Legal supervision clue mining method and system based on full-text retrieval and large model

Abstract

The present invention discloses a method and system for mining legal supervision clues based on full-text retrieval and large models, belonging to the fields of data collection, data retrieval and large model applications, and the method includes: S1. combing supervision rules and retrieval keywords; S2. full-text retrieval, using the keywords combed by the supervision topic and the full-text retrieval system to perform data retrieval, and collecting all relevant graphic and text data within a specified time period from multiple data sources; S3. information distillation and information filtering; S4. building a knowledge base for topic analysis; S5. knowledge base retrieval and sorting; S6. large model intelligent analysis and clue mining. The present invention collects basic data by setting keywords for automatic full-text retrieval, and then through the automatic element parsing and dialogue reasoning analysis of the large model, it can mine valuable legal supervision clues from massive data, thereby improving the efficiency of mining legal supervision clues based on graphic and text electronic data.

Description

Technical Field

[0001] The present invention relates to the technical fields of data collection and data retrieval, image recognition, knowledge base and large model application, and in particular to a legal supervision clue mining method and system based on full-text retrieval and large model. Background Art

[0002] In recent years, with the rapid development of information technology, especially artificial intelligence large language models, it has provided a feasible direction for automatic mining of clues for legal supervision. In 2017, Google proposed a neural network structure based on the self-attention mechanism - the Transformer architecture, which laid the foundation for the large model pre-training algorithm architecture. In 2018, OpenAI and Google released the GPT-1 and BERT large models respectively, and pre-trained large models became the mainstream in the field of natural language processing. At the end of 2022, OpenAI launched its ChatGPT with powerful natural language interaction and generation capabilities, which made a sensation around the world. In 2023, OpenAI's multimodal pre-trained large model GPT-4 was released, which has multimodal understanding and multi-type content generation capabilities. At the same time, domestic AI large models have also emerged, such as Wenxin Yiyan, Tongyi Qianwen, iFlytek Spark, Huawei Pangu, etc. The large model industry is showing a booming development trend. With the commercialization of AI large models launched by many technology companies, various general, industry and end-side large models have achieved remarkable results in many fields, such as finance, medical care, government affairs and other fields. AI large models have become an important means to improve service quality and efficiency. Large language models conduct large-scale pre-training on massive amounts of unlabeled data, allowing the model to learn a large amount of knowledge and fine-tune instructions, thus acquiring general problem-solving capabilities for a variety of tasks. The development of AI technology has been led by technologies related to large language models.

[0003] However, in actual legal supervision work, it has not been linked to information technology, especially artificial intelligence large language models. Traditional working methods are still used, and business documents can only be reviewed manually one by one. The workload is large, the efficiency is low and it is inaccurate, and the use of document data is relatively difficult. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a legal supervision clue mining method and system based on full-text retrieval and large models. Practitioners do not need to read the internal and external documents or file materials of various cases one by one. They can collect basic data by setting keywords for automatic full-text retrieval, and then use the automatic factor parsing and dialogue reasoning analysis of the large model to mine valuable legal supervision clues from massive data.

[0005] The object of the present invention is achieved through the following technical solution: a legal supervision clue mining method based on full-text retrieval and large model, comprising the following steps:

[0006] S1. Sort out supervision rules and search keywords: break down tasks based on legal supervision topics, sort out supervision rules and keywords for full-text search;

[0007] S101. Classify legal supervision issues into thematic categories, including case filing supervision, investigation supervision, adjudication supervision, and execution supervision;

[0008] S102. For different types of legal supervision issues, sort out whether to distinguish the causes of the case. For legal supervision issues that need to distinguish the causes of the case, sort out the supervision rules according to the different types of causes of the case. For legal supervision issues that do not need to distinguish the causes of the case, sort out the general supervision rules;

[0009] S103. Sort out the document types and keywords that need to be retrieved according to the different supervision rules sorted out in S102.

[0010] S2. Full-text search: Use the keywords and full-text search system of the supervised topic combing to search for data, and collect all relevant graphic and text data within a specified time period from multiple data sources;

[0011] S201. Obtain information about the collector of the full-text retrieval system, configure the data source (file path and database address), image recognition service, word segmenter, and configure the professional word library;

[0012] S202. Configure search keywords and time ranges according to different supervision types and case causes, start data collection, assemble into Flink tasks by parsing and collecting SQL, and collect structured data;

[0013] S203. Use image recognition services combined with keywords to retrieve and collect image data;

[0014] S204. Save the collected data in batches to Elasticsearch, and then perform the next batch data collection operation until the collected data is empty.

