant to repay the principal of the loan of 15,000 yuan; an order for the defendant to pay overdue interest (calculated on the basis of 15,000 yuan, from January 1, 2025, with reference to the then one-year loan prime rate until the date of actual repayment); and an order for the defendant to bear all litigation costs, including litigation fees and preservation fees.

[0108] In this legal case, two points of contention can be identified: the first is whether the loan agreement is valid; the second is whether the defendant has repaid part or all of the loan.

[0109] Therefore, by adopting the above-mentioned scheme provided in this specification, the fact generation model can be used to extract structured facts of legal cases by identifying entities and extracting relationships from the complaint, answer, and court transcript, and to determine the focus of the dispute, providing an interpretable factual basis for subsequent processes. At the same time, it can also avoid the omission or improper attribution of the focus of the dispute due to human oversight.

[0110] In some embodiments, when there are multiple identified points of contention, the fact generation model can also rank the importance of each point of contention based on the number of facts contained in each point of contention and its legal impact, thereby determining the dependency relationship between different points of contention and improving the efficiency of subsequent point of contention determination.

[0111] Taking the first point of contention as "whether the loan agreement is valid" and the second point of contention as "whether the defendant has repaid part or all of the loan" as an example, it can be determined that the first point of contention is the most important, and the dependency relationship between the two is that the first point of contention is a prerequisite for the second point of contention.

[0112] In some embodiments, for each point of contention, the fact generation model can also generate a corresponding focus tree, where the root node of the focus tree is the point of contention and the child nodes are the disputed facts corresponding to the point of contention, so as to facilitate the generation of a focus determination conclusion after the reasoning and investigation of the point of contention is completed.

[0113] In this specification, during the process of fact extraction and determination of points of contention, the generation of disputed and undisputed facts, as well as the determination of points of contention, can be achieved by inputting corresponding prompt words into the fact generation model.

[0114] II. Evidence Analysis

[0115] In some embodiments, the original evidence related to a legal case may be multimodal evidence, such as: loan agreements as textual evidence, bank transfer screenshots as image evidence, telephone recordings as audio evidence, and surveillance footage as video evidence. This multimodal original evidence can be input into a multimodal model, which performs semantic extraction on it to obtain semantic information corresponding to each piece of multimodal evidence. Based on a preset format, the semantic information of each piece of multimodal evidence is semantically aligned to output structured evidence.

[0116] The multimodal model can be a dedicated model used to structure data from multiple modalities, specifically converting evidence from multiple modalities into a text structure.

[0117] For example, suppose the textual evidence is a loan agreement stating "loan of 100,000 yuan, May 10, 2023"; the image evidence is a bank transfer screenshot showing "80,000 yuan transferred on May 10, 2023"; the audio evidence is a phone recording in which Li Si says "I repaid 20,000 yuan on May 15"; and the video evidence is surveillance footage showing Li Si visiting the bank on May 15. The multimodal model can extract information such as time and amount from these pieces of evidence and perform semantic alignment processing, resulting in the structured evidence shown in Table 1 below:

[0118] Table 1

[0119]

[0120] In some embodiments, structured evidence from legal cases can be input into an evidence assessment model, which then performs a three-fold assessment of authenticity, legality, and relevance to obtain the assessment results. During this assessment, the evidence assessment model can perform cross-modal consistency verification on the structured evidence to identify temporal, logical, and / or monetary inconsistencies between different modalities of evidence, obtaining consistency verification results, and combining these results with the three-fold assessment.

[0121] Taking the aforementioned four types of multimodal evidence as examples, consistency verification can reveal consistent parts, such as: both the loan agreement and the transfer screenshot mention a loan date of May 10th, indicating a consistent timeframe. Consistency verification can also reveal contradictory parts, such as: the loan agreement states a loan of 100,000, but the transfer screenshot shows "80,000 transferred," a discrepancy in amount; a phone recording mentions "20,000 repaid," but the transfer screenshot shows a loan of 80,000, a logical contradiction. The consistency verification result output by the evidence identification model can include the verification status: partial contradiction, and can also include contradiction details such as "loan amount inconsistency: contract 100,000 vs. transfer 80,000," etc. The consistency verification result can also include suggestions for addressing contradictions such as "the actual loan amount needs to be verified," and can also output a consistency score, etc., without special restrictions.

[0122] In some embodiments, the evidence identification model may use the consistency verification results as a supplement to perform three-dimensional verification of structured evidence in conjunction with the consistency verification results, identify contradictory evidence, and generate acceptance results for contradictory evidence.

[0123] In some embodiments, the results of the three-fold determination of evidence, contradictory evidence, and acceptance results output by the evidence determination model can be displayed through a visual interface for judges to browse. Specifically, contradictory parts, such as details of contradictions in the consistency verification results, can be highlighted to differentiate them, allowing judges to quickly identify key contradictions between pieces of evidence and significantly improve trial efficiency.

[0124] Of course, if the judge believes that one or more of the three-fold determination results, contradictory evidence, and acceptance results output by the model are problematic, they can modify and adjust them, and the judge's modified content can be saved. Furthermore, the judge's modified content can be fed back to the evidence determination model for learning and optimization.

