A method and related device for generating a review discovery text

Abstract

The application discloses a trial finding text generation method and related device, and relates to the technical field of artificial intelligence, which comprises the following steps: normalizing a plurality of evidence data of a target case into a structured evidence list, taking an evidence item in the evidence list as an evidence node, taking a fact element extracted from the evidence list as a fact node, constructing an evidence fact correlation graph, the edge relationship in the evidence fact correlation graph comprises a contradiction relationship, inputting the evidence fact correlation graph into a graph neural network model for feature transmission and aggregation between nodes, obtaining updated features of each fact node, and determining a dispute focus node and a non-dispute node from each fact node according to the fact credibility in the updated features and the contradiction relationship in the evidence fact correlation graph, and generating a trial finding text in a judgment according to the dispute focus node and the non-dispute node. The application simulates the reasoning thinking process of judges in determining facts by the graph node feature transmission process, and improves the accuracy and efficiency of fact determination.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for generating adjudication and investigation text. Background Technology

[0002] In judicial practice, drafting judgments is a core intellectual task for judges. The "Ascertainment of Facts" section serves as a crucial vehicle for determining the facts of a case. The accuracy, completeness, and organization of its content directly affect the solidity of the factual basis and profoundly influence the correctness of subsequent legal application and the social acceptance of the judgment.

[0003] Currently, the drafting of the "Ascertainment of Facts" section still relies heavily on the individual labor of judges. In the contemporary context of emphasizing both judicial efficiency and quality, several practical difficulties are gradually emerging: First, with the continuous increase in the number of cases, analyzing and drafting this section of facts requires a significant investment of time and manpower, resulting in low efficiency and high labor costs; Second, the evidence, the facts to be proven, and the parties' claims form a complex logical network, and manually sorting out the chain of evidence, identifying contradictions in evidence, and discovering points of contention is prone to oversights, leading to deviations in the determination of facts. Summary of the Invention

[0004] In view of the above problems, this application provides a method and related apparatus for generating trial ascertainment texts to solve the problems of high labor costs, low efficiency, and factual deviations due to omission of disputed points that exist in the current method of manually drafting trial ascertainment texts. The specific solution is as follows:

[0005] The first aspect of this application provides a method for generating examination and ascertainment text, including:

[0006] Acquire multiple pieces of evidence data in the target case and organize them into a structured evidence list. The evidence list includes multiple evidence items, with different evidence items used to describe different pieces of evidence data.

[0007] Using the evidence items as evidence nodes and the factual elements extracted from the evidence list as fact nodes, an evidence fact association graph is constructed, and the edge relationships in the evidence fact association graph include contradictory relationships.

[0008] The evidence fact association graph is input into a graph neural network model for feature transfer and aggregation between nodes to obtain updated features of each fact node. Based on the fact credibility in the updated features and the contradictory relationships in the evidence fact association graph, disputed focus nodes and undisputed nodes are determined from each fact node. The fact credibility is the value after feature aggregation of the associated evidence nodes and the fact nodes.

[0009] Based on the disputed key nodes and the undisputed nodes, the trial findings text in the judgment is generated.

[0010] In one possible implementation, constructing an evidence fact association graph using the evidence items as evidence nodes and the fact elements extracted from the evidence list as fact nodes includes:

[0011] The evidence items are used as evidence nodes, and the factual elements extracted from the evidence list are used as fact nodes to obtain a node set;

[0012] Perform relation parsing on the nodes in the node set to determine the nodes with edge relationships and the edge relationships;

[0013] For every two nodes in the node set that have the edge relationship: if they are two evidence nodes, the degree of support each of the two evidence nodes provides to the other is determined as the edge weight between the two evidence nodes; if they are two fact nodes, the correlation strength between the two fact nodes is determined as the edge weight between the two fact nodes; if they are an evidence node and a fact node, the semantic relevance between the evidence node and the fact node and the reliability of the evidence node are determined, and the edge weight between the evidence node and the fact node is determined based on the semantic relevance and the reliability.

[0014] Based on the set of nodes, the edge relationships, and the edge weights, the evidence fact association graph is constructed.

[0015] In one possible implementation, after the fact elements extracted from the evidence list are used as fact nodes, the following is also included:

[0016] The time elements in the fact elements corresponding to each fact node are processed into time information under a unified time standard, so as to determine the temporal relationship between each fact node based on the time information.

[0017] Each fact node is matched with at least one preset logical rule, and the logical relationship between each fact node is determined based on the matching result. The at least one logical rule includes causal rules and process rules. The fact elements also include the subject representing the initiator of the fact action and the object representing the recipient or related object of the fact action. In the first fact node and the second fact node that satisfy the causal rule, the object of the first fact node and the subject of the second fact node have a semantic relationship. The time information of the first fact node precedes the time information of the second fact node. The subjects or objects of multiple fact nodes that satisfy the process rule are consistent, the time sequence is continuous, and the fact actions have a logically progressive relationship.

[0018] Determining the association strength between two fact nodes includes:

[0019] The association strength between the two fact nodes is determined based on the temporal and / or logical relationships between them.

[0020] In one possible implementation, the step of inputting the evidence fact association graph into a graph neural network model for feature transfer and aggregation between nodes to obtain updated features for each fact node, and determining the disputed focus node and undisputed node from each fact node based on the fact credibility in the updated features and the contradictory relationships in the evidence fact association graph, includes:

[0021] The evidence evaluation module of the graph neural network model determines the evidentiary value of each evidence node based on the edge weights between the evidence node and the associated nodes, and determines the initial features of the evidence node based on the evidentiary value. The associated nodes include evidence nodes and fact nodes connected to the evidence node.

[0022] Through the first convolutional module of the graph neural network model, for each fact node, the initial features of the fact node are determined based on the edge weights between the fact node and the associated evidence node and the features of the associated evidence node, wherein the associated evidence node is the evidence node connected to the fact node.

[0023] Through the second convolutional module of the graph neural network model, for each fact node, the updated features of the fact node are determined based on the edge weights between the fact node and the associated fact nodes and the initial features of the fact node. The associated fact nodes are the fact nodes connected to the fact node.

[0024] The node classification module of the graph neural network model determines the disputed focus nodes and undisputed nodes from the fact nodes based on the fact credibility in the updated features of each fact node and the contradictory relationships in the evidence fact association graph.

[0025] In one possible implementation, determining the disputed focus node and undisputed node from each of the fact nodes based on the fact credibility in the updated features of each fact node and the contradictory relationships in the evidence fact association graph includes:

[0026] The final credibility of the fact node is determined based on the fact credibility in the updated features of the fact node and the initial confidence in the fact elements corresponding to the fact node.

[0027] Based on the final credibility of each fact node and the contradictory relationships in the evidence fact association graph, disputed focus nodes and undisputed nodes are determined from each fact node.

[0028] In one possible implementation, determining the disputed focus node and undisputed node from the fact nodes based on the final credibility of each fact node and the contradictory relationships in the evidence fact association graph includes:

[0029] Identify whether there are multiple fact nodes among the fact nodes that contradict each other and whose final credibility is greater than a preset credibility threshold. If so, then identify the multiple fact nodes as the focus of the dispute.