[0015] S3. Information distillation and information filtering: Use big models to perform content summarization, element analysis and information filtering on the collected data;

[0016] S301. Analyze and summarize the text data collected by S2 using the large model, and establish a three-level index of title-summary-paragraph;

[0017] S302. For the text data collected in S2, the key information of the specified type of documents is parsed and structured using a large model. The key information includes the information of the persons involved in the case, the factual elements, and the case circumstances in the legal documents, and then stored in the database. The information of the persons involved in the case and the relationship between the persons involved in the case are parsed and stored in the knowledge graph to provide a data source for the document comparison and personnel relationship mining required for legal supervision clue mining;

[0018] S303. Using OCR and image recognition technology, identify the text content in the image data collected in S2, and then use the large model in step S302 to summarize and structure the text content, and identify and mark key evidence elements such as fingerprints, seals, signatures, etc. in the image;

[0019] S304. Establish screening rules and perform information filtering on the data parsed and marked in the above steps S302 and S303 to screen out the case data that needs to be analyzed;

[0020] S4. Construct a knowledge base for topic analysis: Use the data from the previous large model summary and structured parsing to construct a hierarchically indexed vector knowledge base, a structured data knowledge base, and a graph data knowledge base;

[0021] S401. Based on the data obtained in all steps of S3, a multi-level association table of cases, documents, people and elements is constructed;

[0022] S402. For the text and summary data processed in step S301, use the Ebemding word embedding model to construct a vector database of title-summary-paragraph three-level indexes respectively;

[0023] S403. The structured data and graph data obtained by processing steps S303 and S304 are stored in the corresponding database, and query SQL and graph database query language templates are constructed.

[0024] S5. Knowledge base retrieval and sorting: Based on the vector knowledge base, structured data knowledge base and graph data knowledge base, a joint search method combining vector similarity search, structured query and knowledge graph search is used, and the top N data are extracted using the similarity value threshold, and then the retrieved data is sorted and screened using the rearrangement model;

[0025] S501. According to the needs sorted out in step S102 and the legal supervision rules for document analysis, a sentence for vector library retrieval is sorted out. The retrieval sentence is a natural language for searching target data; for example, "find all judicial documents with a certain plot";

[0026] S502. The vector library search statement constructed in step S501 is vectorized using the Embedding word embedding model. To improve the recall rate of the search, the search question can also be vectorized by using the large model to generate preset answers, and then the vector similarity searcher of the knowledge base is used for search. The knowledge base searcher will perform "document-paragraph-keyword" layer by layer according to the vector similarity search based on the three-layer index established in step S402 above, set the maximum number of returned or similarity threshold, and find the target data that meets the conditions;

[0027] S503. Use the reranking model bge-reranker-large to sort the data retrieved in the previous step S502, and take the top N data most relevant to the target question;

[0028] S504. According to the needs of S102, combined with the legal supervision rules of structured data analysis, build relational database and graph database query rules. For example, for the analysis topic of trial supervision, it is necessary to retrieve cases where the facts, circumstances, and applicable laws found in the prosecution and trial stages are inconsistent. The query template statement will be used to retrieve all case data that meet the conditions from the database built in step S403;

[0029] S505. The data meeting the conditions retrieved in S503 and S504 are integrated and stored using the association table of S401.

[0030] S6. Big model intelligent analysis and clue mining: Construct an analysis decision tree for the big model interactive dialogue, use the big model to analyze the data retrieved in the previous step one by one and draw conclusions, and mark and output the data for which supervision clues are found.

[0031] S601. For the types of legal supervision that need further analysis and filtering of the data obtained in S5 above, different large-model dialogue analysis decision trees are sorted out according to different situations:

[0032] S602. Through the large model multi-round dialogue analysis reasoning decision process constructed in S601 above, the possible results of the answer of each analysis node large model are tested and summarized, and the preset dialogue input and exception handling mechanism for the next decision analysis are prepared;

[0033] S603. Call the big model dialogue interface, use the dialogue mechanism organized by the above steps S601 and S602 to communicate with the big model, and record and warn the cases that the big model finally concludes are listed as legal supervision clues.

