Conversational class case retrieval method and system based on structured perception

Abstract

The invention relates to a structured perception dialogue type class case retrieval method and system, and the method comprises the steps: carrying out the matching of different structure parts of a dialogue and different structure parts of a legal instrument, extracting correlation signals between the dialogue and the legal instrument, carrying out the aggregation of the correlation signals, carrying out the calculation of a correlation score between the dialogue and the legal instrument, and carrying out the retrieval of the dialogue type class case. And sorting the correlation scores between the dialogues and the legal instruments in a descending order from large to small, and selecting the legal instruments ranking the first few for the user to use. According to the method, the dialogue type case retrieval performance can be effectively improved, and the efficiency of a legal information retrieval system is improved. Besides, according to the method, information is extracted and utilized in a complex legal text, so that the dependence on professionals is reduced, and the time cost and the possibility of operation errors are reduced.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and deep learning technology in legal artificial intelligence, and specifically to a structured perception conversational similar case retrieval method and system. Background Art

[0002] In recent years, legal case retrieval has attracted much attention in the legal information processing and information retrieval research communities. Its purpose is to retrieve supporting cases for a given query case and constitutes an important component of legal information systems. In practice, in addition to statutes, precedents of court decisions are the main legal materials in various legal systems. For example, in countries following the common law system, precedents are crucial for lawyers to prepare legal reasoning. Existing studies have shown that automatic systems not only perform search tasks more efficiently than lawyers, but also complete them more efficiently. With existing legal case retrieval systems, users need to issue queries to express their complex information needs, which is a difficult task for them. In legal case retrieval, information needs are more complex than ordinary web searches, and legal domain knowledge is required during the search process.

[0003] Conversational search is an emerging paradigm in information retrieval, which refers to the information retrieval system actively iteratively and interactively optimizing users' requests and search results. Such a system can help users better express their information needs and improve search performance, especially for complex and exploratory search tasks. The conversational search paradigm has been applied in various search scenarios, such as web search, product search, academic search, etc. In legal case retrieval, a typical complex information search scenario, studies have shown that the conversational search paradigm can help improve users' search experience in terms of query formulation, result checking, user satisfaction, and search success rate. To build a practical conversational legal search system, an important issue is how to retrieve relevant legal case documents based on the context of the search conversation.

[0004] In legal search dialogues, different types of speech often focus on different parts of structured legal documents. As the dialogue progresses, the relevant parts of each turn of dialogue are distributed in different paragraphs of case documents. However, most existing legal case retrieval models are unable to capture the complex corresponding structural information of dialogues and case documents and extract accurate relevant signals. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a structured perception conversational similar case retrieval method;

[0006] The present invention aims to effectively utilize case and dialogue structure information to improve the performance of conversational similar case retrieval.

[0007] The present invention also provides a structured-perception conversational similar case retrieval system.

[0008] Terminology explanation:

[0009] 1. Dialogue: The dialogue focused on by the present invention is generated in order to obtain judicial documents that meet the needs of users. The dialogue begins with a judicial question raised by the user (i.e., the initial question part of the dialogue). The subsequent dialogue process includes two types of interactions between the system and the user: the first is that the system generates clarification questions to further understand the user's information needs, and the user answers according to his or her own information needs (i.e., the clarification question and answer part of the dialogue); the second is that the system returns possible similar cases to the user, and the user gives further feedback to the system to obtain more accurate similar cases (i.e., the result of the dialogue - the feedback part). For specific examples, please see the attached. Figure 2 .

[0010] 2. Chinese pre-trained legal model LawFormer: Starting from the famous Chinese pre-trained model chinese-roberta-wwm-ext, Lawformer has been continuously pre-trained, especially optimized for long texts in the legal field. This process involves converting the RoBERTa architecture to the Longformer mechanism that is more suitable for long sequence processing. It can adapt to the contextual understanding and information extraction of large-scale legal texts, showing its great potential in legal terminology recognition, document summarization, and legal reasoning.

