Big model-based judge AI assistant system and data processing method thereof

Abstract

The invention relates to the technical field of information, in particular to a judge AI assistant system based on a large model and a data processing method thereof. According to the scheme, through the data acquisition module and the natural language processing module, the system can quickly collect and analyze a large amount of case data and legal instruments, and the information processing efficiency is greatly improved. Therefore, a judge can obtain key information of the case in a short time, so that a decision can be made more quickly. The natural language processing module adopts an advanced semantic understanding technology and can deeply analyze the content of legal instruments and extract implicit semantic information. The deep understanding ability of the method is far better than that of a traditional keyword matching method, and a more reliable basis is provided for subsequent legal analysis.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a judge AI assistant system based on a large model and a data processing method thereof. Background Art

[0002] With the rapid development of artificial intelligence technology, its application in the legal field is becoming increasingly widespread. Traditional legal assistance systems mainly rely on keyword matching and rule bases. Although they have improved the work efficiency of judges to a certain extent, they still have obvious deficiencies in handling complex cases and understanding the semantics of legal documents.

[0003] Currently, judges face the following challenges when handling cases:

[0004] Information overload: With the increase in the number of cases and the complexity of legal documents, judges need to deal with a large amount of information, which brings them tremendous work pressure.

[0005] Case complexity: Legal cases in modern society often involve multiple fields, requiring judges to have extensive knowledge reserves and in-depth analytical skills.

[0006] Therefore, there is an urgent need for an AI-assisted system that can fully understand legal documents, provide in-depth analysis, and have good explainability and adaptability to better support judges' decision-making process. Summary of the invention

[0007] In order to solve the above problems, the present invention discloses a battery thermal management system and a control method.

[0008] In one aspect, the present application discloses a judge AI assistant system based on a large model, comprising:

[0009] The data collection module is used to collect case data, legal documents and corresponding judicial information to form the first key data;

[0010] A natural language processing module, used for semantically understanding and analyzing the first key data to obtain the second key data;

[0011] The large model machine learning module is used to receive the second key data and input the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

[0012] Among them, the large model machine learning module includes the following trained models: legal document key information extraction model; answer content key information extraction model; bill key information extraction model and similar case recommendation model.

[0013] The architecture of the legal document key information extraction model includes:

[0014] The input layer is used to receive the region information of the context sentence, the context sentence token sequence, and the token sequence of the sentence to be predicted;

[0015] The multi-head attention layer is used to process the output data from the input layer to calculate the relevance and importance between different input elements;

[0016] The feedforward neural network layer is used to perform nonlinear transformation and feature processing on the output data of the multi-head attention layer;

[0017] The output layer is configured to receive the output of the feedforward neural network layer and generate the probability distribution of the sentence to be predicted belonging to each predefined area through a normalized exponential function, thereby achieving accurate regional classification of sentences in legal documents.

[0018] The bill key information extraction model includes a pre-configured dictionary, and can realize the recognition of new bill types by updating the dictionary.

[0019] On the other hand, the present application also discloses a data processing method of a judge AI assistant system, which specifically includes the following steps:

[0020] Collect case data, legal documents and corresponding judicial information to form the first key data;

[0021] Performing semantic understanding and analysis on the first key data to obtain second key data; wherein the semantic understanding and analysis includes performing semantic vectorization using a pre-trained language model based on a converter architecture;

[0022] Receive the second key data and input it into the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

[0023] The above scheme also includes the following steps:

[0024] A general information extraction model based on PaddlePaddle natural language processing is used to extract key information from legal documents. The training method of this general information extraction model includes:

[0025] Use legal documents containing at least judgments and pleadings as training data;

[0026] Divide legal documents into predefined areas;

[0027] Construct a neural network model based on an encoder-decoder architecture. The input of the model includes the sequence number of the region to which the context sentence belongs, the token sequence of the context sentence, and the token sequence of the sentence to be predicted.

[0028] Set the model output to be the probability distribution of the sentence to be predicted belonging to each predefined region;

[0029] Use the cross entropy loss function as the objective function of the model;

[0030] During the training process, a forced teaching method is used to train the model;

[0031] Through iterative training and optimization, the accuracy of the model in the task of extracting key information from legal documents is improved.