[0030] If there is a fact node among the fact nodes whose final credibility is greater than the credibility threshold and whose connected evidence nodes are not contradictory, then the fact node is determined as an undisputed node.

[0031] In one possible implementation, generating the trial findings text in the judgment based on the disputed focal point node and the undisputed node includes:

[0032] For each of the disputed focal nodes and the undisputed nodes, evidence nodes with a correlation greater than a preset correlation threshold are determined from the set of evidence nodes connected to that node. The evidence chain composed of the determined evidence nodes is used as the key evidence chain corresponding to that node, so as to obtain the key evidence chains corresponding to the disputed focal nodes and the undisputed nodes respectively. The correlation is obtained from the updated features of the node.

[0033] From the disputed focus node, the undisputed node, and the key evidence chains corresponding to the disputed focus node and the undisputed node, extract the slot data instances corresponding to each chapter-level template of the trial and investigation template, and fill the corresponding slot data instances into the corresponding slots of the corresponding chapter-level template to obtain at least one chapter-level generation blueprint.

[0034] Each chapter-level generated blueprint and pre-built prompt words are concatenated into a prompt instruction, which is then input into a large language model to obtain the chapter-level text corresponding to each chapter-level generated blueprint.

[0035] The chapter-level text corresponding to the at least one chapter-level generated blueprint is concatenated into a coherent text, which serves as the text for investigation and verification.

[0036] One possible implementation also includes:

[0037] Identify factual statements in the text ascertained during the trial, calculate vector similarity between the factual statements and the disputed key nodes and the undisputed nodes respectively, and determine whether the factual statements contain factual fabrications based on the vector similarity.

[0038] And / or, identify colloquial expressions in the trial investigation text and replace them with professional legal language.

[0039] One possible implementation also includes:

[0040] An evidence number is appended to the factual statement, which is used to link to the evidence fragment on which the factual statement depends and to highlight the evidence fragment in the corresponding evidence data;

[0041] And / or, highlight the segments in the trial investigation text that are related to the disputed key points.

[0042] In one possible implementation, the evidence entry includes one or more of the following evidentiary elements: the witness, the name of the evidence, the substantive content of the evidence, the purpose of the proof, and the location information of the substantive content of the evidence in the corresponding evidence data.

[0043] A second aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the examination and ascertainment text generation method described in the first aspect or any implementation thereof.

[0044] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0045] The memory is used to store computer programs;

[0046] The processor is used to execute the computer program so that the electronic device can implement the trial and investigation text generation method of the first aspect or any implementation thereof.

[0047] The fourth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the examination and investigation text generation method described in the first aspect or any implementation thereof.

[0048] By employing the aforementioned technical solution, the trial investigation text generation method provided in this application acquires multiple pieces of evidence data from a target case and organizes them into a structured evidence list. This evidence list includes multiple evidence items, each describing different pieces of evidence data. To more accurately and comprehensively organize the evidence chain and the facts to be proven, factual elements used to construct the facts are extracted from the evidence list. Then, using evidence items as evidence nodes and extracted factual elements as fact nodes, an evidence-fact relationship graph is constructed, where the edge relationships of the graph include contradictory relationships. By constructing an evidence-fact relationship graph, this application achieves global visualization and explicit identification of contradictions in evidence, facts, and their relationships in complex cases, providing a precise and complete structured data foundation for fact-finding.

[0049] Considering that a judge's thought process in determining facts involves initially collecting relevant evidence and facts, confirming the legality and authenticity of the evidence, examining whether each fact is supported by legal and authentic evidence, and then examining whether there is mutual support between the facts to form a logically consistent and complete narrative loop, thereby determining undisputed facts and points of contention. To simulate this thought process of mutual verification between evidence, between evidence and facts, and between facts themselves, and to overcome the potential for overlooking points of contention in manual analysis, this application can input the evidence-fact association graph into a graph neural network model for feature transfer and aggregation between nodes, obtaining updated features for each fact node. Then, based on the factual credibility in the updated features and the contradictory relationships in the evidence-fact association graph, points of contention and undisputed nodes are determined from each fact node. Since the factual credibility of a fact node is a credibility result obtained after aggregating the features of related evidence nodes and fact nodes, consistent with human thought processes, the accuracy of this factual credibility is higher. Therefore, based on this factual credibility and combined with the contradictory relationships in the graph, the accuracy of determining the two types of nodes is higher. At the same time, the process of determining the two types of nodes does not require human intervention, avoiding the omission of points of contention. Based on this, the trial findings text of the judgment is generated based on the disputed key points and undisputed points, which improves the accuracy and objectivity of the fact-finding content in the trial findings.

[0050] Meanwhile, the entire process of sorting out and analyzing evidence and facts, and generating the trial investigation text, does not require human intervention, which reduces labor costs and improves efficiency. Attached Figure Description

[0051] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0052] Figure 1 A schematic diagram of a system architecture provided for this application;

[0053] Figure 2 A flowchart illustrating a method for generating examination and ascertainment text provided in this application;

[0054] Figure 3 A schematic diagram of a text generation device for adjudication and investigation provided in this application;

[0055] Figure 4 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0056] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0057] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0058] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0059] As described in the background section, manually drafting the trial investigation text is costly and inefficient. Furthermore, manually sorting out evidence and facts can easily lead to the omission of contradictory evidence or key points of contention, resulting in deviations in the determination of facts.

[0060] To achieve automatic generation of judgment texts, one initial approach is to extract content from a structured case information table using preset keyword matching rules and fill it into the corresponding positions in the judgment template. Another approach is to use natural language processing technologies (such as named entity recognition and keyword extraction) to extract surface information from electronic case files, generate an evidence list or information summary, and then use the evidence list or information summary to generate the judgment text.

[0061] However, both of the above approaches are essentially "fill-in-the-blank" automated text generation, which relies entirely on highly structured input information and cannot process unstructured, massive amounts of original evidence. Furthermore, the trial findings text generated in this way is merely a procedural listing of facts or a simple restatement of the litigation claims. It lacks the ability to independently summarize, analyze, and determine the facts of the case from the evidence materials, which may lead to factual deviations in the generated judgment text.

[0062] In addition, Large Language Models (LLMs) can be used to generate trial findings texts using prompt words. However, LLM-generated texts tend to deviate from the strict logical structure of judicial documents (procedural first, then substantive; undisputed first, then disputed), and lack professionalism. In pursuit of fluency, facts or details that do not exist in the evidence may be fabricated, violating the principles of judicial rigor and objectivity.

[0063] To overcome the aforementioned problems, this application provides a method and related apparatus for generating trial findings text, which can be applied to scenarios where judges assist in generating trial findings text when drafting judgments. It should be noted that the above scenarios are merely examples and are not intended to limit this application.

[0064] Optionally, the method for generating trial findings text provided in this application can be applied to, for example... Figure 1 The system architecture shown includes a terminal 100 and a server 200. The server 200 may include one or more servers (…). Figure 1 (This example uses a server as an illustration).

[0065] Either terminal 100 or server 200 can be used independently to execute the trial ascertainment text generation method provided in the embodiments of this application. Alternatively, terminal 100 and server 200 can also be used collaboratively to execute the trial ascertainment text generation method provided in the embodiments of this application.