[0034] A legal supervision clue mining system based on full-text retrieval and large models, including data retrieval module, data analysis module, data storage module, rule configuration module, large model analysis and reasoning module and background management module;

[0035] The data retrieval module, i.e., the full-text retrieval module, includes data collection, data retrieval, image recognition, ES monitoring and operation, and retrieves and collects data related to the legal supervision topic from multiple data sources;

[0036] The data analysis module uses a large model to perform content summarization and element analysis on the data retrieved from the full text, and constructs a vector library, structured data and a graph database;

[0037] The data storage module is used to store various types of data in the process of data retrieval, data parsing, rule configuration and large model analysis in the legal supervision clue mining system;

[0038] The rule configuration module is used to configure the keywords of full-text search, the elements of large model analysis, and the dialogue flow and output conclusion rules of large model intelligent analysis;

[0039] The big model analysis and reasoning module uses the analysis and reasoning capabilities of the big model to summarize the content of the document, analyze the elements, and mine legal supervision clues based on the configured dialogue analysis process;

[0040] The background management module includes configuration management, data source management, word library management, and large model management. The configuration management unit is responsible for various system settings and parameter adjustments; the data source management unit ensures effective access and integration of data; the word library management unit maintains the update of professional terms and keywords; and the large model management unit supports the deployment and optimization of multiple large models, jointly ensuring the efficient operation and flexibility of the platform.

[0041] The beneficial effects of the present invention are: (1) The present invention supports the collection of multiple data sources and real-time data collection functions, and also supports the "search for images by text" function;

[0042] (2) The present invention can perform retrieval based on SQL statements and also provides relatively rich multi-combination keyword retrieval and multi-aggregation retrieval functions;

[0043] (3) The present invention uses the technology of the natural language big model to realize the automatic analysis of documents without annotation training, including the basic information of various cases, the facts determined, the case elements, etc.

[0044] (4) The present invention realizes the automatic analysis and judgment of the constituent elements of relevant legal supervision content by means of large-scale model dialogue reasoning, which greatly reduces the workload of business personnel in reading documents. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The overall flow chart constructed for the system of the present invention;

[0046] Figure 2 is a flow chart of the system operation of the present invention;

[0047] Figure 3 It is a full-text search acquisition flow chart of the present invention;

[0048] Figure 4 A flow chart of knowledge base construction and data screening for the present invention;

[0049] Figure 5 It is a flow chart of large model intelligent analysis of the present invention;

[0050] Figure 6 It is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0052] The present invention utilizes the powerful language pan-simplification ability of the big model, combined with the supervision rules sorted out by legal experts, to realize the automatic mining of legal supervision clues, greatly save manpower, improve work efficiency, and also make effective use of massive data resources; firstly, based on the full-text retrieval module, the electronic material data of various data sources and data formats are automatically collected with the keywords sorted out by the user, and then the big model is used to perform content summary, element analysis and information filtering on the collected data, and construct a hierarchically indexed vector knowledge base, structured data knowledge base and graph data knowledge base; then, based on the natural language questions and key elements raised by business experts, the knowledge base adopts a method combining vector similarity retrieval and knowledge graph search for joint retrieval to obtain accurately matched data; finally, according to the process of sorting out specific legal supervision topics, an analysis decision tree of the big model interactive dialogue is constructed, and the data retrieved from the knowledge base are analyzed one by one using the big model to draw conclusions, and finally the data with supervision problems are marked and output.

[0053] The full-text search module is a graphic and text search system that can search documents, pictures and database fields containing relevant content from a data source according to text or keywords. Baidu search can be used, or a self-developed graphic and text search system can be used. In the embodiment of the present application, the self-developed graphic and text search system includes:

[0054] The data collection module is used to obtain business data, document data and file data, and the collected data is used to provide support for data retrieval; the data collection module supports multi-data source collection and embedded object mode collection, and provides full collection, incremental collection, real-time collection, retry after collection failure, dirty data processing, and collection log recording functions;

[0055] Data retrieval module, used to provide data retrieval functions, perform multi-combination keyword query, multi-aggregation query and retrieval based on SQL statements;

[0056] The data source management module is used to manage structured and unstructured data sources, and also supports dynamic expansion of data source types through plug-ins;