[0011] The technical solution of the present invention is:

[0012] A structured perception conversational similar case retrieval method, comprising:

[0013] Match the different structural parts of the conversation with the different structural parts of the legal document respectively, extract the correlation signals between the conversation and the legal document, aggregate the correlation signals, calculate the correlation scores between the conversation and the legal document, sort the correlation scores between the conversation and the legal document in descending order from large to small, and select the top-ranked legal documents for users to use.

[0014] Preferably, according to the present invention, different structural parts of the conversation and different structural parts of the legal document are matched respectively, and correlation signals between the conversation and the legal document are extracted; including:

[0015] The candidate legal documents are structured into: basic case facts, adjudication analysis process and judgment results;

[0016] The initial question part of the dialogue and the adjudication analysis process part of the candidate legal document are concatenated and input into the first relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part.

[0017] The clarification question and answer part of the conversation and the basic case facts part of the candidate legal document are concatenated and input into the second relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part.

[0018] The result of the dialogue - the feedback part is also divided into three parts: basic case facts, adjudication analysis process and judgment results. They are spliced ​​with the basic case facts, adjudication analysis process and judgment results of the candidate legal document respectively, and then input into the third relevant signal extractor respectively. The vector representation of each [CLS] is retained in turn, and the three obtained vector representations of [CLS] are merged through the maximum pooling layer as the relevant signal of the result of the dialogue - the feedback part.

[0019] Further preferably, the first relevant signal extractor, the second relevant signal extractor and the third relevant signal extractor are all Chinese pre-trained legal models LawFormer.

[0020] Further preferably, the correlation signal is obtained by a first correlation signal extractor as shown below:

[0021]

[0022] Wherein, o represents the connection of two sequences, e1 represents the relevant signal representation extracted from the initial question of the dialogue and the adjudication analysis process of the legal case, IQRE() represents the first relevant signal extractor, and u1 represents the text of the initial question of the dialogue; [SEP] is used as a special marker in natural language processing to indicate the separation between sentences or texts; c h Represents the text of the judicial analysis process of a legal case.

[0023] Further preferably, the correlation signal is obtained by a second correlation signal extractor as shown below:

[0024]

[0025] Among them, e i is the correlation signal representation extracted from the i-th system-user utterance pair and the basic case facts, QARE() represents the second correlation signal extractor; a i-1 is the system's i-1th response, u i is the user's i-th reply; c f It refers to the basic facts of the case part of the candidate legal document;

[0026] Further preferably, if the i-th system-user utterance pair is part of a result-feedback pair, then a i-1 It is a legal case document, including the basic facts of the case. f i-1、Referee analysis processa h i-1 and the judgment result d i-1 ; User's i-1 Unsatisfied, therefore, provide feedbacku i , a i-1 Each part, user feedback i and the corresponding part of the candidate legal case c are connected together as the input of the third relevant signal extractor; as shown below:

[0027]

[0028] Wherein, RFRE() is the third correlation signal extractor, It is the relevant signal representation extracted from the basic case facts of the returned result case, user feedback, and the basic case facts of the candidate case; It is the basic case facts for returning the result; It is the relevant signal representation extracted from the referee analysis process part of the returned result case, the user feedback and the referee analysis process part of the candidate case; It is the referee analysis process of the returned result case; c h It is the adjudicative analysis process of candidate cases; It is the relevant signal representation extracted from the judgment result part of the returned result case, the user feedback and the judgment result part of the candidate case; It is the judgment result part of the returned result case; c d It is the judgment result part of the candidate case;

[0029] The first [CLS] embedding represents the relevance signals extracted from user feedback and the basic case facts of the result document and the candidate document; the other two [CLS] embeddings represent the relevance signals from user feedback and the judgment analysis process / judgment result part respectively;

[0030] Combine the above embeddings through a max pooling layer To obtain the final correlation signal e i ; as shown below:

[0031]

[0032] Preferably, according to the present invention, the correlation signal is aggregated; comprising:

[0033] Summarize the relevant signal e1 extracted from the initial question and the relevant signal e extracted from the i-th system-user utterance pair i ; Then, calculate the correlation signal (e1,…,e |q|), |q| is the number of turns in the conversation, expressed as the aggregated correlation:

[0034] r avg =AvgPooling(e1,…,e |q| )

[0035] Among them, r avg Indicates that the aggregated correlation is loud, and AvgPooling() represents the average pooling layer, which is used to obtain the average vector of the correlation signal;

[0036] The summarized relevance representation is input into a two-layer multi-layer perceptron MLP with nonlinearity to calculate the final relevance score f(c|q), which is the relevance score between the conversation and the legal document:

[0037] f(c|q)=g a (W2·g a (W1·r avg +b1)+b2)

[0038] Among them, W1, W2 are weight matrices, b1, b2 are bias vectors, g a is the tanh function.

[0039] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a structured-aware conversational similar case retrieval method when executing the computer program.

[0040] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a structured-aware conversational similar case retrieval method.

[0041] A structured perception conversational similar case retrieval system, comprising:

[0042] The correlation signal extraction module is configured to: respectively match different structural parts of the conversation with different structural parts of the legal document, and extract correlation signals between the conversation and the legal document;

[0043] A correlation score calculation module is configured to: aggregate the correlation signals and calculate the correlation score between the conversation and the legal document;

[0044] The recommendation module is configured to sort the correlation scores between the conversations and the legal documents in descending order, and select the top-ranked legal documents for users to use.

[0045] The beneficial effects of the present invention are:

[0046] The present invention can effectively improve the performance of conversational similar case retrieval and enhance the efficiency of the legal information retrieval system. In addition, by extracting and utilizing information from complex legal texts, the method reduces the reliance on professionals, reduces time costs and the possibility of operational errors. By improving the model's understanding of documents, the method makes the retrieved legal documents not only correct in form, but also highly consistent with the user's needs at the legal knowledge level. The method is applicable to a variety of legal dialogue application scenarios, such as similar case retrieval, judgment prediction and legal consultation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of a structured perception conversational similar case retrieval method of the present invention;

[0048] Figure 2 This is an example of the dialogue mentioned in the present invention; DETAILED DESCRIPTION

[0049] The present invention will be further defined below in conjunction with the accompanying drawings and embodiments, but is not limited thereto.

[0050] Example 1

[0051] A structured perceptual conversational case retrieval method, such as Figure 1 As shown, including:

[0052] Match the different structural parts of the conversation with the different structural parts of the legal document respectively to more accurately extract the correlation signals between the conversation and the legal document. Aggregate the correlation signals, calculate the correlation scores between the conversation and the legal document, sort the correlation scores between the conversation and the legal document in descending order from large to small, and use them as the basis for correlation sorting. Select the top-ranked legal documents for users to use.

[0053] Example 2

[0054] The difference between the structured perception conversational similar case retrieval method described in Example 1 is that:

[0055] Match different structural parts of the conversation with different structural parts of the legal document to more accurately extract the correlation signal between the conversation and the legal document; including:

[0056] The candidate legal documents are structured and divided into: basic case facts, judicial analysis process and judgment results; these three parts are retained in the document.

[0057] The initial question part of the dialogue (see the attached example) Figure 2) and the judgment analysis process part of the candidate legal document are spliced ​​and input into the first relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part. "[CLS]" is usually used as a special marker in natural language processing to indicate the beginning of a sentence or text. When training the model, a special "[CLS]" marker is added to the beginning of the input text to help the model recognize the overall meaning of the sentence and perform classification tasks.

[0058] The clarification question and answer part of the dialogue (see the attached example) Figure 2 ) is concatenated with the basic case facts of the candidate legal document and input into the second relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part.

[0059] The result of the dialogue - the feedback part (see the attached example) Figure 2 ) is also divided into three parts: basic case facts, adjudication analysis process and judgment results, which are spliced ​​with the basic case facts, adjudication analysis process and judgment results of the candidate legal documents respectively, and then input into the third relevant signal extractor respectively, and retain their respective [CLS] vector representations in turn. The obtained three [CLS] vector representations are merged through the maximum pooling layer as the result of the dialogue - the relevant signal of the feedback part.