[0066] The following description Figure 1 The product form of the mid-terminal 100;

[0067] The terminal 100 in this application embodiment can be a mobile phone, tablet computer, wearable device, vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any restrictions on it.

[0068] To enable those skilled in the art to better understand this application, the method for generating examination and ascertainment text according to embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0069] Reference Figure 2 , Figure 2 This application provides a flowchart illustrating a method for generating examination and verification text, as shown in the embodiments below. Figure 2 As shown, the method for generating the trial investigation text may include:

[0070] Step S201: Obtain multiple pieces of evidence data for the target case and organize them into a structured evidence list. The evidence list includes multiple evidence items, and different evidence items are used to describe different pieces of evidence data.

[0071] Here, the original evidence data is unstructured data, and multiple pieces of evidence data may be data from different sources and in different formats. For example, multiple pieces of evidence data may include unstructured document data such as image data and scanned copies.

[0072] To facilitate subsequent processing, multiple pieces of evidence data can be organized into a structured evidence list.

[0073] In one possible implementation, firstly, optical character recognition technology can be used to extract text from multiple unstructured pieces of evidence data. Then, a document layout analysis and cataloging engine can be used to automatically identify index data such as type, title, and page number in the evidence data that can be used to locate text positions, and establish the relationship between the index data and the extracted text to form a structured and searchable electronic evidence library.

[0074] Considering that the evidence may contain noisy data that is irrelevant or of low relevance to the case proceedings, in order to remove the noisy data and retain only the evidentiary elements that are highly relevant to the case proceedings and can construct and prove legal facts, this application can further refine the analysis and processing of the text in the aforementioned electronic evidence database to generate an evidence list.

[0075] Specifically, a deep learning model, pre-trained on massive amounts of legal texts and fine-tuned with data from the judicial field, can be used to automatically identify and extract key entities in the text corresponding to each piece of evidence that are highly relevant to the case trial process and can construct and prove legal facts. These entities include, but are not limited to: the witness (such as the plaintiff, defendant, third party, etc.), the evidence item (i.e. the specific name of the evidence), and the substantive content of the evidence (such as time, amount, clause content, etc.).

[0076] In addition, when extracting the substantive content of evidence, in order to clarify the location of the substantive content in the evidence data and facilitate subsequent review, the location information of the substantive content of the evidence in the corresponding evidence data can be automatically recorded, such as page number, paragraph, line number, and text number.

[0077] In addition, by combining the extracted key entities, we can analyze and infer the purpose of the evidence submitted by the party presenting the evidence, such as proving the establishment of a contract, proving a breach of contract, or proving the amount of loss.

[0078] Therefore, this embodiment can generate an evidence entry for each piece of evidence data, such as the witness, evidence name, substantive content of the evidence, purpose of proof, and location information of the substantive content of the evidence in the corresponding evidence data. Through the various evidence elements contained in the evidence entry, the different perspectives of the evidence data can be described more completely and accurately, realizing the preliminary semantic understanding and digital modeling of the evidence data.

[0079] Finally, all the evidence items are combined to form an evidence list.

[0080] It should be noted that the above process of organizing evidence into a list is merely an example and is not intended to limit this application. In addition, the five dimensions of evidence items, including the witness, the name of the evidence, the substantive content of the evidence, the purpose of the proof, and the location information of the substantive content of the evidence in the corresponding evidence data, are merely examples. In addition, evidence items may include more or less than five dimensions of evidence elements, depending on the application scenario, and this application does not make specific limitations.

[0081] Step S202: Using evidence items as evidence nodes and fact elements extracted from the evidence list as fact nodes, construct an evidence fact association graph. The edge relationships in the evidence fact association graph include contradictory relationships.

[0082] Understandably, the core content of the trial investigation is to ascertain the facts. Therefore, it is necessary to first determine the various facts involved in the target case. In order to analyze the relationship between facts and evidence more objectively, and then deduce the points of contention and undisputed facts, this embodiment can combine the court transcript of the target case to pre-organize multiple atomic facts at the finest granularity, such as Party A sending a "Demand Letter" to Party B at time T1.

[0083] It is worth noting that the atomic facts compiled from the court transcripts are facts to be proven, and there may be facts claimed by the plaintiff, defendant or third party (which may be false facts without evidence). In order to determine whether each atomic fact is a legal fact supported by evidence, this embodiment can extract the factual elements used to construct each atomic fact from the evidence list.

[0084] Optionally, the deep learning model, which was pre-trained on massive amounts of legal texts and fine-tuned with data from the judicial field, can be used to extract factual elements from the evidence list to construct the individual atomic facts. Of course, other deep learning models can also be used, and no specific limitations are made here.

[0085] Optionally, the factual elements used to construct an atomic fact include: fact ID (Identity), factual content, subject, object, time element, location element, factual source, and initial confidence level. Here, factual content refers to the core predicate phrase in the atomic fact, i.e., the factual action, such as signing a contract, making payment, or delivering goods; subject refers to the initiator of the factual action, such as plaintiff Company A or defendant Mr. Zhang; object refers to the recipient or related object of the factual action, such as contract number xxx, a loan of 500,000 yuan, or a batch of building materials; time element refers to the time when the event occurred, which can be an absolute time, such as January 1, 2026, or a relative time, such as within three days after the contract was signed; location element refers to the location where the event occurred, such as through bank transfer or at Party A's business location; source refers to the location from which the relevant evidence of the atomic event originates from the original evidence data; and the initial confidence level can be predicted by the model.

[0086] In order to transform fragmented evidence items and atomic facts into interconnected and semantically structured knowledge, and to facilitate the subsequent analysis of the mutual support between evidence, facts, and evidence, this application embodiment can construct an evidence-fact relationship graph using evidence items as evidence nodes and extracted fact elements as fact nodes.

[0087] Among them, the edge relationships in the evidence fact correlation graph include contradictory relationships. Optionally, the edge relationships may also include direct supporting relationships, indirect corroborating relationships, etc., which are not specifically limited in this application.

[0088] Step S203: Input the evidence fact association graph into the graph neural network model to perform feature transfer and aggregation between nodes, obtain the updated features of each fact node, and determine the dispute focus node and undisputed node from each fact node based on the fact credibility in the updated features and the contradictory relationship in the evidence fact association graph.

[0089] The credibility of a fact is the value obtained by aggregating the associated evidence nodes and features of the fact nodes.

[0090] In a judge's process of determining facts, after examining the legality and authenticity of evidence, each fact is examined to see if it is supported by legally valid and authentic evidence. Then, the relationships between these facts are examined to form a logically consistent and complete narrative loop. Through this review process, it is possible to more accurately determine whether the multiple atomic facts obtained earlier can constitute credible legal facts. By combining these credible legal facts with the contradictory relationships between credible evidence, the undisputed facts and the points of contention can be identified.

[0091] To simulate the thought process of the above analysis and reasoning, this embodiment can input the evidence fact association graph into the graph neural network model, so that feature transfer and aggregation between nodes can be realized through multi-layer graph convolution operations in the model, and then the credibility of each atomic event (i.e. fact node) can be determined based on the aggregation result.

[0092] In this embodiment, the updated features of each fact node have undergone feature aggregation of associated evidence nodes and associated fact nodes. Therefore, the fact credibility contained in the updated features is the credibility result after the associated evidence and fact verification. The more credible evidence and facts support the atomic facts corresponding to a fact node, the higher the fact credibility of that fact node.