[0032] The predefined multiple areas include but are not limited to the header, litigation request, facts and reasons, list of evidence and fact determination.

[0033] The method of using forced teaching to train the model specifically includes:

[0034] In the training phase, the actual region sequence number in the training set is used as the input of the region to which the context sentence belongs;

[0035] In the prediction stage, the region sequence number of the first sentence is initialized to 0, and the region sequence number of the subsequent sentences uses the region corresponding to the maximum probability of the prediction result of the previous sentence.

[0036] The scheme of this application uses a data collection module and a natural language processing module, so the system can quickly collect and analyze a large amount of case data and legal documents, greatly improving the efficiency of information processing. This enables judges to obtain key information of the case in a short period of time and make decisions more quickly. The natural language processing module uses advanced semantic understanding technology, which can deeply analyze the content of legal documents and extract implicit semantic information. This deep understanding capability far exceeds traditional keyword matching methods and provides a more reliable foundation for subsequent legal analysis. The large model machine learning module integrates multiple professional models, including a legal document key information extraction model, an answer content key information extraction model, a bill key information extraction model, and a similar case recommendation model. This multi-model collaborative working method can analyze cases from multiple angles and provide comprehensive and in-depth legal insights. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the battery thermal management system in Example 1 of the present application;

[0038] Figure 2 This is a flow chart of the data processing method of the judge AI assistant system in Example 2 of this application;

[0039] Figure 3 This is a flow chart of the training method of the model in Example 2 of the present application;

[0040] Figure 4 This is a flow chart of the method for model training using a forced teaching method in Example 2 of the present application. DETAILED DESCRIPTION

[0041] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The principles and features of the present invention are described below in conjunction with the drawings. The examples are only used to explain the present invention and are not used to limit the scope of the present invention.

[0042] The terms "including" and other expressions with similar meanings in the specification or claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions, such as a process, method or system, or device comprising a series of steps or units is not limited to the listed steps or units.

[0043] Example 1: Figure 1 As shown, a judge AI assistant system based on a large model, the system includes:

[0044] The data collection module is used to collect case data, legal documents and corresponding judicial information to form the first key data; the data collection module can specifically collect data through multi-source data collection channels, including but not limited to the court case management system, electronic archives, public judgment document network, etc.;

[0045] A natural language processing module is used to perform semantic understanding and analysis on the first key data to obtain the second key data; a pre-trained language model based on the Transformer architecture (such as BERT, RoBERTa, etc.) can be used for semantic vectorization; named entity recognition (NER) is implemented to extract key entities in the legal text, and dependency parsing technology is used to understand the sentence structure;

[0046] The large model machine learning module is used to receive the second key data and input the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

[0047] Among them, the large model machine learning module includes the following trained models: legal document key information extraction model; answer content key information extraction model; bill key information extraction model and similar case recommendation model.

[0048] In a possible implementation, the legal document key information extraction model may be a sequence labeling model BiLSTM-CRF (bidirectional long short-term memory network + conditional random field);

[0049] The answer content key information extraction model can adopt text classification + named entity recognition model, such as BERT + fully connected layer (for classification) + Bi LSTM-CRF (for entity recognition)

[0050] The bill key information extraction model can use image text recognition + information extraction model

[0051] For example, CNN+LSTM+Attention (for OCR)+BERT+Pointer Network (for information extraction)

[0052] Similar case recommendation model: text similarity calculation + ranking model can be used, for example, Simes BERT network + Learning to Rank

[0053] In addition, the above models can share a unified knowledge base or vector database, including:

[0054] Shared word embeddings: All models can use the same set of word embeddings pre-trained in the legal field to ensure consistency in semantic understanding;

[0055] Knowledge Graph: The information extracted from each model can be aggregated into a unified legal knowledge graph for use by all models.

[0056] Cross-validation: The output of one model can be used as validation data for another model, helping to identify and correct errors.

[0057] For example: cross-validation of the legal document key information extraction model and the similar case recommendation model:

[0058] The legal document model extracts key information of the case (such as case type, parties involved, focus of dispute, etc.).

[0059] The similar case model recommends relevant cases based on this information.

[0060] If the recommended precedent is significantly different from the current case, it may mean that the information extracted by the legal instrument model is incorrect.