[0060] The first relevant signal extractor, the second relevant signal extractor and the third relevant signal extractor are all Chinese pre-trained legal models LawFormer.

[0061] First, extract the correlation signal between the initial question and the legal case c. h are opinions and possible rulings on case issues. In most cases, legal issues arise when people want to understand or interpret possible opinions or rulings on a particular legal case. In the case study, it was observed that the relevant case documents are usually related to the initial legal issue u1. Therefore, they are concatenated together to obtain the correlation signal through the first correlation signal extractor as shown below:

[0062]

[0063] Among them, o represents the connection of two sequences, e1 represents the relevant signal representation extracted from the initial question of the dialogue and the judgment analysis process of the legal case, IQRE() represents the first relevant signal extractor, and the existing LawFormer model is directly used here. "[CLS]" is usually used as a special marker in natural language processing to indicate the beginning of a sentence or text; u1 represents the text of the initial question of the dialogue; [SEP] is used as a special marker in natural language processing to indicate the separation between sentences or texts; c hThe present invention uses a first relevance signal extractor and uses its [CLS] embedding as the relevance signal extracted from the initial question and the legal case document.

[0064] When the i-th system-user utterance pair (i.e., the combination of the i-1th response and the i-th response of the user) is a clarification question-answer pair, the basic case facts part is considered. f is a description of the case, such as where, when, and how the case occurred. When asking the user a clarification question, the answer is likely to be matched in the "facts" section of the relevant case file. Therefore, they are concatenated together to obtain the relevance signal through the second relevance signal extractor, as shown below:

[0065]

[0066] Among them, e i is the relevant signal representation extracted from the i-th system-user utterance pair and the basic case facts, QARE() represents the second relevant signal extractor, and the existing LawFormer model is also directly used here; i-1 is the system's response to paragraph i-1. In this section, this is a clarification question; u i is the user's ith response; in this case, it is the answer to the clarification question, c f refers to the basic case fact part of the candidate legal document; its [CLS] embedding is used as the relevance signal extracted from the clarification question-answer pairs and the legal case document.

[0067] If the i-th system-user utterance pair is part of a result-feedback pair, then a i-1 It is a legal case document, including the basic facts of the case. f i-1 、Referee analysis processa h i-1 and the judgment result d i-1 ; User's i-1 Unsatisfied, therefore, provide feedbacku i , in order to let the agent know or help the agent improve the results. From experience, it is found that it is difficult to predict how users will create feedback on legal case documents provided by agents. In other words, users may have multiple reasons to be dissatisfied with various parts of the document. For this reason, a i-1 Each part, user feedback i and the corresponding part of the candidate legal case c are connected together as the input of the third relevant signal extractor; as shown below:

[0068]

[0069] Among them, RFRE() is the third related signal extractor, and the existing LawFormer model is also directly used here; It is the relevant signal representation extracted from the basic case facts of the returned result case, user feedback, and the basic case facts of the candidate case; It is the basic case facts for returning the result; It is the relevant signal representation extracted from the referee analysis process part of the returned result case, the user feedback and the referee analysis process part of the candidate case; It is the referee analysis process of the returned result case; c h It is the adjudicative analysis process of candidate cases; It is the relevant signal representation extracted from the judgment result part of the returned result case, the user feedback and the judgment result part of the candidate case; It is the judgment result part of the returned result case; c d is the judgment result part of the candidate case; the first [CLS] embedding represents the relevance signal extracted from user feedback and the basic case facts of the result document and the candidate document; the other two [CLS] embeddings represent the relevance signals from user feedback and the judgment analysis process / judgment result part respectively;

[0070] Combine the above embeddings through a max pooling layer To obtain the final correlation signal e i ; as shown below:

[0071]

[0072] Aggregation of correlation signals; including:

[0073] After extracting relevance signals from each part of the dialogue sentence, we need to aggregate this information and create a single relevance score for each candidate case document for retrieval. Specifically, first, we aggregate the relevance signal e1 extracted from the initial question and the relevance signal e extracted from the i-th system-user utterance pair. i ; Then, calculate the correlation signal (e1,…,e |q| ), |q| is the number of turns in the conversation, expressed as the aggregated correlation:

[0074] r avg =AvgPooling(e1,…,e |q| )

[0075] Among them, r avg Indicates that the aggregated correlation is loud, and AvgPooling() represents the average pooling layer, which is used to obtain the average vector of the correlation signal;

[0076] Finally, the summarized correlation representation is input into a two-layer multilayer perceptron MLP with nonlinearity, which is an artificial neural network structure commonly used in machine learning tasks. This network structure consists of an input layer, at least one hidden layer, and an output layer. Each layer contains multiple neurons, where the hidden layer and the output layer usually introduce nonlinear activation functions. In a typical two-layer MLP, the input layer receives the raw data, passes the data to the hidden layer, the hidden layer applies a nonlinear activation function to introduce a nonlinear feature transformation, and then passes the output to the output layer, which may apply a different activation function to output the final prediction result. By introducing nonlinear activation functions, the multilayer perceptron can learn complex nonlinear relationships, enabling it to adapt to various data patterns and tasks. Calculate the final correlation score f(c|q), which is the correlation score between the conversation and the legal document:

[0077] f(c|q)=g a (W2·g a (W1·r avg +b1)+b2)

[0078] Among them, W1, W2 are weight matrices, b1, b2 are bias vectors, g a is the tanh function.

[0079] Example 3

[0080] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a structured-aware conversational similar case retrieval method described in Example 1 or 2 are implemented.

[0081] Example 4

[0082] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a structured-aware conversational similar case retrieval method described in Example 1 or 2 are implemented.

[0083] Example 5

[0084] A structured perception conversational similar case retrieval system, comprising:

[0085] The correlation signal extraction module is configured to: respectively match different structural parts of the conversation with different structural parts of the legal document to more accurately extract the correlation signal between the conversation and the legal document;

[0086] A correlation score calculation module is configured to: aggregate the correlation signals and calculate the correlation score between the conversation and the legal document;

[0087] The recommendation module is configured to sort the correlation scores between the conversations and the legal documents in descending order from large to small as a basis for correlation sorting, and select the top-ranked legal documents for use by the user.

Claims

1. A structured perception conversational similar case retrieval method, characterized in that: include: Match the different structural parts of the conversation with the different structural parts of the legal document respectively, extract the correlation signals between the conversation and the legal document, aggregate the correlation signals, calculate the correlation scores between the conversation and the legal document, sort the correlation scores between the conversation and the legal document in descending order from large to small, and select the top-ranked legal documents for users to use.

2. According to claim 1, a structured perception conversational similar case retrieval method is characterized by: Match different structural parts of the conversation with different structural parts of the legal document, and extract the correlation signal between the conversation and the legal document; including: The candidate legal documents are structured into: basic case facts, adjudication analysis process and judgment results; The initial question part of the dialogue and the part of the adjudication analysis process of the candidate legal document are concatenated and input into the first relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part; The clarification question and answer part of the dialogue and the basic case facts part of the candidate legal document are concatenated and input into the second relevant signal extractor, and the vector representation of [CLS] is retained as the relevant signal of this part; The result of the dialogue - the feedback part is also divided into three parts: basic case facts, adjudication analysis process and judgment results. They are spliced ​​with the basic case facts, adjudication analysis process and judgment results of the candidate legal document respectively, and then input into the third relevant signal extractor respectively. The vector representation of each [CLS] is retained in turn, and the three obtained vector representations of [CLS] are merged through the maximum pooling layer as the relevant signal of the result of the dialogue - the feedback part.

3. A structured perception conversational similar case retrieval method according to claim 2, characterized in that: The first relevant signal extractor, the second relevant signal extractor and the third relevant signal extractor are all Chinese pre-trained legal models LawFormer.