[0093] Understandably, for an atomic fact to become a point of contention, two necessary conditions are required: high credibility and a contradictory relationship with other atomic facts. Similarly, for an atomic fact to become an undisputed point of contention, two necessary conditions are required: high credibility and consistent, uncontradictory related evidence. Therefore, this embodiment can determine the point of contention and undisputed nodes from each fact node based on the factual credibility of the fact nodes and the contradictory relationships in the evidence fact association graph. Here, the atomic facts corresponding to the point of contention nodes are the points of contention facts, and the atomic facts corresponding to the undisputed nodes are the undisputed facts.

[0094] Step S204: Based on the disputed key nodes and undisputed nodes, generate the trial findings text in the judgment.

[0095] Once the key points of contention and the undisputed points are accurately identified without omission, the trial findings text in the judgment can be accurately generated based on the key points of contention and the undisputed facts.

[0096] The method for generating trial findings text provided in this application acquires multiple pieces of evidence data from a target case and organizes them into a structured evidence list. This evidence list includes multiple evidence items, each describing different pieces of evidence data. To more accurately and comprehensively analyze the chain of evidence and the facts to be proven, factual elements used to construct the facts are extracted from the evidence list. Then, using evidence items as evidence nodes and extracted factual elements as fact nodes, an evidence-fact relationship graph is constructed, where the edge relationships of the graph include contradictory relationships. By constructing this evidence-fact relationship graph, this application achieves global visualization and explicit identification of contradictions in evidence, facts, and their relationships in complex cases, providing a precise and complete structured data foundation for fact-finding.

[0097] Considering that a judge's thought process in determining facts involves initially collecting relevant evidence and facts, confirming the legality and authenticity of the evidence, examining whether each fact is supported by legal and authentic evidence, and then examining whether there is mutual support between the facts to form a logically consistent and complete narrative loop, thereby determining undisputed facts and points of contention. To simulate this thought process of mutual verification between evidence, between evidence and facts, and between facts themselves, and to overcome the potential for overlooking points of contention in manual analysis, this application can input the evidence-fact association graph into a graph neural network model for feature transfer and aggregation between nodes, obtaining updated features for each fact node. Then, based on the factual credibility in the updated features and the contradictory relationships in the evidence-fact association graph, points of contention and undisputed nodes are determined from each fact node. Since the factual credibility of a fact node is a credibility result obtained after aggregating the features of related evidence nodes and fact nodes, consistent with human thought processes, the accuracy of this factual credibility is higher. Therefore, based on this factual credibility and combined with the contradictory relationships in the graph, the accuracy of determining the two types of nodes is higher. At the same time, the process of determining the two types of nodes does not require human intervention, avoiding the omission of points of contention. Based on this, the trial findings text of the judgment is generated based on the disputed key points and undisputed points, which improves the accuracy and objectivity of the fact-finding content in the trial findings.

[0098] Meanwhile, the entire process of sorting out and analyzing evidence and facts, and generating the trial investigation text, does not require human intervention, which reduces labor costs and improves efficiency.

[0099] In some embodiments of this application, the process of step S202, "constructing an evidence fact association graph using evidence items as evidence nodes and fact elements extracted from the evidence list as fact nodes", is described in detail.

[0100] In one possible implementation, this embodiment can use evidence items as evidence nodes and factual elements extracted from the evidence list as fact nodes to obtain a node set. ,in, , , This represents the m-th evidence node. This represents the nth fact node.

[0101] Next, relationship parsing can be performed on the nodes in the node set to determine the nodes with edge relationships and the edge relationships. For example, pairwise semantic parsing can be performed on each evidence node to determine the direct support, indirect corroboration, and contradiction equilateral relationships between evidence nodes; pairwise semantic parsing can be performed on evidence nodes and fact nodes to determine the direct support, indirect corroboration, and contradiction equilateral relationships between evidence nodes and fact nodes; logical parsing can be performed on fact nodes based on preset logical rules to determine fact nodes with logical relationships, and then the direct support, indirect corroboration, and contradiction equilateral relationships between fact nodes can be determined according to the specific logical relationships, and / or, the temporal relationship between fact nodes can be determined based on the time element in the fact elements, and then the direct support, indirect corroboration, and contradiction equilateral relationships between fact nodes can be determined based on the temporal relationship.

[0102] Of course, the above process of resolving relationships is merely an example and is not intended to limit this application.

[0103] In this embodiment, the edge weight can also be determined for every two nodes in the node set that have an edge relationship.

[0104] First, if two nodes with an edge relationship are two evidence nodes, then the degree of support each evidence node provides to the other is determined as the edge weight between the two evidence nodes. Taking evidence nodes a and b as an example, the edge weight includes: the degree of support that evidence node a provides to evidence node b, and / or, the degree of support that evidence node b provides to evidence node a.

[0105] Second, if two nodes with an edge relationship are two fact nodes, then the association strength between the two fact nodes is determined as the edge weight between the two fact nodes.

[0106] Optionally, in this embodiment, after the fact elements extracted from the evidence list are used as fact nodes, the fact nodes are sorted and grouped in chronological order, and then the correlation strength between two fact nodes is determined based on the chronological sorting results (such as a chronological dependency graph) and the grouping results.

[0107] In order to facilitate chronological sorting, this embodiment can process the time elements in the fact elements corresponding to each fact node into time information under a unified time standard, and then determine the chronological relationship between each fact node based on the time information.

[0108] For example, optionally, a joint model based on rules and deep learning can be used to identify time expressions in time elements, such as autumn of 2025, March of the following year, and within five days after receiving the notification. Then, a time inference engine can be used to normalize all relative and fuzzy times in the time elements into absolute times based on the baseline time point of the target case (such as the date of case filing), or normalize them to a comparable time interval to obtain the time information corresponding to each fact node. Based on the time information corresponding to each fact node, the temporal relationship between each fact node can be determined.

[0109] For example, for two fact nodes and Their normalized time information is as follows: and Then its temporal relationship It can be:

[0110] Formula (1).

[0111] Optionally, in order to record the temporal relationship more clearly, this embodiment can use each fact node as a node, the determined temporal relationship as a directed edge, and use a topological sorting strategy to globally sort each fact node to construct a temporal dependency graph.

[0112] In addition, this embodiment can preset at least one logical rule and match each fact node with the at least one logical rule, and determine the logical relationship between each fact node based on the matching result.

[0113] Optionally, the aforementioned "at least one logical rule" includes causal rules and procedural rules; correspondingly, logical relationships include causal relationships and procedural relationships within the same transaction (such as "negotiation," "drafting," "signing" -> "the process of both parties entering into a contract"). Specifically, for the first and second fact nodes that satisfy the causal rule, the object of the first fact node and the subject of the second fact node are semantically related, and the time information of the first fact node precedes the time information of the second fact node; for multiple fact nodes that satisfy the procedural rule, the subjects or objects of these multiple fact nodes are consistent, their temporal sequence is continuous, and the factual actions contained in their factual content have a logically progressive relationship.