[0061] Example:

[0062] The legal document model mistakenly identified a civil dispute as a criminal case. The similar precedent model found that the recommended precedents were all civil cases, indicating that there might be an error.

[0063] Cross-validation of the bill key information extraction model and the legal document key information extraction model:

[0064] The bill model extracts key information on the bill (such as amount, date, and parties involved).

[0065] The legal document model extracts relevant information from case descriptions.

[0066] The information in both should be consistent.

[0067] Example:

[0068] The bill model extracted an amount of 1 million yuan, while the legal document model extracted an amount of 100,000 yuan from the case description. This inconsistency suggests that one of the models may be wrong.

[0069] Through this multi-dimensional cross-validation, the system can better identify potential errors and improve overall accuracy and reliability. Each model can serve as a verification tool for other models, forming a network that proofreads and complements each other. This not only helps correct errors in individual models, but also improves the robustness and consistency of the entire system, ultimately providing judges with more accurate and reliable auxiliary information.

[0070] In the specific scheme, the workflow of the large model machine learning module includes:

[0071] Receive and pre-process case-related documents;

[0072] Extract key information using models in the Big Model Machine Learning module, which includes:

[0073] a) Key information extraction model for legal documents;

[0074] b) Answer content key information extraction model;

[0075] c) Bill key information extraction model;

[0076] d) Similar case recommendation model;

[0077] Based on the extracted key information, the following operations are performed in parallel:

[0078] a) Use similar case recommendation models to recommend relevant historical cases;

[0079] b) Match relevant legal provisions;

[0080] c) Identify the main controversial issues in the case;

[0081] d) Analyze and assign weight to evidence;

[0082] Based on the above parallel operation results, legal reasoning is carried out to obtain preliminary legal analysis results;

[0083] Assess the potential risks of different sentencing options;

[0084] Generate legal analysis recommendations;

[0085] Convert analysis results into natural language reports;

[0086] An interactive question-and-answer function is provided using a key information extraction model to allow users to inquire about the report content.

[0087] Based on user feedback and final judgment results, each model in the large model machine learning module is continuously learned and updated;

[0088] Among them, each step collaborates through information flow, parallel processing, cross-validation and iterative optimization. Each model provides a confidence score for its output results, and finally provides users with a comprehensive analysis report containing case summary, key points of dispute, relevant legal provisions, evidence analysis, similar precedents and their implications, possible judgment plans and their basis, risk warnings and recommended reasons for judgment.

[0089] In a possible implementation, the architecture of the legal document key information extraction model includes:

[0090] The input layer is used to receive the region information of the context sentence, the context sentence token sequence, and the token sequence of the sentence to be predicted. By receiving the context sentence information through the input layer, the model can understand the context of the sentence to be predicted in the entire document.

[0091] The multi-head attention layer is used to process the output data from the input layer to calculate the relevance and importance between different input elements. The multi-head attention layer can capture the relevance between different input elements. Through this layer, the model can better understand the logical relationship and dependency between sentences.

[0092] The feedforward neural network layer is used to perform nonlinear transformation and feature processing on the output data of the multi-head attention layer; the feedforward neural network layer performs nonlinear transformation and feature processing on the data, which enhances the model's ability to recognize complex patterns;

[0093] The output layer is configured to receive the output of the feedforward neural network layer and generate the probability distribution of the sentence to be predicted belonging to each predefined area through the normalized exponential function, so as to achieve accurate regional classification of sentences in legal documents. The output layer uses the normalized exponential function (softmax) to generate the probability distribution, which can not only give the most likely classification result, but also provide the probability of each possible classification. This method provides more detailed classification information, which is helpful for the subsequent decision-making process.

[0094] The bill key information extraction model includes a pre-configured dictionary, which can be used to identify new bill types by updating the dictionary. The pre-configured dictionary provides the model with domain-specific prior knowledge, which helps improve the accuracy of key information extraction. In particular, for rare or special bill types, the dictionary can provide key identification clues.