[0057] The vocabulary management module is used to configure the word segmentation vocabulary for data retrieval and provide vocabulary isolation function to facilitate the use of different vocabulary configurations on different indexes;

[0058] ES monitoring and operation and maintenance module, which is used to provide a monitoring panel for ES cluster services and to perform dynamic expansion and reduction operations on ES clusters;

[0059] The system management module is used to implement user rights management of the system and provide data isolation function to isolate data between different businesses to avoid malicious access and the generation of junk data;

[0060] An image recognition module is used to identify target items in document files based on an image recognition model obtained through training of procuratorial industry image data, and realize the function of searching images by documents in combination with data retrieval function; the target items include a combination of one or more of fingerprints, seals, and signatures;

[0061] In this application, the method for performing image and text retrieval using an image and text retrieval system includes the following steps:

[0062] The first step is to deploy the image and text retrieval service:

[0063] 1. Prepare the basic environment;

[0064] 2. Environment installation. First, install and configure the JDK environment. Then install the MySQL service and initialize the database information of the image and text retrieval system. Then install the Elasticsearch service and initialize the account and password of the Elasticsearch service. Then install the image recognition service. At this time, you need to execute the image recognition training steps to obtain the trained image recognition model and configure the image recognition service address in the image recognition extension plug-in. Then configure the MySQL connection information, Elasticsearch connection information, and extension plug-in path information for the image and text retrieval service. Finally, start the image and text retrieval service.

[0065] Step 2: System user configuration: After logging into the system as a system administrator, you first need to create a role and a user, then log out and log in with the newly created user to avoid confusion in data retrieval between businesses, as well as dirty and useless data when the system is connected to multiple business environments;

[0066] Step 3: Data source configuration: configure the structured data source information and unstructured data source information that need to be collected. When using the image recognition function, you need to select an unstructured extension plug-in with image recognition function for unified management;

[0067] Step 4: Data collection configuration: Configure the structured data source and the associated unstructured data source to be collected in the data collector. In addition, the image file address needs to be specified in the data collection SQL. By default, the field value ending with '__URL' is used as the unstructured document address. The image service address also needs to be configured in the unstructured data source plug-in that supports image recognition.

[0068] Step 5: Data collection: Verify whether the database has archive logs enabled. If enabled, execute the real-time synchronization step based on Flink. Otherwise, execute the data extraction step, and then traverse the collected data to obtain the image files that need to be identified and save them locally.

[0069] Step 6: Image recognition result storage: Use the image recognition plug-in of the unstructured data source to call the image recognition service to identify the image content and save the recognition result data;

[0070] Step 7: Data retrieval: After waiting for data collection to be completed, query the required data information through the data retrieval step. If the object to be retrieved is an image index, the image address that matches the keyword will be returned.

[0071] Step 8. When image preview is required, deploy an nginx service locally and mount the image save address. Use the nginx proxy address plus the local address of the hit image to implement image preview.

[0072] In the above embodiment, the data extraction step includes:

[0073] A1. Obtain collector information and data source information;

[0074] A2, determine the current acquisition task is a real-time acquisition task or a scheduled acquisition task, if it is real-time acquisition, execute step A3, if it is a scheduled acquisition, execute step A4;

[0075] A3. In real-time collection, SQL is parsed first, the parsed SQL information is converted into Flink tasks, and Flink is used to synchronize data in real time.

[0076] A4. In scheduled data collection, first obtain the last data collection time as the start time of this data collection, then assemble SQL and use the SPI mechanism to load the corresponding structured data source driver, execute the data collection SQL to obtain the collected data, and then load the unstructured data source driver to extract the unstructured data;

[0077] Then save the collected data in batches to Elasticsearch, and then collect the next batch of data until the collected data is empty; if an error occurs during this process, record the collection failure information, and re-execute step A4 to retry the collection failure. If the number of failures of the current collection task is greater than 5, record the failure of the current collection task and manually check the cause; if not, update the last collection time and wait for the next collection task to be triggered.