[0114] For example, in order to determine the causal relationship between each fact node, this embodiment can group every two fact nodes together, take the node with earlier time information as the first fact node, and the other node as the second fact node, calculate the semantic relevance between the object of the first fact node and the subject of the second fact node, if the semantic relevance is greater than the preset semantic relevance threshold, then it is determined that there is a semantic association between the object of the first fact node and the subject of the second fact node, that is, there is a causal relationship between the first fact node and the second fact node.

[0115] For example, in order to determine the same transaction process relationship between each fact node, this embodiment can construct each fact node and the process rule "that is, multiple fact nodes that satisfy the process rule, which are consistent in subject or object, sequential in time and have logically progressive relationship in fact actions contained in fact content", as prompt instructions, and then input them into a deep learning model that has been pre-trained with massive legal texts and fine-tuned with judicial data to obtain at least one set of fact nodes output by the model, and the fact nodes in each set of fact nodes have the same transaction process relationship.

[0116] Based on this, the optional process of "determining the association strength between two fact nodes" may include: determining the association strength between two fact nodes based on the temporal and / or logical relationships between them.

[0117] For example, for two fact nodes and If a causal relationship or a relationship within the same transaction process exists, then the fact node is determined. and correlation strength If it is 0.9, Time information before ,but It is 0.8.

[0118] Of course, the above correlation strength is merely an example and is not intended to limit this application.

[0119] Third, if it is an evidence node and a fact node, then determine the semantic relevance between the evidence node and the fact node and the reliability of the evidence node. Based on the semantic relevance between the evidence node and the fact node and the reliability of the evidence node, determine the edge weight between the evidence node and the fact node.

[0120] Optionally, the semantic relevance between the evidence node and the fact node and the reliability of the evidence node can be weighted and summed to obtain the edge weight between the evidence node and the fact node, for example, the following formula (2).

[0121] Formula (2);

[0122] in, Indicates evidence node With fact nodes Edge weights between them Indicates evidence node With fact nodes Semantic relevance between them Indicates evidence node Reliability, Let two weight coefficients satisfy... .

[0123] Optional, evidence nodes With fact nodes The semantic relevance between evidence nodes can be determined by analyzing the evidence nodes. With fact nodes It is obtained by calculating cosine similarity, vector Euclidean distance, etc.

[0124] Optional, evidence nodes The reliability can be determined based on the type of evidence, such as documentary evidence > witness testimony, with a value range of [0,1].

[0125] Optional, It can be determined by experts in the judicial field.

[0126] Having obtained the node set, edge relationships, and edge weights in the previous section, we can construct an evidence fact association graph.

[0127] This embodiment constructs an evidence-fact relationship map, which can process discrete evidence and facts into structured information and more clearly record the relationship between evidence and facts, laying the foundation for subsequent analysis of the credibility of evidence and facts.

[0128] In some other embodiments of this application, the process of step S203 described above, which involves "inputting the evidence fact association graph into a graph neural network model to perform feature transfer and aggregation between nodes, obtaining updated features of each fact node, and determining the disputed focus node and undisputed node from each fact node based on the fact credibility in the updated features and the contradictory relationships in the evidence fact association graph", will be described.

[0129] In one possible implementation, the graph neural network model may include an evidence evaluation module, a first convolutional module, a second convolutional module, and a node classification module.

[0130] This embodiment can use an evidence evaluation module to determine the evidentiary validity of each evidence node based on the edge weights between the evidence node and associated nodes, and determine the initial characteristics of the evidence node based on the evidentiary validity. The associated nodes include evidence nodes and fact nodes connected to the evidence node in the evidence fact association graph, and the initial characteristics of the evidence node can characterize the credibility of the evidence node.

[0131] For example, for each piece of evidence, its evidentiary value can be assessed based on the degree of support provided by other evidence nodes connected to it, as well as the strength of its association with the factual nodes connected to it. In this embodiment, the evidentiary value can characterize the probative force and importance of the evidence node.

[0132] Optionally, the evidentiary value of the evidence node can be directly used as the initial feature of the evidence node.

[0133] Optionally, considering that each evidence node has its own unique characteristics in addition to its relationship with other nodes, such as the reliability of the evidence itself, the type of evidence, and the evidence item itself, this embodiment can also determine the initial characteristics of the evidence node based on its evidentiary validity, pre-marked self-attributes (such as reliability and type of evidence), and evidence item.

[0134] Of course, there may be other ways to determine the initial features, which will not be described in this application.

[0135] It is understandable that the higher the evidentiary validity and reliability of an evidence node, the more credible the fact node supported by that evidence node is. Therefore, in order to determine the credibility of each fact node in the graph, this embodiment can make the features of the evidence node propagate to the fact nodes it supports.

[0136] More specifically, in this embodiment, the first convolution module can determine the initial features of each fact node based on the edge weights of the fact node and the associated evidence nodes and the features of the associated evidence nodes. The associated evidence nodes are the evidence nodes connected to the fact node in the evidence fact association graph.

[0137] Considering that within the set of evidence nodes connected to a fact node, some evidence nodes may strongly support that fact node, while others provide weak support, determining the initial characteristics of the fact node after fusing evidence based on these weakly supporting evidence nodes (i.e., noisy evidence nodes) could be misleading and reduce the credibility of the fact node. Therefore, alternatively, when determining the initial characteristics of the fact node, only those evidence nodes that strongly support it can be considered; for example, only those evidence nodes whose edge weights to the fact node are greater than a preset weight threshold (e.g., 0.5).

[0138] For example, for any fact node (represented by fact node f) in the evidence fact association graph, its initial characteristics can be calculated using the following formula:

[0139] Formula (3);

[0140] in, The initial characteristics of fact node f are represented. This represents the set of evidence nodes connected to fact node f in the evidence-fact association graph. Indicates evidence node initial characteristics, This represents the activation function. Indicates the first bias term. This represents a minute constant, with values ​​ranging from 1 to 10. , Indicates evidence node The edge weight between the node and the fact node f.

[0141] As mentioned earlier regarding the judge's reasoning, in order to determine whether a fact is credible, in addition to evidence, there also needs to be a logical relationship between the facts, that is, there also needs to be a supporting relationship between the facts. In order to obtain the credibility of a fact that integrates both the value of evidence and the support of facts, this embodiment can make the characteristics of the fact node propagate to the fact node that supports it.

[0142] That is, through the second convolution module, for each fact node, the updated features of the fact node are determined based on the edge weights of the fact node and the associated fact nodes and the initial features of the fact node. The associated fact nodes are the fact nodes connected to the fact node in the evidence fact association graph.

[0143] Optionally, the update characteristics of fact node f can be calculated using the following formula (4):

[0144] Formula (4);

[0145] in, This represents the update characteristics of fact node f. This represents the set of fact nodes connected to fact node f (such as the set of fact nodes that have a temporal or logical relationship with fact node f). Representing fact node f and The strength of the association.

[0146] Finally, this embodiment can use the node classification module to determine the disputed focus nodes and undisputed nodes from each fact node based on the fact credibility in the update characteristics of each fact node and the contradictory relationships in the evidence fact association graph.