4. A structured perception conversational similar case retrieval method according to claim 2, characterized in that: The correlation signal is obtained by the first correlation signal extractor as follows: Wherein, o represents the connection of two sequences, e1 represents the relevant signal representation extracted from the initial question of the dialogue and the adjudication analysis process of the legal case, IQRE() represents the first relevant signal extractor, and u1 represents the text of the initial question of the dialogue; [SEP] is used as a special marker in natural language processing to indicate the separation between sentences or texts; c h Represents the text of the judicial analysis process of a legal case.

5. The structured perception conversational similar case retrieval method according to claim 2, characterized in that: The correlation signal is obtained by the second correlation signal extractor as follows: Among them, e i is the correlation signal representation extracted from the i-th system-user utterance pair and the basic case facts, QARE() represents the second correlation signal extractor; a i-1 is the system's i-1th response, u i is the user's i-th reply; c f It refers to the basic case facts part of the candidate legal document.

6. A structured perception conversational similar case retrieval method according to claim 2, characterized in that: If the i-th system-user utterance pair is part of a result-feedback pair, then a i-1 It is a legal case document, including the basic facts of the case. f i-1 、Referee Analysis Processa h i-1 and the judgment result d i-1 ; User's i-1 Unsatisfied, therefore, provide feedbacku i , a i-1 Each part, user feedback i and the corresponding part of the candidate legal case c as the input of the third relevant signal extractor; as shown below: Wherein, RFRE() is the third correlation signal extractor, It is the relevant signal representation extracted from the basic case facts of the returned result case, user feedback, and the basic case facts of the candidate case; It is the basic case facts for returning the result; It is the relevant signal representation extracted from the referee analysis process part of the returned result case, the user feedback and the referee analysis process part of the candidate case; It is the referee analysis process of the returned result case; c h It is the adjudicative analysis process of candidate cases; It is the relevant signal representation extracted from the judgment result part of the returned result case, the user feedback and the judgment result part of the candidate case; It is the judgment result part of the returned result case; c d It is the judgment result part of the candidate case; The first [CLS] embedding represents the relevance signals extracted from user feedback and the basic case facts of the result document and the candidate document; the other two [CLS] embeddings represent the relevance signals from user feedback and the judgment analysis process / judgment result part respectively; Combine the above embeddings through a max pooling layer To obtain the final correlation signal e i ; as shown below:

7. A structured perception conversational similar case retrieval method according to claim 6, characterized in that: Aggregation of correlation signals; including: First, we aggregate the relevant signal e1 extracted from the initial question and the relevant signal e extracted from the i-th system-user utterance pair. i ; Then, the correlation signal (e1,…,e |q| ), |q| is the number of turns in the conversation, expressed as the aggregated correlation: r avg =AvgPooling(e1,…,e |q| ) Among them, r avg Indicates that the aggregated correlation is loud, and AvgPooling() represents the average pooling layer, which is used to obtain the average vector of the correlation signal; Finally, the summarized relevance representation is input into a two-layer multi-layer perceptron MLP with nonlinearity to calculate the final relevance score f(c|q), which is the relevance score between the conversation and the legal document: f(c∣q)=g a (W2·g a (W1·r avg +b1)+b2) Among them, W1, W2 are weight matrices, b1, b2 are bias vectors, g a is the tanh function.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a structured-aware conversational similar case retrieval method described in any one of claims 1-7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a structured-aware conversational similar case retrieval method described in any one of claims 1-7 are implemented.

10. A structured perception conversational similar case retrieval system, characterized by: include: The correlation signal extraction module is configured to: respectively match different structural parts of the conversation with different structural parts of the legal document, and extract correlation signals between the conversation and the legal document; A correlation score calculation module is configured to: aggregate the correlation signals and calculate the correlation score between the conversation and the legal document; The recommendation module is configured to sort the correlation scores between the conversations and the legal documents in descending order, and select the top-ranked legal documents for users to use.

Images