[0078] In the above embodiment, the Flink-based real-time synchronization step includes:

[0079] B1. Determine whether the database supports archive log reading and whether archive log recording is enabled. If not, the current task cannot perform real-time synchronization tasks, but performs scheduled synchronization tasks;

[0080] B2. Create a thread to obtain the archive logs of the database, parse the logs, extract the table structure and data information required for data collection, and save them in memory to wait for the Flink task to consume the data. At the same time, check regularly whether the saved data will be consumed. If not, it means that the data is garbage data and needs to be discarded to avoid occupying a large amount of memory and causing memory overflow;

[0081] B3. Load the collector information, assemble it into a Flink task by parsing the collection SQL, consume the archived data, and save it to Elasticsearch.

[0082] In the above embodiment, the data retrieval step includes:

[0083] C1, determine whether the call is a SQL query, a multi-combination keyword search or a multi-aggregate search function, if it is a SQL query function, execute step C2; if it is a multi-combination keyword search function, execute step C3; if it is a multi-aggregate search function, execute step C4;

[0084] C2, obtain the retrieved SQL statement, perform SQL parsing, extract the query fields, associated indexes, filter conditions and query numbers, assemble them into DSL syntax for execution, and finally return the result data;

[0085] The default SQL search only supports single index and basic search, and does not support associated search and word segmentation search, so you need to install a SQL search plug-in to enhance the SQL search function;

[0086] C3, obtain multiple combinations of search conditions, support precise filtering, range filtering, word segmentation search, term search, exact match search, embedded object search, multi-level search, search sorting, search highlighting functions, encapsulate the search conditions into DSL syntax, perform search, and return result data;

[0087] C4. Obtain multi-aggregation search conditions, support indicator aggregation, and perform secondary aggregation operations on the aggregation results. Then, the aggregation query will be encapsulated into DSL syntax for aggregate retrieval.

[0088] In the above embodiment, the image recognition model training step includes:

[0089] D1. Before image recognition, you need to prepare training data and test data. The training data is a set of images that have been labeled, while the test data is unlabeled images used to evaluate the performance of the model.

[0090] D2. Start model training: In machine learning, a neural network is used as a model to learn the patterns and features of input data through training;

[0091] D3. After model training, use the trained model to perform image prediction. Given a new image, the model outputs a prediction result, indicating the probability that the image belongs to each category. Parameters will also be optimized and retrained based on the prediction results.

[0092] The vector knowledge base is a data structure and system that uses vectors to represent graphics and text, usually mathematical vectors in high-dimensional space, which can convert complex information into numerical form for easy calculation and comparison.

[0093] The natural language refers to a language that humans use for natural communication, such as Chinese and English, which is the main tool for human communication and thought description.

[0094] The large model refers to a deep learning model with ultra-large-scale parameters (usually more than one billion) (such as Zhipu ChatGLM3-9B, Baichuan2-13B, Puyuinternlm2.5-20B, Qianwenqwen1.5-72B, etc.), which has a complex computing structure that provides it with powerful expression, learning and generalized language understanding capabilities. It can handle more complex tasks and data such as reading comprehension in natural language processing, machine translation, text summarization, intelligent question answering, knowledge reasoning, etc.

[0095] like Figure 1 As shown, a general flow chart of a legal supervision clue mining method and system based on full-text retrieval and large model intelligent analysis:

[0096] First, the user sorts out the business rules and search keywords of the legal supervision topic, and then uses the full-text retrieval system to search data with keywords, and collects original graphic and text data from multiple data sources; then uses the big model to summarize the content, mark information and filter information of the collected data, and build a hierarchical indexed vector knowledge base, structured data and graph data knowledge base; then uses vector retrieval and knowledge graph search methods to search the knowledge base and filter out the most useful data; finally, uses the logical reasoning ability of the big model combined with the business logic of the legal supervision topic to analyze and judge one by one, and mark and output the supervision clues after they are found.

[0097] Combined with Figure 2 , a legal supervision clue mining method and system operation flow chart based on full-text retrieval and large model intelligent analysis:

[0098] After logging in, the user selects the topic and data source of legal supervision, and then sorts out the business rules and search keywords, and obtains all related original data from the full-text retrieval system by entering keywords; then enter the elements that need to be analyzed for knowledge distillation, and can filter out irrelevant data by configuring keywords or material names, or manually delete data that does not meet the conditions; then choose the type of subject knowledge base to be built, which can be vector, relational database and graph database; then the user enters the data filtering rules, and can directly enter natural language for the vector knowledge base, enter the element name for the relational database, or multi-table query SQL, and enter the entity and relationship names to be queried for the graph database; finally, based on the business rules, the user configures a large model analysis decision tree, and uses the large model to perform batch analysis on the filtered data and output the results.

