Lawyer recommendation system based on artificial intelligence

Abstract

The invention discloses a lawyer recommendation system and method based on artificial intelligence. According to the system, user case description and evidence files are obtained, and comprehensive case characterization is generated and vectorized after processing. On one hand, case similarity scores are calculated in a vector database through mixed search; and on the other hand, the historical agent cases of the lawyers are scored through a case scoring agent. And finally, the system fuses the similarity score, the case score and the user evaluation score, generates a comprehensive score, and recommends the most suitable lawyer to the user according to the comprehensive score. According to the lawyer recommendation method, the crossing from passive matching to active evaluation is realized, and the lawyer recommendation accuracy and the intelligent level are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and smart legal services technology, and specifically relates to a lawyer recommendation system and method that uses a hybrid search of vector databases and intelligent agent scoring. Background Technology

[0002] Currently, users seeking legal services face information asymmetry and high screening costs when searching for lawyers. Existing online legal platforms largely rely on lawyers' self-filled tags and users' keyword searches, resulting in a crude recommendation mechanism. Users, especially individual users, often need to bring complex evidence materials (such as contracts, invoices, photos of injuries, and PDF files) for consultation, but existing systems cannot effectively understand and process this unstructured information. Technically, although single-modal AI technologies (such as text classification and image recognition) are relatively mature, in the legal field, there is a lack of technical solutions that can deeply integrate text, voice, images, and documents for unified case understanding. Furthermore, existing lawyer recommendations are mostly based on static lawyer profiles, lacking an intelligent mechanism that can dynamically assess lawyer suitability for specific cases. Vector databases and similarity calculation technologies offer new ideas for case retrieval; for example, existing technologies include methods for case text matching based on models such as BERT. However, when these technologies are applied to lawyer recommendations, they often only address the question of "who has handled similar cases," failing to further answer the core question of "who is most suitable to handle my case." The latter requires incorporating the similarity between the user's cases and the lawyer's past cases, the lawyer's past case ratings, and the user's rating. Therefore, this invention aims to achieve precise lawyer recommendations based on "having handled similar cases," "having handled similar cases successfully," and "user satisfaction." The purpose of this invention is to provide a system and method capable of understanding complex case information, assessing the match between lawyers and cases, thereby achieving precise and personalized lawyer recommendations. Summary of the Invention

[0003] The core concept of this invention is to construct an intelligent lawyer recommendation system. This system can not only interact with users through natural language and comprehensively obtain case information, but also compare the case information with the judgments in the vector database for similarity. It also introduces a scoring agent that can score the lawyer's performance in the case based on the judgment. Finally, it combines the similarity score, the agent's score, and the user's evaluation score to recommend a lawyer.

[0004] Compared with the prior art, the present invention has the following significant advantages: More comprehensive information utilization: It breaks through the limitations of relying solely on text descriptions and can extract key information from user-uploaded images, PDFs, and other evidence, enabling a comprehensive and accurate understanding of the case.

[0005] The recommendation logic is more intelligent: a "scoring agent" has been introduced, which is no longer a simple matching of historical cases, but a dynamic assessment of the lawyer's potential to handle the case, making the recommendation reasons more persuasive.

[0006] More accurate recommendations: Hybrid search ensures the comprehensiveness and accuracy of case retrieval, while agent scoring ensures the suitability of lawyer selection. The combination of these two methods makes the recommendations more reliable. Detailed Implementation

[0007] The workflow of this system is as follows: Intelligent Agent Case Scoring and Data Entry: Preparation Phase: Input a massive amount of historical judgment documents in electronic format. Using a pre-trained model, identify and extract structured key legal elements; generate a high-dimensional representation vector for each case and cluster all case vectors. Generate Benchmark Representation: Calculate and generate the benchmark representation vector for each benchmark. Target Case Classification: Input a judgment document of a lawyer-represented case to be evaluated, extract the key legal elements of the case and quantify them. Calculate the similarity between this vector and the center of all benchmarks, and classify it into the benchmark with the highest similarity. Intelligent Comparison and Scoring: The intelligent agent compares the actual result of the target case with the benchmark representation vector of its respective benchmark item by item. Finally, output the case score and save the score in the database.

[0008] Information Input and Interaction Stage: Users describe the case background in text or voice through the system's multimodal interface and upload relevant evidence, such as photos of the scene, scanned contract PDFs, and videos. The system converts this non-textual information into processable data and guides users to supplement key information through dialogue. Case Understanding and Vectorization Stage: Using a model fine-tuned with legal domain data, the system encodes the fused text (including user descriptions, dialogue records, and evidence extraction text) to generate a semantically rich comprehensive case fact vector. Simultaneously, the legal element enhancement extraction module accurately extracts legal elements such as "case details," "place of occurrence," "evidence," and "amount in dispute" from this comprehensive representation and vectorizes them separately.

[0009] Database retrieval stage: Hybrid search: The hybrid search module uses extracted legal elements as keywords for initial screening, and then uses comprehensive case fact vectors to perform semantic similarity calculations in the candidate judgment set, finally outputting a set of similar historical judgments and their similarity scores.

[0010] Comprehensive Decision-Making and Recommendation Phase: The comprehensive recommendation module receives "similarity scores" from the hybrid search module and "case scores" from the agent scoring module. The system then follows a pre-set mathematical model... Calculation, assuming a lawyer is found One case, The professional expertise of the representing lawyer in the area where the user's case falls. It is an intelligent agent's score of the lawyer's performance in each case, which can intuitively reflect the gap between the lawyer's performance in that case and that of their peers. This is the semantic similarity value between the user's case and the lawyer's historical cases, reflecting the degree of matching between the user's case and the lawyer. It is a time decay function, which gradually decreases over time. Recent cases contribute more to the overall score, while the impact of older cases gradually decreases. It is a user rating score that reflects the lawyer's service level. It is a weighting of professional competence. This is a weighting of service capabilities. Finally, a comprehensive score is obtained for each candidate lawyer, and the system presents a list of recommended lawyers to the user in descending order of score.