[0095] Example 2: Figure 2 As shown, a data processing method of a judge AI assistant system specifically includes the following steps:

[0096] S101: Collect case data, legal documents and corresponding judicial information to form the first key data;

[0097] This step involves collecting various information related to the legal case. The collected data includes case data (which may include case description, party information, etc.), legal documents (such as indictments, defense statements, etc.) and relevant judicial information (such as relevant legal provisions, previous precedents, etc.);

[0098] S102: performing semantic understanding and analysis on the first key data to obtain second key data; wherein the semantic understanding and analysis includes performing semantic vectorization using a pre-trained language model based on a converter architecture;

[0099] Semantic vectorization can be performed using pre-trained language models based on transformer architectures (such as BERT, GPT, etc.). This means that the system converts text into numerical vectors that can be understood and processed by machines; this process is used to enable the system to understand the meaning of the text, not just identify keywords; the data processed in this step is called "second key data", which contains the semantic information of the original data;

[0100] S103: Receive the second key data and input it into the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

[0101] Receive the second key data and input it into the pre-trained legal models. These legal models may include specialized models such as case classification, legal text matching, and judgment prediction. The model processes this data and generates legal analysis results and recommendations. These outputs are intended to assist judges in case evaluation and judgment making, and may include recommendations for relevant legal texts, analysis of similar cases, possible judgment directions, etc.

[0102] like Figure 3 As shown, the above scheme also includes the following steps:

[0103] A general information extraction model based on PaddlePaddle natural language processing is used to extract key information from legal documents. The training method of this general information extraction model includes:

[0104] S201: taking legal documents including at least a judgment and a complaint as training data;

[0105] The training data includes legal documents such as judgments and complaints. This ensures that the model can learn the characteristics of various types of legal documents;

[0106] S202: Divide the legal document into a plurality of predefined areas;

[0107] Divide legal documents into predefined areas, such as header, claim, facts and reasons, etc., so that the model can understand the structure of the document;

[0108] S203: constructing a neural network model based on an encoder-decoder architecture, wherein the input of the model includes the sequence number of the region to which the context sentence belongs, the token sequence of the context sentence, and the token sequence of the sentence to be predicted;

[0109] The model is built using an encoder-decoder architecture, with input including the regional information of the context sentence, the context sentence, and the token sequence of the sentence to be predicted; this allows the model to take into account the context and position information of the sentence;

[0110] S204: Setting the model output as the probability distribution of the sentence to be predicted belonging to each predefined region; the model output is the probability distribution of the sentence to be predicted belonging to each predefined region; allowing the model to make a prediction on the region belonging to each sentence;

[0111] S205: Use the cross entropy loss function as the objective function of the model;

[0112] Use the cross entropy loss function as the objective function for multi-class prediction problems.

[0113] S206: During the training process, a mandatory teaching method is used to train the model;

[0114] To accelerate the convergence of the model and improve training efficiency;

[0115] S207: Improve the accuracy of the model in the task of extracting key information from legal documents through iterative training and optimization.

[0116] like Figure 4 As shown in FIG. 1 , the method of using the forced teaching method for model training specifically includes:

[0117] S301: In the training phase, the actual region sequence number in the training set is used as the input of the region to which the context sentence belongs;

[0118] During the training phase, the model uses the actual (true) region sequence numbers in the training set as input; no matter what prediction the model made in the previous step, the correct region sequence number will be used as input for the next step; this allows the model to learn the correct sequence and avoid error accumulation caused by wrong predictions in the early stages of training; this method can speed up the training process and help the model learn the correct pattern faster;

[0119] S302: In the prediction stage, the region sequence number of the first sentence is initialized to 0, and the region sequence number of the subsequent sentences uses the region corresponding to the maximum probability of the prediction result of the previous sentence.

[0120] In the actual prediction (inference) phase, the model no longer has access to the true region sequence number; for the first sentence of the document, the region sequence number is initialized to 0. This is because we do not have previous context information; for each subsequent sentence, the model uses the region with the highest probability in the previous sentence prediction result as the region sequence number input for the current sentence. This approach allows the model to continue with subsequent classification based on its own predictions without the true label.

[0121] The exemplary above legal document key information extraction model can be represented by the following mapping:

[0122] p=h(ctx_section,ctx_token_ids,input_token_ids)

[0123] Model input:

[0124] Divide a legal document (target documents include lawsuits, appeals, defenses and judgments) into the following areas (Sect i on):

[0125] S=['Header', 'Lawsuit', 'Facts and Reasons', 'List of Evidence', 'Fact Finding']

[0126] ctx_section:

[0127] The sequence number corresponding to the region to which the context sentence belongs.