[0099] Combined with Figure 3 , a legal supervision clue mining method and system based on full-text retrieval and large model intelligent analysis full-text retrieval data collection flow chart:

[0100] In the incremental data collection process, a collection task is created first, and when it is created, it is determined whether incremental log synchronization is supported. If it is supported, the Flink task is started to assemble the data and save it to the downstream ES. If it is not supported, it waits for the collection task to be triggered and determines whether the last collection time is before the collector creation time. If so, the full collection process is executed.

[0101] If the last collection time is after the creation time of the collector, SQL-based incremental synchronization is performed. First, the main query is used for data synchronization. When synchronizing, the associated query and embedded query need to be bound. Then, the incremental update SQL is executed for the associated query and embedded query respectively to obtain the updated data. The ES document ID is obtained through the associated field. If it is obtained, the data is updated. If it is not obtained, the data is discarded. The reason why the data that cannot be reversed is discarded here is that this means that the main query has not collected the associated master data. When waiting for the associated master data to be collected, all the associated query and embedded query data will be directly obtained, and there is no need to reversely obtain the data of the main query.

[0102] When a data source is added or modified, the archiving thread will be started. First, check whether the archive log can be obtained correctly. If not, record that the data source does not have the archive log enabled. If it can, record the archive log enabled by the data source and read the archive log. The archive log is read by using the debezium service to read the archive log of the data source in real time, save the DML data, and wait for the Flink task to consume the data. Figure 4 , a legal supervision clue mining method and system based on full-text retrieval and large model intelligent analysis, knowledge base construction and data screening flow chart:

[0103] For image data, you can first read the text through the OCR tool, and then call the image recognition tool to parse the image feature data. For example, fingerprints, seals, handwritten signatures, etc. can be directly stored in the structured database; for text data, the text summary and paragraph content can be obtained through large model analysis, and then the word embedding model can be used to build a hierarchical index vector to obtain a vector knowledge base; the large model prompt engineering can also realize text element analysis, extract various factual elements, entities and relationships in the text, and store them in the structured database and graph database respectively.

[0104] For vector databases, we input natural language to perform retrieval, and then filter out the top n data that best meet the conditions based on vector similarity and rearrangement models; for graph data knowledge bases that store entities and relationships, we can query relationships of specific types and directions, or nodes with specific labels and attributes; for relational data knowledge bases, we can combine multiple key elements for querying, and ultimately filter out data that is useful for mining legal supervision clues.

[0105] Combined with Figure 5 , a legal supervision clue mining method and system based on full-text retrieval and large model intelligent analysis flow chart of large model intelligent analysis:

[0106] Relying on the large model to realize the logical reasoning of the thinking chain through multiple rounds of dialogue, based on the useful data screened out by the knowledge base, the questions of each round of dialogue nodes are pre-configured in combination with the business logic. If the answer of each round of documents is negative, it ends directly. If the answer is positive, it enters the next round of dialogue. The input data of each round of dialogue may not be necessary until the conclusion is finally reached. Each round of dialogue is marked as a clue according to the actual business logic. After the dialogue ends, the next piece of data is processed. In general, in order to ensure the reliability of the conclusion of the large model dialogue, the number of dialogue rounds generally does not exceed 5 rounds.

[0107] Taking case filing supervision as an example, multiple cases that have not been filed are screened out through the previous steps from the case summary data. Assuming that there are three conditions for the filing of a certain type of case, a case should be filed if one of them is met, a large model with multiple rounds of dialogue is set up according to the case filing standards to make judgments on each piece of data.

[0108] Set up the first round of dialogue for the large model and enter:

[0109] "Known case facts: {Case summary}, condition 1 for filing a case is: {Case filing condition 1}, question: Does the case facts meet the conditions for filing a case?"

[0110] If the big model answers affirmatively, the case ends and condition 1 is met, and it is marked as a clue;

[0111] If negative, enter the second round of dialogue of the large model and enter:

[0112] "Condition 2 for filing a case is: {Condition 2 for filing a case}. Question: Does the above case meet the conditions for filing a case?"

[0113] If the big model answers affirmatively, the process ends, and case filing condition 2 is met, and the case is marked as a clue;

[0114] If the answer is no, enter the third round of dialogue of the large model and enter:

[0115] "Condition 3 for filing a case is: {Condition 3 for filing a case}. Question: Does the above case meet the conditions for filing a case?"