[0147] In one possible implementation, considering the process of feature propagation between nodes mentioned above, only the propagation of evidence to facts and the propagation between facts are considered. However, facts themselves can be credible or uncredible. For example, the fact that "cows fly in the sky" has extremely low credibility, but the fact that "cows are plowing the field" is relatively credible. Therefore, this embodiment also needs to comprehensively consider the credibility of the facts themselves (i.e., the initial confidence level predicted by the model mentioned above) in order to obtain a more realistic and reliable credibility for the fact nodes.

[0148] In other words, in this embodiment, when determining the disputed focus nodes and undisputed nodes from each fact node based on the fact credibility in the update features of each fact node and the contradictory relationships in the evidence fact association graph, the final credibility of the fact node can be determined first based on the fact credibility in the update features of the fact node and the initial confidence level in the fact elements corresponding to the fact node. Then, the disputed focus nodes and undisputed nodes can be determined from each fact node based on the final credibility of each fact node and the contradictory relationships in the evidence fact association graph.

[0149] For example, optionally, the final confidence level can be obtained by weighted summing of the fact credibility in the updated features of the fact node and the initial confidence level in the fact elements corresponding to the fact node, as shown in the following formula (5).

[0150] Formula (5);

[0151] in, This represents the final credibility of fact node f. This represents the credibility of the facts in the update features of fact node f. This represents the initial confidence level of the fact element corresponding to fact node f. This represents the weighting coefficient, which can be obtained through training a graph neural network model.

[0152] Furthermore, in order to identify the key points of contention, this embodiment can identify whether there are multiple contradictory factual nodes among the factual nodes that ultimately have a credibility greater than a preset credibility threshold. If so, these multiple factual nodes are determined as key points of contention. That is, this embodiment can automatically identify those highly credible yet contradictory factual nodes that are strongly supported by evidence from different witnesses; these nodes are potential key points of contention.

[0153] Furthermore, to obtain undisputed nodes, this embodiment can identify whether there are fact nodes among each fact node whose final credibility is greater than a credibility threshold and whose connected evidence nodes are uncontradictory. If so, the fact node is determined as an undisputed node. That is, this embodiment can identify those highly credible fact nodes that are strongly supported by the evidence of the plaintiff and defendant or two or more independent pieces of evidence. If these evidences of the plaintiff and defendant or two or more independent pieces of evidence are uncontradictory, then the highly credible fact node is an undisputed node.

[0154] Therefore, it can be seen that the node feature transmission process in the graph neural network model of this embodiment covers evidence feature aggregation, evidence propagation to facts, and propagation between facts, which is consistent with the judge's reasoning. Through this node feature transmission process, in-depth optimization of factual features is achieved, making the representation of factual nodes more consistent with the coherence of the case's factual chain and improving the accuracy of factual credibility.

[0155] By comprehensively considering the initial confidence level of the fact node itself, the credibility of the fact node after association with evidence, and the credibility of the fact after fact fusion, the final credibility level of the fact node is obtained. This final credibility level is more accurate and reliable. Then, based on this final credibility level, the focus of the dispute and the undisputed facts are identified, which improves the accuracy of the identification results.

[0156] In some other embodiments of this application, the process of step S204, "generating the trial findings text in the judgment based on the disputed focus nodes and undisputed nodes," is described.

[0157] In one possible implementation, the disputed key points and undisputed points to be decided can be marked on the evidence fact association graph. At the same time, the key evidence chains associated with the disputed key points and undisputed points can also be marked to obtain the fact determination graph.

[0158] Optionally, the key evidence chains associated with the disputed focus nodes and undisputed nodes can be determined in the following way: In this embodiment, in addition to the aforementioned factual credibility, the updated features also include the correlation between the factual node and each connected evidence node. Therefore, for each node among the disputed focus nodes and undisputed nodes, this embodiment can determine the evidence nodes whose correlation with the node is greater than a preset correlation threshold from the set of evidence nodes connected to the node. The evidence chain composed of the determined evidence nodes is used as the key evidence chain corresponding to the node, so as to obtain the key evidence chains corresponding to the disputed focus nodes and undisputed nodes respectively.

[0159] Optionally, the process of marking disputed nodes, undisputed nodes, and key evidence chains on the evidence fact association graph can be completed within the graph neural network model mentioned above, in which case the graph neural network model outputs the fact determination graph; of course, the marking process can also be completed outside the graph neural network model, in which case the graph neural network model outputs disputed nodes and undisputed nodes.

[0160] If the annotation process is completed within the graph neural network model, then after obtaining the fact-finding graph, this embodiment can structure it and encapsulate it into a hierarchical information package that can be deeply understood by the machine. The hierarchical information package includes at least: disputed focus nodes, undisputed nodes, and key evidence chains corresponding to each of the disputed focus nodes and undisputed nodes.

[0161] Optionally, the hierarchical information package may also include: index information of each piece of evidence in the key evidence chain corresponding to the point of contention, and index information of each piece of evidence in the key evidence chain corresponding to the undisputed point.

[0162] In one possible implementation, prompt words can be directly constructed based on "disputed key nodes, undisputed nodes, and the key evidence chains corresponding to each of the disputed key nodes and undisputed nodes" and input into a large language model (LLM) to generate the trial investigation text.

[0163] However, considering that formal trial investigation texts have a relatively fixed logical structure (such as procedural first, then substantive, undisputed first, then disputed), if this logical structure is not defined for LLM, the content generated by LLM may deviate from this inherent logical structure and lack professionalism.

[0164] To overcome this deficiency and ensure that the generated trial findings text conforms to both the inherent logical structure (such as procedural before substantive, undisputed before disputed) and actual needs, in another preferred implementation, this embodiment can pre-configure a trial findings template. This template is not static text, but a hierarchical configuration file consisting of logical rules, content slots, and generation strategies, including the following two-layer structure.

[0165] The first layer of structure is a macro-structure template used to define the overall chapter flow generated by the trial investigation. For example, [Procedural Matters] -> [Facts Undisputed by the Parties] -> [Dispute Focus 1: Presentation and Examination of Evidence and Fact Determination] -> ... -> [Comprehensive Summary of Facts]. Each square bracket "[]" represents a chapter. In this embodiment, the activation and order of each chapter can be automatically adjusted according to the case type.

[0166] The second layer of structure: chapter-level templates, used to define the internal logical structure of each chapter. Taking the previous "Dispute Focus 1: Presentation, Examination, and Fact Determination" as an example, the chapter-level content planning could be: "Regarding the issue of [Summary of Disputed Points]. The plaintiff [claimant] asserts [claim content], and submits the following evidence: [List and Explanation of Evidence]. The defendant [defendant] argues [defense content], and submits the following evidence: [List and Explanation of Evidence]. In response to the evidence presented by both parties, this court analyzes as follows: [Comparative Analysis Based on Evidence Validity and Factual Credibility], [A Tendency Factual Judgment is Arrived]."

[0167] At the same time, it is also necessary to define the generation specifications for the slots of each sentence in the second-level structure, so as to better understand what content should be filled in the slots and the format requirements. For example, for the slot "Listing and Explanation of Evidence", the generation specifications can be: list the evidence in descending order of its evidentiary weight, and the explanation of each piece of evidence must include "Evidence Name", "Purpose of Proof" and "Substantive Content of Evidence". In addition, the language of the statement should be an objective paraphrase and should not contain subjective words such as "we believe".