[0125] Therefore, by adopting the above-mentioned scheme provided in this specification, original evidence of different modalities such as text, images, audio, and video can be uniformly encoded into the same vector space through a multimodal model, realizing the construction of evidence chains and consistency verification of cross-modal evidence. Compared with the existing technology that mainly processes text evidence, this represents a breakthrough in the ability to process multimodal evidence.

[0126] In this specification, during the above evidence analysis process, the evidence identification model can be used to perform the three-fold identification of evidence in the above legal cases, consistency verification, identification of contradictory evidence, and determination of the acceptance result by inputting corresponding prompt words into the evidence identification model.

[0127] III. Fact Determination

[0128] In some embodiments, after the admissibility, authenticity, and legality of evidence are determined and accepted, the investigation of the legal facts concerning the points of contention can proceed. Specifically, based on the established evidentiary foundation, each point of contention can be evaluated and reasoned to generate a legally definitive conclusion on the determination of the points of contention. For example, the structured evidence of the legal case, the results of the admissibility, authenticity, and legality determination, the results of acceptance, and the points of contention can be input into a fact-finding model. The fact-finding model then reasons and investigates the points of contention to generate a conclusion on the determination of the points of contention for each point of contention, i.e., a legal fact-finding conclusion for each point of contention.

[0129] Please refer to Figure 3 The process by which the fact-finding model generates a focus determination conclusion may include the following steps:

[0130] Step 301: Determine whether a conclusion on the focus of the investigation can be inferred based on the structured evidence, the results of the three-dimensional assessment, and the results of acceptance.

[0131] In some embodiments, the fact-finding model can determine whether it can directly deduce the conclusions regarding the key issues of each dispute based on the structured evidence, the results of the three-fold assessment, and the acceptance results. Generally, for some relatively simple issues of dispute, the fact-finding model can directly deduce the corresponding conclusions regarding the key issues.

[0132] For example, the central issue in the dispute is whether the lease agreement continued until December 31, 2021. Specifically, the plaintiff claims that the lease agreement continued until December 31, 2021, but the plaintiff has not provided any valid evidence to prove this. The fact-finding model can directly determine that the plaintiff, as the party bearing the burden of proof, failed to provide valid evidence to prove this factual claim and should bear the consequences of failing to provide evidence. Therefore, the plaintiff's claim that the lease agreement continued until December 31, 2021 lacks factual basis and is not supported.

[0133] In some embodiments, certain points of contention are complex and the fact-finding model cannot directly deduce the corresponding conclusions regarding these points of contention; in such cases, step 302 can be continued.

[0134] Step 302: If the conclusion on the focus of the dispute cannot be deduced, formulate multiple investigation plans for the focus of the dispute.

[0135] Step 303: Based on the structured evidence, the results of the three-indication assessment, and the acceptance results, deduce each investigation plan to obtain the corresponding investigation plan results.

[0136] In some embodiments, when the fact-finding model cannot directly deduce the conclusion of the focus of the dispute, it can formulate multiple investigation plans for the focus of the dispute, that is, split the focus of the dispute into multiple investigation plans that need to be investigated, and then reason about each investigation plan according to the structured evidence, the results of the three-dimensional assessment, and the results of acceptance, so as to obtain the corresponding investigation results of the plan, thereby realizing the split reasoning of complex points of dispute.

[0137] Among these, disputes where a conclusion cannot be directly deduced are often complex issues involving insufficient evidence, unclear facts, significant contradictions between pieces of evidence, or requiring professional judgment. For example, suppose a dispute concerns the validity of a back-to-back clause and whether the payment conditions have been met. This can be broken down into three investigation plans: Investigation Plan 1: Has the general contract been settled? Investigation Plan 2: Has the defendant received the corresponding payment from the owner? Investigation Plan 3: The validity of the back-to-back clause and whether the defendant failed to assert its rights with the owner.

[0138] Step 304: Determine the focus identification conclusion based on the plan identification results corresponding to the multiple investigation plans.

[0139] In some embodiments, the conclusion on the determination of the focus of the dispute can be determined based on the results of the planned investigations of multiple investigation plans corresponding to the focus of the dispute. For example, regarding the focus of the dispute "the validity of the back-to-back clause and whether the payment conditions have been met", the corresponding conclusion on the determination of the focus can be determined based on the results of the planned investigations of the above three investigation plans: the back-to-back clause is valid, and the payment conditions are considered to have been met.

[0140] In some embodiments, after reasoning is completed, the focus determination conclusion can be generated based on the focus tree corresponding to the focus of the dispute and the planned ascertainment results of the multiple ascertainment plans. Specifically, the focus tree includes the disputed facts corresponding to the focus of the dispute, and specific explanations of whether the claims in the focus determination conclusion are valid can be generated based on these disputed facts.