[0128] ctx_token_ids: Where m represents the maximum sequence length. The token sequence obtained by word segmentation of the context sentence.

[0129] input_token_ids: Where m represents the maximum sequence length. The token sequence obtained by segmenting the sentence to be predicted.

[0130] Model output:

[0131] The probability that the sentence to be predicted belongs to each region.

[0132] Objective function and training method:

[0133] Objective function: Cross-entropy loss:

[0134]

[0135] Among them, N is the number of samples or batch size, which refers to the number of samples processed in one iteration.

[0136] M is the number of categories, which indicates the number of predefined areas (such as header, claim, facts and reasons, etc.)

[0137] The training method draws on the teacher-forcing technology commonly used in NMT models.

[0138] In the training phase, the ctx_section in the model input comes from the corresponding field in the training set. In the prediction phase, the initial value of ctx_section is 0 (representing the "header"), and the value of ctx_section thereafter is the model prediction result corresponding to the previous sentence (argmax). This solves the problem of overfitting very well. After optimizing the legal document element extraction model, the model recognition accuracy has increased from the initial 92.41% to the current 99.73%.

[0139] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims

1. A judge AI assistant system based on a large model, characterized in that: include: The data collection module is used to collect case data, legal documents and corresponding judicial information to form the first key data; A natural language processing module, used for semantically understanding and analyzing the first key data to obtain the second key data; The large model machine learning module is used to receive the second key data and input the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

2. The system according to claim 1, characterized in that The large model machine learning module includes the following trained models: a legal document key information extraction model; an answer content key information extraction model; a bill key information extraction model and a similar case recommendation model.

3. The system according to claim 2, characterized in that The architecture of the legal document key information extraction model includes: The input layer is used to receive the region information of the context sentence, the context sentence token sequence, and the token sequence of the sentence to be predicted; The multi-head attention layer is used to process the output data from the input layer to calculate the relevance and importance between different input elements; The feedforward neural network layer is used to perform nonlinear transformation and feature processing on the output data of the multi-head attention layer; The output layer is configured to receive the output of the feedforward neural network layer and generate the probability distribution of the sentence to be predicted belonging to each predefined area through a normalized exponential function, thereby achieving accurate regional classification of sentences in legal documents.

4. The system according to claim 2, characterized in that The bill key information extraction model includes a pre-configured dictionary, and can realize the recognition of new bill types by updating the dictionary.

5. A data processing method for a judge AI assistant system, characterized in that: The following steps are involved: Collect case data, legal documents and corresponding judicial information to form the first key data; Performing semantic understanding and analysis on the first key data to obtain second key data; wherein the semantic understanding and analysis includes performing semantic vectorization using a pre-trained language model based on a converter architecture; Receive the second key data and input it into the trained legal model to calculate and output legal analysis and suggestions to assist judges in case evaluation and judgment making.

6. The method according to claim 5, characterized in that The following steps are also included: A general information extraction model based on PaddlePaddle natural language processing is used to extract key information from legal documents. The training method of this general information extraction model includes: Use legal documents containing at least judgments and pleadings as training data; Divide legal documents into predefined areas; Construct a neural network model based on an encoder-decoder architecture. The input of the model includes the sequence number of the region to which the context sentence belongs, the token sequence of the context sentence, and the token sequence of the sentence to be predicted. Set the model output to be the probability distribution of the sentence to be predicted belonging to each predefined region; Use the cross entropy loss function as the objective function of the model; During the training process, a forced teaching method is used to train the model; Through iterative training and optimization, the accuracy of the model in the task of extracting key information from legal documents is improved.

7. The training method according to claim 6, characterized in that: The predefined multiple areas include but are not limited to the header, litigation request, facts and reasons, list of evidence and fact determination.

8. The training method according to claim 6, characterized in that: The method of using forced teaching to train the model specifically includes: In the training phase, the actual region sequence number in the training set is used as the input of the region to which the context sentence belongs; In the prediction stage, the region sequence number of the first sentence is initialized to 0, and the region sequence number of the subsequent sentences uses the region corresponding to the maximum probability of the prediction result of the previous sentence.

Images