[0168] Furthermore, in accordance with the generation specifications, data instances of each slot corresponding to each chapter-level template of the trial and investigation template can be extracted from the disputed focus nodes, undisputed nodes, and the key evidence chains corresponding to the disputed focus nodes and undisputed nodes, and the data instances of each slot can be filled into the corresponding slots of the corresponding chapter-level template to obtain at least one chapter-level generation blueprint.

[0169] It should be noted that, due to the diverse nature of case types, the logical structure of the trial findings text may differ slightly between different case types. Therefore, different trial findings templates may be configured for different case types. Thus, when extracting data instances from each slot, the trial findings template used earlier must be the one corresponding to the target case. To obtain the trial findings template corresponding to the target case, this embodiment can match the trial findings template based on the case data of the target case (such as first-instance cases, second-instance cases, cause of action, etc.) as the corresponding trial findings template for the target case.

[0170] It should be understood that a chapter-level generation blueprint is essentially an operation sequence consisting of instructions, data fragments, and format requirements. Compared to "disputed focal points, undisputed nodes, and the key evidence chains corresponding to each of the disputed focal points and undisputed nodes," chapter-level generation blueprints are more standardized data that meet the requirements. Therefore, each chapter-level generation blueprint and pre-built prompt words can be concatenated into prompt instructions and input into a large language model to obtain the chapter-level text corresponding to each chapter-level generation blueprint.

[0171] Optionally, prompts may include: system role instructions, specific task instructions, structured context (i.e., chapter-level generated blueprints), and output format and style constraints.

[0172] Optionally, the specific task instruction is as follows: while ensuring that the logical structure of the chapter-level generation blueprint remains unchanged, perform textual processing on the structured evidence items and factual elements in each slot of the chapter-level generation blueprint to obtain the chapter-level text. In other words, the large language model performs conditional, fill-in-the-blank generation tasks under the strong constraints of the chapter-level generation blueprint, which greatly reduces the risk of the large language model's free interpretation and fabrication.

[0173] As mentioned earlier, the fact nodes in the evidence fact association graph are the finest-grained atomic facts. However, in actual application scenarios, there may be relationships between the atomic facts, such as causal relationships and relationships within the same transaction process. If each atomic fact is introduced one by one in the chapter-level text, on the one hand, the text may be too long, resulting in poor readability. On the other hand, the text readability may be further reduced by ignoring the relationships between the atomic facts.

[0174] Therefore, optional specific task instructions may also include: aggregating atomic facts that have causal relationships and the same transaction process relationship, and generating chapter-level text based on the aggregated facts.

[0175] Finally, the chapter-level text corresponding to at least one chapter-level generated blueprint is concatenated into a coherent text, which serves as the text for review and investigation. The concatenation methods include, but are not limited to, adding conjunctions and adjusting pronoun relationships.

[0176] This embodiment ensures that the large language model generates tasks according to the prescribed logical structure and structured evidence items and factual elements by configuring the trial ascertainment template and prompt words, thereby improving the accuracy and reliability of the generated trial ascertainment text.

[0177] In one optional embodiment, after generating the trial findings text, this embodiment can perform at least one of the following checks: factual consistency and legal language professionalism, to ensure the quality of the generated text.

[0178] This embodiment can first identify factual statements in the text ascertained during the trial. Specifically, sentences derived from factual elements are taken as factual statements.

[0179] Furthermore, in order to verify the consistency of facts, this embodiment can calculate the vector similarity between the factual statement and the disputed focus node and the undisputed node, and determine whether the factual statement is fabricated based on the vector similarity.

[0180] By retracing back to the points of contention and those without contention, it can be ensured that there are no fabricated or distorted expressions of the original facts in the text ascertained during the trial, thereby improving the objectivity and authenticity of the text ascertained during the trial.

[0181] To verify the professionalism of legal language, this embodiment can identify colloquial expressions in the trial investigation text and replace them with professional legal language. For example, by using a legal terminology database and style filters, the trial investigation text is scanned to identify colloquial expressions and replace them with professional legal language, ensuring that the terminology is standardized, objective, and rigorous.

[0182] To further improve the readability of the trial findings text, this embodiment can also perform at least one of the following post-processing steps on the trial findings text generated above:

[0183] The first method involves appending an evidence number after the factual statement, linking it to the corresponding evidence fragment. When a user clicks on the evidence number to view the fragment, the fragment is highlighted within the relevant evidence data. This method of linking evidence fragments by evidence number enables one-click tracing and improves the user's viewing experience of the trial findings.

[0184] The second approach is to highlight segments in the trial investigation text that are related to the key points of contention. This allows judges and other users to focus on these segments when reviewing the trial investigation text generated by the large language model, thus enabling them to promptly identify and correct errors when they occur at the key points of contention.

[0185] In addition, this embodiment can also provide a text editing interface, allowing judges and other users to modify, confirm, or request the regeneration of specified parts of the trial findings text through the text editing interface, such as correcting errors in segments related to the points of contention. Through human-computer collaborative processing, the trial findings text can better meet user expectations.

[0186] The above describes a method for generating trial and ascertainment text according to embodiments of this application. The following describes the apparatus for performing the above-described method for generating trial and ascertainment text.

[0187] Please see Figure 3 , Figure 3 This is a schematic diagram of a text generation device for adjudication and investigation provided in an embodiment of this application. Figure 3 As shown, the trial-founded text generation device may include:

[0188] Evidence organization unit 301 is used to acquire multiple pieces of evidence data in the target case and organize them into a structured evidence list. The evidence list includes multiple evidence items, and different evidence items are used to describe different pieces of evidence data.

[0189] The graph generation unit 302 is used to construct an evidence fact association graph with evidence items as evidence nodes and fact elements extracted from the evidence list as fact nodes. The edge relationships in the evidence fact association graph include contradictory relationships.

[0190] The graph update unit 303 is used to input the evidence fact association graph into the graph neural network model to perform feature transfer and aggregation between nodes, obtain the updated features of each fact node, and determine the dispute focus node and undisputed node from each fact node based on the fact credibility in the updated features and the contradictory relationship in the evidence fact association graph. The fact credibility is the value after the association evidence node and the feature aggregation of the fact node.

[0191] The text generation unit 304 is used to generate the trial findings text in the judgment based on the disputed focus nodes and undisputed nodes.

[0192] The aforementioned investigation revealed that each module in the text generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the operations corresponding to each module.

[0193] This application also provides an electronic device, which may include at least one processor and a memory connected to the processor, wherein:

[0194] Memory is used to store computer programs;

[0195] The processor is used to execute computer programs to enable electronic devices to implement any of the trial and ascertainment text generation methods provided in the embodiments of this application.

[0196] refer to Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0197] like Figure 4 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0198] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have instead.

[0199] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the adjudication and ascertainment text generation methods provided in this application.

[0200] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the trial and investigation text generation methods provided in this application.