[0141] As can be seen from the above description, when the fact-finding model cannot directly deduce the conclusion on the focus of the dispute, it can formulate multiple investigation plans for the disputed focus, deduce the investigation results of each plan separately, and then determine the conclusion on the focus based on the investigation results of multiple plans. This process, for complex disputes, uses a staged reasoning approach to determine the conclusion on the focus, significantly improving the accuracy of legal fact-finding conclusions.

[0142] In some embodiments, when the dependency relationship between the first and second points of contention is a premise that the first point of contention is the second point of contention, the fact-finding model first infers and ascertains the first point of contention to obtain the first point of contention determination conclusion, and then determines the second point of contention determination conclusion based on the first point of contention determination conclusion.

[0143] Taking the first point of contention as "whether the loan contract is valid" and the second point of contention as "whether the defendant has repaid part or all of the loan" as an example, if the conclusion of the first point of contention is "the contract is invalid", since the first point of contention is a premise of the second point of contention, and since the premise is no longer valid, there is no need to reason about the second point of contention. The fact-finding model can directly output the conclusion of the second point of contention. For example: given that the contract is invalid, there is no need to find out whether the defendant has repaid part or all of the loan.

[0144] Of course, in other cases, if the conclusion of the first point of contention is that "the contract is valid", the fact-finding model can continue to reason about the second point of contention to generate the corresponding conclusion.

[0145] In this specification, during the process of determining the focus of the aforementioned dispute, the conclusion of the focus determination can be achieved by inputting corresponding prompt words into the fact-finding model.

[0146] In some embodiments, the investigation plan, the investigation results, and the focus determination conclusion are displayed in the form of a mind map through a visual interface for judges to browse. The focus of the dispute is the main node of the mind map, the investigation plan is a child node of the main node, the investigation results are child nodes of the corresponding investigation plan, and the focus determination conclusion is a summary node of the multiple investigation results.

[0147] Taking the aforementioned point of contention, "the validity of back-to-back clauses and whether the payment conditions have been met," as an example, please refer to... Figure 4 In the example, the point of contention is the main node, the three investigation plans are its child nodes, the child nodes of each investigation plan are its investigation results, and the conclusion of the point of contention determination is the summary node of the investigation results of these three plans.

[0148] Optionally, the amount of evidence required for the reasoning and further investigation can be indicated at each investigation plan section, such as... Figure 4 The number "5" displayed after the investigation plan 3: the validity of the back-to-back clause and whether the defendant failed to assert its rights to the owners is the number of pieces of evidence required to deduce this investigation plan. When the judge clicks on this number, the corresponding evidence can be displayed through pop-up windows or other means.

[0149] IV. Selection of Applicable Legal Provisions

[0150] Please refer to Figure 5 The process of selecting applicable legal provisions may include the following steps:

[0151] Step 501: Based on the cause of action of the legal case, the legal provisions in the legal provisions database are screened to obtain the first-level candidate legal provisions.

[0152] In some embodiments, the legal provisions database can be pre-built, storing mapping relationships between legal provisions and causes of action. When filtering applicable legal provisions, the database can be searched first based on the cause of action of the legal case, and the found provisions can be used as primary candidate provisions. The cause of action may be derived from court hearing transcripts.

[0153] Step 502: Input the first-level candidate legal provisions, the points of contention, and the conclusion of the point of contention into the legal provision screening model. The legal provision screening model determines the legal relationship of the legal case based on the conclusion of the factual determination. It then selects second-level candidate legal provisions that match the legal relationship from the first-level candidate legal provisions and determines the second-level candidate legal provisions required to adjudicate the points of contention as the applicable legal provisions.

[0154] In some embodiments, the selected primary candidate legal provisions, points of contention, and conclusions on the determination of points of contention can be input into a legal provision screening model. The legal provision screening model can first identify the legal relationship of the legal case based on the points of contention and the conclusions on the determination of points of contention. The legal relationship can be composed of core elements. Then, based on the legal relationship, the aforementioned primary candidate legal provisions can be filtered to select secondary candidate legal provisions that match the legal relationship.

[0155] Assuming the core elements to be identified include: loan agreement (IOU), loan delivery (Alipay transfer), borrower's overdue payment, claim for principal + overdue interest + fees, etc., these core elements can determine that the parties in the case are all natural persons, belonging to "loan agreements between natural persons". This allows filtering out irrelevant or less relevant clauses, such as: pre-deduction of interest clauses (Article 670, which does not exist in this case), and the remaining first-level candidate legal provisions can be identified as second-level candidate legal provisions.

[0156] In some embodiments, the legal provision screening model can also perform reverse reasoning on the selected secondary candidate legal provisions to determine whether the selected secondary candidate legal provisions are correct, and to determine the secondary candidate legal provisions to be cited in adjudicating the focus of the dispute as the applicable legal provisions for the legal case.

[0157] Taking the aforementioned dispute as an example, based on the focus of the dispute, we can extract some factual elements that the ruling needs to include, such as: whether the defendant should repay the principal, whether the plaintiff's claim for overdue interest is legal, and whether the starting point and standard for calculation are reasonable. Based on these factual elements, we can determine whether the selected secondary candidate legal provisions are accurate. Taking whether the defendant should repay the principal as an example, it is necessary to prove that the loan contract was established and that the plaintiff has fulfilled its payment obligations, corresponding to Article 679 of the Criminal Law.