[0116] If the big model answers in the negative, it ends; if it answers in the affirmative, it is marked as a clue, which meets the case filing situation 3.

[0117] After the above three rounds of dialogue, it can be analyzed whether the case summary is a legal supervision clue of "should be filed but was not filed".

[0118] Combined with Figure 6 , a legal supervision clue mining method and system framework diagram based on full-text retrieval and large model intelligent analysis:

[0119] The bottom layer of the system is the infrastructure layer, which can use virtual cloud service resources or physical machine servers with GPU accelerator cards, configure sufficient storage and computing power resources, and adapt domestic CPUs and GPUs as needed according to actual conditions.

[0120] Above the infrastructure layer is the data layer, which can connect to a variety of structured databases, vector databases, and graph databases to store vector knowledge base data, structured data, and graph data with entities and relationships.

[0121] Above the data layer is the operation layer, which can provide the ability to develop and process structured data in various industries through standardized capabilities. DB-Manager implements access management of various databases (such as DM, MySQL, neo4j, etc.); Elasticsearch has the ability to quickly retrieve and collect massive data; LLM-API provides unified scheduling of all large models (including parameter configuration and unified large model API); Docker provides a virtualized software environment for the deployment and installation of multiple tools; the entire series of technical capabilities provide external service capabilities based on K8S unified hosting.

[0122] Above the operation layer is the business layer, which provides internal and external interfaces such as configuration management, large model management, data management, SQL proofreading, ES management, vocabulary management, rule configuration, and resource monitoring.

[0123] The top layer of the system is the interface layer. Customers can verify large model dialogues, data queries, and data configurations on the interactive pages of the application system. The system displays text data, structured data, and knowledge graph data during the entire process.

[0124] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A legal supervision clue mining method based on full-text retrieval and large models, characterized by: The following steps are involved: S1. Sort out supervision rules and search keywords: break down tasks based on legal supervision topics, sort out supervision rules and keywords for full-text search; S2. Full-text search: Use the keywords and full-text search system of the supervised topic combing to search for data, and collect all relevant graphic and text data within a specified time period from multiple data sources; S3. Information distillation and information filtering: Use big models to perform content summarization, element analysis and information filtering on the collected data; The step S3 comprises: S301. Analyze and summarize the text data collected in S2 using the big model, and establish a three-level index of title-summary-paragraph; S302. For the text data collected in S2, the key information of the specified type of documents is parsed and structured using a large model. The key information includes the information of the persons involved in the case, the factual elements, and the case circumstances in the legal documents, and then stored in the database. The information of the persons involved in the case and the relationship between the persons involved in the case are parsed and stored in the knowledge graph to provide a data source for the document comparison and personnel relationship mining required for legal supervision clue mining; S303. Using OCR and image recognition technology, identify the text content in the image data collected in S2, and then use the large model to parse the key information of the text content and structure it according to S302, and identify and mark the key evidence elements; the key evidence elements include fingerprints, seals, and signatures in the image; S304. Establish screening rules and perform information filtering on the data parsed and marked in S302 and S303 to screen out the case data that needs to be analyzed; S4. Construct a knowledge base for thematic analysis: Use the data from the document content summary, element analysis, and information filtering in the previous step to construct a hierarchically indexed vector knowledge base, a structured data knowledge base, and a graph data knowledge base; The step S4 comprises: S401. Based on the data obtained in S3, a multi-level association table of cases, documents, people and elements is constructed; S402. For the text and summary data processed in S301, use the Ebemding word embedding model to construct a vector database of title-summary-paragraph three-level indexes respectively; S403. The structured data and graph data obtained by processing in S303~S304 are stored in the corresponding database, and query SQL and graph database query language templates are constructed; S5. Knowledge base retrieval and sorting: Based on the vector knowledge base, structured data knowledge base and graph data knowledge base, a joint search method combining vector similarity search, structured query and knowledge graph search is used, and the top N data are extracted using the similarity value threshold, and then the retrieved data is sorted and screened using the rearrangement model; The step S5 comprises: S501. According to the supervision rules, the vector library search statements are sorted out, and the search statements are natural languages ​​that need to find the target data; S502. The vector library search sentence constructed in S501 is vectorized using the Embedding word embedding model. To improve the recall rate of the search, the search question is vectorized by using the large model to generate a preset answer. Then, the vector similarity searcher of the knowledge base is used for the search. The knowledge base searcher will perform a "document-paragraph-keyword" layer by layer based on the vector similarity search according to the three-layer index established in S402, set the maximum number of returned items or the similarity threshold, and find the target data that meets the conditions; S503. Sort the data retrieved in S502 using the reranking model bge-reranker-large, and take the top N data most relevant to the target question; S504. According to the needs sorted out in S102 and the legal supervision rules for structured data analysis, build relational database and graph database query rules, and retrieve all case data that meet the conditions from the database built in S403; S505. The qualified data retrieved from S503 to S504 are integrated and stored using the association table of S401; S6. Big model intelligent analysis and clue mining: Build an analysis decision tree for the big model interactive dialogue, use the big model to analyze the data retrieved in the previous step one by one and draw conclusions, and mark and output the data for which supervision clues are found; The step S6 comprises: S601. For the legal supervision types obtained from S5, further analysis and filtering of the data is performed, and different large-model dialogue analysis decision trees are sorted out according to different situations: S602. Through the large model multi-round dialogue analysis reasoning decision process built in S601, the possible results of the answer of the large model at each analysis node are tested and summarized, and the preset dialogue input and exception handling mechanism for the next step of decision analysis are prepared; S603. Call the big model dialogue interface, use the dialogue mechanism organized by S601~S602 to communicate with the big model, and record and warn the cases that the big model finally concludes are listed as legal supervision clues.