[0201] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0202] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0203] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0204] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for generating trial investigation and verification text, characterized in that, include: Acquire multiple pieces of evidence data in the target case and organize them into a structured evidence list. The evidence list includes multiple evidence items, with different evidence items used to describe different pieces of evidence data. Using the evidence items as evidence nodes and the factual elements extracted from the evidence list as fact nodes, an evidence fact association graph is constructed, and the edge relationships in the evidence fact association graph include contradictory relationships. The evidence fact association graph is input into a graph neural network model for feature transfer and aggregation between nodes to obtain updated features of each fact node. Based on the fact credibility in the updated features and the contradictory relationships in the evidence fact association graph, disputed focus nodes and undisputed nodes are determined from each fact node. The fact credibility is the value after feature aggregation of the associated evidence nodes and the fact nodes. Based on the disputed key nodes and the undisputed nodes, the trial findings text in the judgment is generated.

2. The method for generating trial investigation text according to claim 1, characterized in that, The construction of an evidence fact association graph, using the evidence items as evidence nodes and the factual elements extracted from the evidence list as fact nodes, includes: The evidence items are used as evidence nodes, and the factual elements extracted from the evidence list are used as fact nodes to obtain a node set; Perform relation parsing on the nodes in the node set to determine the nodes with edge relationships and the edge relationships; For every two nodes in the node set that have the edge relationship: if they are two evidence nodes, the degree of support each of the two evidence nodes provides to the other is determined as the edge weight between the two evidence nodes; if they are two fact nodes, the correlation strength between the two fact nodes is determined as the edge weight between the two fact nodes; if they are an evidence node and a fact node, the semantic relevance between the evidence node and the fact node and the reliability of the evidence node are determined, and the edge weight between the evidence node and the fact node is determined based on the semantic relevance and the reliability. Based on the set of nodes, the edge relationships, and the edge weights, the evidence fact association graph is constructed.

3. The method for generating trial investigation text according to claim 2, characterized in that, Following the statement that the factual elements extracted from the evidence list will be used as factual nodes, the following is also included: The time elements in the fact elements corresponding to each fact node are processed into time information under a unified time standard, so as to determine the temporal relationship between each fact node based on the time information. Each fact node is matched with at least one preset logical rule, and the logical relationship between each fact node is determined based on the matching result. The at least one logical rule includes causal rules and process rules. The fact elements also include the subject representing the initiator of the fact action and the object representing the recipient or related object of the fact action. In the first fact node and the second fact node that satisfy the causal rule, the object of the first fact node and the subject of the second fact node have a semantic relationship. The time information of the first fact node precedes the time information of the second fact node. The subjects or objects of multiple fact nodes that satisfy the process rule are consistent, the time sequence is continuous, and the fact actions have a logically progressive relationship. Determining the association strength between two fact nodes includes: The association strength between the two fact nodes is determined based on the temporal and / or logical relationships between them.

4. The method for generating trial investigation text according to any one of claims 1-3, characterized in that, The step of inputting the evidence fact association graph into a graph neural network model for feature transfer and aggregation between nodes to obtain updated features for each fact node, and determining disputed focus nodes and undisputed nodes from each fact node based on the fact credibility in the updated features and the contradictory relationships in the evidence fact association graph, includes: The evidence evaluation module of the graph neural network model determines the evidentiary value of each evidence node based on the edge weights between the evidence node and the associated nodes, and determines the initial features of the evidence node based on the evidentiary value. The associated nodes include evidence nodes and fact nodes connected to the evidence node. Through the first convolutional module of the graph neural network model, for each fact node, the initial features of the fact node are determined based on the edge weights between the fact node and the associated evidence node and the features of the associated evidence node, wherein the associated evidence node is the evidence node connected to the fact node. Through the second convolutional module of the graph neural network model, for each fact node, the updated features of the fact node are determined based on the edge weights between the fact node and the associated fact nodes and the initial features of the fact node. The associated fact nodes are the fact nodes connected to the fact node. The node classification module of the graph neural network model determines the disputed focus nodes and undisputed nodes from the fact nodes based on the fact credibility in the updated features of each fact node and the contradictory relationships in the evidence fact association graph.

5. The method for generating trial investigation text according to claim 4, characterized in that, The step of determining disputed nodes and undisputed nodes from each of the fact nodes based on the fact credibility in the updated features of each fact node and the contradictory relationships in the evidence fact association graph includes: The final credibility of the fact node is determined based on the fact credibility in the updated features of the fact node and the initial confidence in the fact elements corresponding to the fact node. Based on the final credibility of each fact node and the contradictory relationships in the evidence fact association graph, disputed focus nodes and undisputed nodes are determined from each fact node.

6. The method for generating trial investigation text according to claim 5, characterized in that, The step of determining the disputed and undisputed nodes from the fact nodes based on the final credibility of each fact node and the contradictory relationships in the evidence fact association graph includes: Identify whether there are multiple fact nodes among the fact nodes that contradict each other and whose final credibility is greater than a preset credibility threshold. If so, then identify the multiple fact nodes as the focus of the dispute. If there is a fact node among the fact nodes whose final credibility is greater than the credibility threshold and whose connected evidence nodes are not contradictory, then the fact node is determined as an undisputed node.

7. The method for generating trial investigation text according to claim 1, characterized in that, The process of generating the trial findings text in the judgment based on the disputed key nodes and the undisputed nodes includes: For each of the disputed focal node and the undisputed node, evidence nodes with a correlation greater than a preset correlation threshold are determined from the set of evidence nodes connected to that node. The evidence chain composed of the determined evidence nodes is used as the key evidence chain corresponding to that node, so as to obtain the key evidence chains corresponding to the disputed focal node and the undisputed node respectively. The correlation is obtained from the update features of the node. From the disputed focus node, the undisputed node, and the key evidence chains corresponding to the disputed focus node and the undisputed node, extract the slot data instances corresponding to each chapter-level template of the trial and investigation template, and fill the corresponding slot data instances into the corresponding slots of the corresponding chapter-level template to obtain at least one chapter-level generation blueprint. Each chapter-level generated blueprint and pre-built prompt words are concatenated into a prompt instruction, which is then input into a large language model to obtain the chapter-level text corresponding to each chapter-level generated blueprint. The chapter-level text corresponding to the at least one chapter-level generated blueprint is concatenated into a coherent text, which serves as the text for investigation and verification.

8. The method for generating trial investigation text according to claim 1, characterized in that, Also includes: Identify factual statements in the text ascertained during the trial, calculate vector similarity between the factual statements and the disputed key nodes and the undisputed nodes respectively, and determine whether the factual statements contain factual fabrications based on the vector similarity. And / or, identify colloquial expressions in the trial investigation text and replace them with professional legal language.

9. The method for generating trial investigation text according to claim 8, characterized in that, Also includes: An evidence number is appended to the factual statement, which is used to link to the evidence fragment on which the factual statement depends and to highlight the evidence fragment in the corresponding evidence data; And / or, highlight the segments in the trial investigation text that are related to the disputed key points.

10. The method for generating trial investigation text according to claim 1, characterized in that, The evidence item includes one or more of the following evidentiary elements: the person presenting the evidence, the name of the evidence, the substantive content of the evidence, the purpose of the proof, and the location information of the substantive content of the evidence in the corresponding evidence data.

11. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the trial ascertainment text generation method as described in any one of claims 1 to 10.

12. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is configured to execute the computer program to enable the electronic device to implement the trial ascertainment text generation method as described in any one of claims 1 to 10.

13. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the trial and investigation text generation method as described in any one of claims 1 to 10.

Images