[0158] Through the above process, the applicable legal provisions can be finally selected as: Articles 675, 676, and 679 of the *** Law.

[0159] Therefore, the scheme provided in this manual firstly filters the legal provisions in the legal provisions database based on the cause of action of the legal case to obtain first-level candidate provisions, including all potentially relevant provisions, thus avoiding omissions due to a limited search scope and ensuring the comprehensiveness of the provision selection from the outset. Next, a second-level candidate provision is obtained by using a legal provision screening model based on the focus of the dispute and the conclusion of the focus determination. This effectively solves the problems of the illusion of a large model and errors in the content and direction of the legal provisions cited. Through model reasoning, the model can anchor the provisions that are indispensable and directly corresponding to resolving the core dispute of the legal case. Then, through reverse reasoning, the provisions required to adjudicate the focus of the dispute in the second-level candidate provisions are determined as the applicable provisions selected, ensuring the accuracy of the applied provisions. Thus, by determining the applicable provisions for the legal case through the above three-level screening, the accuracy, rigor, consistency, and efficiency of the application of law in the judgment can be effectively guaranteed.

[0160] In this specification, during the above-mentioned screening of applicable legal provisions, the screening of secondary candidate legal provisions and applicable legal provisions can be achieved by inputting corresponding prompt words into the legal provision screening model.

[0161] V. Generation of Judgment

[0162] In some embodiments, a content generation model can be used to generate the findings of the trial, the court's opinion, and the judgment items in a legal case judgment, thereby generating the final judgment.

[0163] In some embodiments, the undisputed facts, points of contention, conclusions on the determination of points of contention, results of the determination of the three elements and results of acceptance of the facts can be input into the content generation model, and the content generation model can be used to generate and output the trial findings in the judgment.

[0164] The content generation model first reiterates the points of contention and clarifies the scope of facts to be proven. Specifically, it can first review the identified points of contention (such as whether the contract was established, whether a breach of contract occurred, and the calculation of the amount of damages), and then clarify the specific facts to be proven under each point of contention, excluding facts that are not substantially related to the case or are not disputed by the parties. An example of the result could be: "The points of contention in this case are: 1. Whether a legally valid sales contract exists between the parties; 2. Whether the defendant has breached the contract by delaying delivery; 3. Whether the plaintiff's claimed losses are true and should be borne by the defendant."

[0165] Next, the content generation model can output the following example based on the results of the three-fold verification and the acceptance results: "The original 'Purchase and Sale Contract' submitted by the plaintiff has been examined by the defendant without objection. This court confirms its authenticity, legality, and relevance to the first point of contention, and accepts it as credible."

[0166] Then, the content generation model can output the following example based on the conclusion of the focus of the dispute: "Based on the delivery note, acceptance record and WeChat communication records between the two parties, it is sufficient to prove that the defendant's actual delivery date was later than 15 days after the contract, which constitutes a delay in performance."

[0167] Finally, the content generation model, in accordance with the requirements of judicial documents, generates paragraph-style outputs of the findings of the trial in neutral, accurate, and concise language. For example: "The court found that... On May 1, 2023, the plaintiff and defendant signed an 'Equipment Purchase and Sale Contract,' stipulating that the defendant should deliver the equipment before July 1, 2023. On June 28 of the same year, the plaintiff paid the full amount. It was found that the defendant did not complete the delivery until July 16, 2023, exceeding the agreed deadline by 15 days. During this period, the plaintiff repeatedly urged delivery, but the defendant did not raise any legitimate reason to defend itself..."

[0168] In some embodiments, the content ascertained during the trial, output by the content generation model, may be displayed to the judge for modification and adjustment.

[0169] In some embodiments, the findings of the legal case and the selected applicable legal provisions can be input again into the content generation model to generate the court's opinion in the judgment. Specifically, the content generation model can output the court's opinion according to the applicable legal provisions and the findings of the trial, following the structure of the court's opinion, for example: "The court holds that... In conclusion, the plaintiff's claims are legally justified and are supported by this court."

[0170] In some embodiments, the content output by the content generation model, which is considered by the court, may be displayed to the judge for modification and adjustment.

[0171] In some embodiments, the aforementioned findings of the trial, the court's opinion, the applicable legal provisions, and the claims can be input into the content generation model, which then generates the judgment items in the judgment. Specifically, the content generation model generates specific judgment items based on facts and reasoning. For example, if the claim is for repayment of the principal, the applicable legal provisions (i.e., the legal basis) are Articles 675 and 679 of the Criminal Law, and the factual support is an IOU and transfer vouchers, indicating that delivery has been completed and the deadline has expired, the judgment item "The defendant shall return the principal of RMB 15,000 to the plaintiff within ten days from the date this judgment becomes effective" can be generated.