2. The legal supervision clue mining method based on full-text retrieval and large model according to claim 1 is characterized by: The step S1 comprises: S101. Classify legal supervision issues into thematic categories, including case filing supervision, investigation supervision, adjudication supervision, and execution supervision; S102. For different types of legal supervision issues, sort out whether to distinguish the causes of the case. For legal supervision issues that need to distinguish the causes of the case, sort out the supervision rules according to the different types of causes of the case. For legal supervision issues that do not need to distinguish the causes of the case, sort out the general supervision rules; S103. Sort out the document types and keywords that need to be retrieved according to the different supervision rules sorted out in S102.

3. The legal supervision clue mining method based on full-text retrieval and large model according to claim 2 is characterized by: The step S2 comprises: S201 obtains the information of the collector of the full-text retrieval system, configures the data source, image recognition service, word segmenter, and configures the professional vocabulary; the data source includes the file path and database address; S202. Configure search keywords and time ranges according to different supervision types and case causes, start data collection, assemble into Flink tasks by parsing and collecting SQL, and collect structured data; S203. Use image recognition services combined with keywords to retrieve and collect image data; S204. Save the collected data in batches to Elasticsearch, and then perform the next batch data collection operation until the collected data is empty.

4. A legal supervision clue mining system based on full-text retrieval and large models, using the method described in any one of claims 1 to 3, characterized in that: It includes data retrieval module, data parsing module, data storage module, rule configuration module and large model analysis and reasoning module; The data retrieval module, i.e., the full-text retrieval module, is used to retrieve and collect data related to the legal supervision topic from a variety of data sources; The data analysis module uses a large model to perform content summarization and element analysis on the data retrieved from the full text, and constructs a vector library, structured data and a graph database; The data storage module is used to store various types of data in the process of data retrieval, data parsing, rule configuration and large model analysis in the legal supervision clue mining system; The rule configuration module is used to configure the keywords of full-text search, the elements of large model analysis, and the dialogue flow and output conclusion rules of large model intelligent analysis; The big model analysis and reasoning module uses the analysis and reasoning capabilities of the big model to summarize the content of the document and parse the elements, and mines legal supervision clues based on the configured dialogue analysis process.

5. The legal supervision clue mining system based on full-text retrieval and large model according to claim 4 is characterized by: The legal supervision clue mining system also includes a background management module: The background management module includes a configuration management unit, a data source management unit, a vocabulary management unit, and a large model management unit. The configuration management unit is responsible for various system settings and parameter adjustments; the data source management unit ensures effective access and integration of data; the vocabulary management unit maintains the update of professional terms and keywords; the large model management unit supports the deployment and optimization of multiple large models, jointly ensuring the efficient operation and flexibility of the platform.

Images