[0172] In some embodiments, after generating the judgment content, the content generation model may further perform legal consistency verification, factual closure verification, and procedural compliance verification on the judgment content, and output the judgment content after the judgment content passes the legal consistency verification, the factual closure verification, and the procedural compliance verification.

[0173] Legal consistency verification typically refers to the fact that all aspects of the judgment are supported by explicit legal provisions, such as Articles 675, 676, and 679 of the Criminal Law and procedural laws. Factual closure verification may include monetary verification, such as whether the principal amount matches the Alipay transfer record and the amount on the IOU; for defenses without a statute of limitations, it is assumed that the statute of limitations has not expired. Procedural compliance verification usually involves verifying the litigation procedure; for example, for small claims proceedings, it should be noted that "one instance, final judgment"; for cases of defendant absence, it should be noted that "having refused to appear in court without justifiable reason after being summoned, this court renders a default judgment in accordance with the law," etc.

[0174] After the judgment content passes the legal consistency verification, factual closure verification, and procedural compliance verification, the judgment content can be output. Before outputting the judgment content or during its generation, the language can be optimized and polished using standard template sentences from judicial documents to ensure consistent subject designations (e.g., "Plaintiff Zhang San," "Defendant Li Si"), accurate timeframes (e.g., "from January 1, 2025 until the date of actual payment"), clear performance deadlines (e.g., "to be performed within ten days from the date this judgment becomes effective"), and no omissions of liability scope (e.g., "including but not limited to case acceptance fees, publication fees (if any), and preservation fees (if any)").

[0175] For example, the output judgment could include: In accordance with Articles 675, 676, and 679 of the Criminal Law, the judgment is as follows: 1. The defendant…, this judgment is final.

[0176] In some embodiments, the judgment content output by the content generation model can be displayed to the judge for modification and adjustment.

[0177] In some embodiments, a judgment for the legal case may be generated based on the findings of the trial, the court's opinion, and the judgment items.

[0178] In this specification, during the generation of the aforementioned judgment content, the content ascertained during the trial, the court's opinion, and the judgment items can be generated by inputting corresponding prompt words into the content generation model.

[0179] In some embodiments, a judgment can be generated using the methods described above in this specification. Then, one or more of the evidence acceptance results, key point determinations, and applicable legal provisions determined during the judgment generation process can be displayed to the judge through a visual interface, allowing the judge to review the specific reasoning process during judgment generation. If the judge feels that one or more of these items need modification or adjustment, they can make the corresponding modifications and adjustments, and then regenerate the judgment based on the adjusted content.

[0180] For example, when the acceptance results are adjusted and updated, the fact-finding model can be reused to infer and identify the points of contention based on structured evidence, the results of the three-dimensional assessment, and the updated acceptance results, so as to generate a new conclusion on the points of contention.

[0181] For example, when the conclusion on the focus of the case is adjusted or updated, the legal provision screening model is reused to determine the new applicable legal provision for the case based on the updated conclusion. This update can be done manually by the judge, or it can be a result of the judge manually updating the accepted findings, after which the fact-finding model re-determines the focus of the case based on the new accepted findings.

[0182] For example, if any input to the content generation model is updated, the updated data is re-inputted into the content generation model to generate a new judgment. The inputs to the content generation model may include undisputed facts, disputed facts, points of contention, conclusions on the determination of those points, results of the three-fold determination, results of acceptance, findings of the trial, the court's opinion, and judgment items.

[0183] In other words, in the process of generating a judgment using chain thinking in this instruction manual, if the input of any step is updated, the subsequent steps will be regenerated based on the updated data.

[0184] In some embodiments, during the generation of the aforementioned judgment, the model outputs can be presented to the judge sequentially in a chain-like manner. Once the judge confirms the outputs, the enforcement process can proceed. For example, after the judge confirms the admissibility and acceptance results of the evidence determination model, the fact-finding model can then be used to generate the key findings conclusions.

[0185] As can be seen from the above description, the judgment generation scheme provided in this specification has the following beneficial effects:

[0186] (1) Improve the controllability and interpretability of the judgment generation process.

[0187] The above-described solution in this specification employs a chain-like thinking process, breaking down the judge's trial process into five distinct stages: "fact structuring, identification of points of contention, evidence analysis, application of law, and judgment generation." Each stage has clearly defined inputs and outputs, and the output of the previous stage can serve as the input for the next, thus achieving precise control and traceability of the reasoning process. Compared to existing methods that directly generate large models, this approach can precisely guide the model's attention allocation, avoiding the randomness and unpredictability of the reasoning path.

[0188] (2) Reduce the illusions and factual errors in the model.

[0189] The above-mentioned scheme in this specification adopts a phased verification and structured output mechanism. In the fact structuring stage, structured facts are extracted through entity recognition and relation extraction technology. In the evidence analysis stage, contradictory evidence is identified through a multimodal consistency verification algorithm. In the judgment generation stage, logical verification is used to ensure the accuracy of the judgment, thereby effectively controlling the occurrence of illusions and factual errors.

[0190] (3) Ensuring the rigor of legal logic

[0191] The above-mentioned scheme in this manual employs an automated and structured logical binding of "legal provisions (major premise)," "focus determination conclusion (minor premise)," and "judgment conclusion" to ensure that each judgment's main text has clear legal basis and factual support. Simultaneously, through responsive reasoning regarding the points of contention, the generated "Court's Opinion" section is logically rigorous and well-reasoned, achieving the level of argumentation found in professional judicial documents.

[0192] (4) Significantly enhanced adaptability to complex and novel cases.

[0193] The above-described solution in this manual is based on the general reasoning ability of a large model, combined with trial thinking mode modeling technology. It can handle various complex, non-standardized, and novel cases. It does not rely on a rule base preset for specific causes of action. It can flexibly handle various non-standardized, complex, and even novel legal relationships cases through semantic understanding. It breaks through the limitation of traditional technology that can only handle simple and repetitive cases, and greatly expands the application boundaries of intelligent trial.

[0194] (5) Breakthrough in multimodal evidence processing capabilities

[0195] The above-described solution in this specification encodes original evidence from different modalities, such as text, images, audio, and video, into the same vector space through a multimodal model. This enables the construction of evidence chains and consistency verification of cross-modal evidence. Compared with existing technologies that mainly process text evidence, this solution represents a breakthrough in multimodal evidence processing capabilities.

[0196] (6) Improved maintainability and scalability of the system

[0197] The above-described solution in this manual adopts a layered architecture and standardized interface design, with loose coupling between functions. Compared with the traditional model that requires the separate development of templates (element-based) or the construction of graphs (knowledge graphs) for each case, the maintenance cost is greatly reduced and the scalability is greatly enhanced.

[0198] (7) Processing efficiency has been greatly improved

[0199] The above-described scheme in this manual adopts an automated trial process, which can significantly improve processing efficiency compared to traditional manual trial methods.

[0200] Figure 6 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 6 At the hardware level, the device includes a processor 602, an internal bus 604, a network interface 606, memory 608, and non-volatile memory 610, and may also include other hardware required for its functions. One or more embodiments of this specification can be implemented in software, such as the processor 602 reading the corresponding computer program from the non-volatile memory 610 into memory 608 and then running it. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0201] Please refer to Figure 7 This specification also provides a judgment generation system 700 to implement the technical solution of this specification. The judgment generation system 700 may include:

[0202] The fact generation module 701 is used to input the complaint, answer, and court transcript of a legal case into the fact generation model, and use the fact generation model to perform structured processing on the facts of the legal case through entity recognition and relation extraction, generate the disputed facts and undisputed facts involved in the legal case, and determine the focus of the dispute.

[0203] The evidence identification module 702 is used to use the evidence identification model to identify the authenticity, legality, and relevance of the structured evidence in the legal case, obtain the three-fold identification results, and identify contradictory evidence and determine the acceptance result.

[0204] The fact-finding module 703 is used to use the fact-finding model to reason and ascertain the points of contention based on the structured evidence, the results of the three-fold determination, and the results of acceptance, so as to generate a conclusion on the point of contention.

[0205] The legal provision screening module 704 is used to screen legal provisions based on the disputed issues and the conclusions of the issues identification using a legal provision screening model, so as to obtain the applicable legal provisions for the legal case.

[0206] The content generation module 705 is used to generate the trial findings content using a content generation model based on the undisputed facts, the points of contention, the conclusion on the determination of the points of contention, the results of the determination of the three elements (relevance, authenticity, and legality), and the results of acceptance; generate the court's opinion content using the content generation model based on the trial findings content and the applicable legal provisions; generate the judgment item content using the content generation model based on the trial findings content, the court's opinion content, the applicable legal provisions, and the litigation claims; and generate the judgment of the legal case based on the trial findings content, the court's opinion content, and the judgment item content.

[0207] Optionally, the process by which the fact generation model determines the focus of the dispute includes:

[0208] Calculate the semantic similarity between different disputed facts;

[0209] The disputed facts are clustered according to a preset semantic similarity threshold and the semantic similarity, so as to merge semantically related disputed facts into the same disputed cluster, resulting in one or more disputed clusters;

[0210] The focus of the dispute is determined based on one or more of the dispute clusters.

[0211] Optionally, the method further includes:

[0212] Based on the cause of action of the aforementioned legal cases, the legal provisions in the legal provisions database are filtered to obtain first-level candidate legal provisions;

[0213] The legal provision screening model is used to screen legal provisions based on the points of contention and the conclusions reached regarding those points of contention, thereby obtaining the applicable legal provisions for the legal case, including:

[0214] The primary candidate legal provisions, the points of contention, and the conclusion of the point of contention are input into the legal provision screening model. The legal provision screening model determines the legal relationship of the legal case based on the conclusion of the factual determination. It then selects secondary candidate legal provisions from the primary candidate legal provisions that match the legal relationship and determines the secondary candidate legal provisions required to adjudicate the points of contention as the applicable legal provisions.

[0215] Optionally, after generating the judgment content, the content generation model performs legal consistency verification, factual closure verification, and procedural compliance verification on the judgment content; and outputs the judgment content after the judgment content passes the legal consistency verification, the factual closure verification, and the procedural compliance verification.

[0216] Optionally, the original evidence in the legal case is multimodal evidence; the method further includes:

[0217] The original multimodal evidence is input into the multimodal model, which performs semantic extraction on the multimodal evidence to obtain the semantic information corresponding to each multimodal evidence. The semantic information of each multimodal evidence is then semantically aligned based on a preset format to output the structured evidence.

[0218] Optionally, when the evidence identification model performs the three-fold identification of authenticity, legality, and relevance of the structured evidence in the legal case, it also performs cross-modal consistency verification on the structured evidence to identify temporal, logical, and / or monetary contradictions between different modal evidence, obtain consistency verification results, and combine the consistency verification results to perform the three-fold identification.

[0219] Optionally, the fact generation model is also used to generate a corresponding focus tree for each point of contention, wherein the root node of the focus tree is the point of contention, and the child nodes are the disputed facts corresponding to the point of contention;

[0220] The process by which the fact-finding model generates the focus determination conclusion of the disputed focus includes: after reasoning and identifying the disputed focus, generating the focus determination conclusion based on the focus tree.

[0221] Optionally, the method further includes:

[0222] The results of the acceptance, the conclusions on the determination of the focus, and one or more of the applicable legal provisions are displayed through a visual interface.

[0223] If the presented acceptance results are updated, the fact-finding model is reused to infer and ascertain the points of contention based on the structured evidence, the three-fold assessment results, and the updated acceptance results, so as to generate a new conclusion on the points of contention.

[0224] If the focus determination conclusion is updated, the legal provision screening model is reused to determine the new applicable legal provision for the legal case based on the updated focus determination conclusion.

[0225] If any input to the content generation model is updated, the updated data is re-inputted into the content generation model to regenerate the judgment.

[0226] Optionally, the process by which the fact-finding model infers and ascertains the points of contention based on the structured evidence, the results of the three-fold verification, and the acceptance results, to generate a conclusion on the points of contention, includes:

[0227] Based on the structured evidence, the results of the three-fold assessment, and the acceptance results, determine whether the conclusion on the focus of the assessment can be inferred;

[0228] In the absence of a conclusion on the identified focus, multiple investigation plans should be developed for the disputed focus.

[0229] Based on the structured evidence, the results of the three-dimensional identification, and the results of acceptance, each investigation plan is inferred to obtain the corresponding investigation plan results;

[0230] The focus identification conclusion is determined based on the results of the investigations corresponding to the multiple investigation plans.

[0231] Optionally, the method further includes:

[0232] The investigation plan, the investigation results, and the focus identification conclusion are displayed in the form of a mind map through a visual interface. The focus of the dispute is the main node of the mind map, the investigation plan is a child node of the main node, the investigation results are child nodes of the corresponding investigation plan, and the focus identification conclusion is a summary node of the multiple investigation results.

[0233] Optionally, the fact generation model is further used to determine the dependency relationship between the first and second points of contention when there are multiple points of contention; the fact determination model is further used to, under the premise that the first point of contention is the second point of contention, first perform reasoning and investigation on the first point of contention to obtain the first point of contention determination conclusion, and determine the second point of contention determination conclusion based on the first point of contention determination conclusion.

[0234] The aforementioned fact generation module 701, evidence identification module 702, fact identification module 703, legal provision screening module 704, and content generation module 705 can be applied to the aforementioned Figure 6This specification does not impose any special limitations on the devices shown.

[0235] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0236] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0237] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

Claims

1. A method for generating a judgment, characterized in that, The method includes: The complaint, answer, and court transcript of a legal case are input into a fact generation model. The fact generation model is then used to perform structured processing of the facts of the legal case through entity recognition and relation extraction, generating disputed and undisputed facts related to the legal case and identifying the focus of the dispute. The evidence identification model is used to conduct a three-fold assessment of the authenticity, legality, and relevance of structured evidence in the legal case, and to obtain the results of the three-fold assessment, as well as to identify contradictory evidence and determine the acceptance of the evidence. The fact-finding model is used to infer and ascertain the points of contention based on the structured evidence, the results of the three-fold verification, and the results of acceptance, so as to generate a conclusion on the points of contention. The applicable legal provisions for the legal case are obtained by using a legal provision screening model based on the points of contention and the conclusions reached regarding those points of contention. The content ascertained during the trial is generated using a content generation model based on the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three aspects (relevance, authenticity, and legality), and the results of the acceptance of the findings. The content generation model is used to generate the court's opinion based on the content ascertained during the trial and the applicable legal provisions; The content generation model is used to generate judgment items based on the content ascertained during the trial, the content held by the court, the applicable legal provisions, and the litigation claims. The judgment for the legal case is generated based on the findings of the trial, the opinion of this court, and the judgment items.

2. The method according to claim 1, characterized in that, The process by which the fact generation model determines the focus of the dispute includes: Calculate the semantic similarity between different disputed facts; The disputed facts are clustered according to a preset semantic similarity threshold and the semantic similarity, so as to merge semantically related disputed facts into the same disputed cluster, resulting in one or more disputed clusters; The focus of the dispute is determined based on one or more of the dispute clusters.

3. The method according to claim 1, characterized in that, The method further includes: Based on the cause of action of the aforementioned legal cases, the legal provisions in the legal provisions database are filtered to obtain first-level candidate legal provisions; The legal provision screening model is used to screen legal provisions based on the points of contention and the conclusions reached regarding those points of contention, thereby obtaining the applicable legal provisions for the legal case, including: The primary candidate legal provisions, the points of contention, and the conclusion of the point of contention are input into the legal provision screening model. The legal provision screening model determines the legal relationship of the legal case based on the conclusion of the factual findings. It then selects secondary candidate legal provisions from the primary candidate legal provisions that match the legal relationship and determines the secondary candidate legal provisions required to adjudicate the points of contention as the applicable legal provisions.

4. The method according to claim 1, characterized in that, After generating the judgment content, the content generation model performs legal consistency verification, factual closure verification, and procedural compliance verification on the judgment content; and outputs the judgment content after the judgment content passes the legal consistency verification, factual closure verification, and procedural compliance verification.

5. The method according to claim 1, characterized in that, The original evidence in the legal case is multimodal evidence; the method also includes: The original multimodal evidence is input into the multimodal model, which performs semantic extraction on the multimodal evidence to obtain the semantic information corresponding to each multimodal evidence. The semantic information of each multimodal evidence is then semantically aligned based on a preset format to output the structured evidence.

6. The method according to claim 1, characterized in that, The method further includes: The results of the acceptance, the conclusions on the determination of the focus, and one or more of the applicable legal provisions are displayed through a visual interface. If the presented acceptance results are updated, the fact-finding model is reused to infer and ascertain the points of contention based on the structured evidence, the three-fold assessment results, and the updated acceptance results, so as to generate a new conclusion on the points of contention. If the focus determination conclusion is updated, the legal provision screening model is reused to determine the new applicable legal provision for the legal case based on the updated focus determination conclusion. If any input to the content generation model is updated, the updated data is re-inputted into the content generation model to regenerate the judgment.

7. The method according to claim 1, characterized in that, The fact-finding model, based on the structured evidence, the results of the three-fold verification, and the acceptance results, deduces and ascertains the points of contention to generate a conclusion on the points of contention, comprising: Based on the structured evidence, the results of the three-fold assessment, and the acceptance results, determine whether the conclusion on the focus of the assessment can be inferred; In the absence of a conclusion on the identified focus, multiple investigation plans should be developed for the disputed focus. Based on the structured evidence, the results of the three-dimensional identification, and the results of acceptance, each investigation plan is inferred to obtain the corresponding investigation plan results; The focus identification conclusion is determined based on the results of the investigations corresponding to the multiple investigation plans.

8. The method according to claim 7, characterized in that, The method further includes: The investigation plan, the investigation results, and the focus identification conclusion are displayed in the form of a mind map through a visual interface. The focus of the dispute is the main node of the mind map, the investigation plan is a child node of the main node, the investigation results are child nodes of the corresponding investigation plan, and the focus identification conclusion is a summary node of the investigation results of multiple plans.

9. The method according to claim 1, characterized in that, The fact generation model is also used to determine the dependency relationship between the first and second points of contention when there are multiple points of contention; the fact determination model is also used to, under the premise that the first point of contention is the second point of contention, first perform reasoning and investigation on the first point of contention to obtain the first point of contention determination conclusion, and determine the second point of contention determination conclusion based on the first point of contention determination conclusion.

10. A judgment generation system, characterized in that, The system includes: The fact generation module is used to input the complaint, answer, and court transcript of a legal case into the fact generation model. The fact generation model uses entity recognition and relation extraction to perform structured processing on the facts of the legal case, generate the disputed and undisputed facts involved in the legal case, and determine the focus of the dispute. The evidence identification module is used to use the evidence identification model to identify the authenticity, legality, and relevance of the structured evidence in the legal case, obtain the results of the three-fold identification, and identify contradictory evidence and determine the acceptance result. The fact-finding module is used to use a fact-finding model to reason and ascertain the points of contention based on the structured evidence, the results of the three-fold verification, and the results of acceptance, so as to generate a conclusion on the point of contention. The legal provision screening module is used to screen legal provisions based on the points of contention and the conclusions of the point of contention using a legal provision screening model, so as to obtain the applicable legal provisions for the legal case. The content generation module is used to generate trial findings content based on the undisputed facts, the points of contention, the conclusions on the determination of the points of contention, the results of the determination of the three elements (relevance, authenticity, and legality), and the results of acceptance of the findings; to generate the court's opinion content based on the trial findings content and the applicable legal provisions; to generate judgment items based on the trial findings content, the court's opinion content, the applicable legal provisions, and the litigation claims; and to generate the judgment of the legal case based on the trial findings content, the court's opinion content, and the judgment items.

Images