[0011] Evaluation phase: After the service is completed, users can evaluate the lawyer, and the system will generate a user evaluation score based on the user evaluation. Example

[0012] Mr. Wang sought help due to a housing purchase dispute. He described the issue via voice on a mobile app: "After taking possession of my house, I discovered a serious water seepage problem," and uploaded a PDF of his purchase contract, screenshots of a video showing the seepage, and screenshots of his chat history with the developer. Process: The system converted the voice to text and extracted contract terms, the location of the seepage, and the timeline of communication from the PDF and images. The model understood this to be a "commercial housing sales contract dispute," with the core issue being "liability for quality defects," and key evidence being the "on-site photos" and the "quality guarantee clauses in the contract." Search and scoring: The system retrieved a batch of judgments related to "housing quality issues" from its vector database. Judgment A (similarity 0.92, score 98) was won by Attorney Zhang representing the homeowner; Judgment B (similarity 0.88, score 40) was lost by Attorney Liu due to insufficient evidence. Ultimately, after comprehensive calculation, the system recommended Attorney Zhang first to Mr. Wang.

[0013] The above description is merely a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural transformations made based on the inventive concept of this application and the content of this specification, or direct / indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. Claims An artificial intelligence-based lawyer recommendation system, characterized in that: include: The multimodal interaction and information processing module engages in dialogue with the user via text and voice to obtain a natural language description of the case and receives evidence files uploaded by the user, including images, documents, and videos. It performs an integrated understanding of the case description and evidence files, converting voice input into text and extracting text information from images, videos, and PDFs. It uses a model to generate a comprehensive case fact representation; extracts key legal elements from the comprehensive case fact representation, automatically prompting for answers when key elements are missing; and transforms the legal elements and the comprehensive case fact representation into a high-dimensional vector representation. The vector database stores a large number of legal document vectors. Each legal document vector is associated with an original case fact vector, a legal element vector, a case scoring label (generated by a scoring agent), and information on the court and the attorney representing the case. The hybrid search and similarity calculation module is used to combine keyword matching and vector similarity calculation to retrieve similar judgments from the vector database and generate a similarity score for each similar legal document. The case scoring agent module has a built-in scoring agent that is used to quantitatively analyze the litigation results, the submission of evidence and the adoption of defense opinions by the court based on the comprehensive factual representation and legal elements of the case to be processed, and generate a lawyer's ability score. The comprehensive recommendation module is used to integrate the similarity score, case score, and user evaluation score to generate a comprehensive score, and recommend the most suitable lawyer to the user based on this score.

2. The system according to claim 1, characterized in that, The case scoring agent module operates as follows: Preparation Phase: Inputting a massive amount of historical judgment documents into electronic files, using a pre-trained model to identify and extract structured key legal elements; generating a high-dimensional representation vector for each case and clustering all case vectors. Benchmark Representation Generation: Calculating the benchmark representation vector for each comparison benchmark. Target Case Classification: Inputting a judgment document of a lawyer-represented case to be evaluated, extracting the key legal elements of the case and quantifying them, calculating the similarity between this vector and the centers of all comparison benchmarks, and classifying it into the benchmark with the highest similarity. Intelligent Comparison and Scoring: The agent compares the actual result of the target case with the benchmark representation vector of its corresponding benchmark. Finally, output the score.

3. The system according to claim 1, characterized in that, The hybrid search and similarity calculation module employs a multi-stage hybrid retrieval strategy: Phase 1: Based on the information extracted from user descriptions, such as the cause of action, location of the court with jurisdiction, case details, and evidence, a preliminary search is conducted to construct a candidate set; Second stage: Use the comprehensive case fact vector to perform nearest neighbor search in the candidate set and calculate semantic similarity score; The third stage: generating the final semantic similarity score.

4. The system according to claim 1, characterized in that, The model in the multimodal information processing module is a pre-trained model. This model is fine-tuned by constructing a dataset in the corresponding format and setting fine-tuning parameters to understand the specific structure and semantics of legal documents and to act as a pre-sales legal service professional to communicate with users.

5. The system according to claim 1, characterized in that, The overall score of the comprehensive recommendation module is calculated using the following formula: Suppose a search finds a lawyer One case, The professional expertise of the representing lawyer in the area where the user's case falls. It is an intelligent agent's score of the lawyer's performance in each case, which can intuitively reflect the gap between the lawyer's performance in that case and that of their peers. This is the semantic similarity value between the user's case and the lawyer's historical cases, reflecting the degree of matching between the user's case and the lawyer. It is a time decay function, which gradually decreases over time. Recent cases contribute more to the overall score, while the impact of older cases gradually decreases. These are user reviews, reflecting the lawyer's service level. It is a weighting of professional competence. It is the service capability weight.

6. A lawyer recommendation method based on multimodal interaction and agent scoring, characterized in that, Includes the following steps: Obtain user case descriptions and evidence documents through interactive means; Processing and fusing multimodal information to generate a comprehensive representation of the facts of the case; Extract key legal elements from comprehensive representation and quantify them; Similar judgments are found in the vector database using a hybrid search method, and similarity scores are calculated. The case scoring AI is used to score the cases handled by lawyers. The final recommendation list is generated by combining similarity scores, case scores, and user review scores.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the lawyer recommendation method as described in any one of claims 1-6.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor, when executing the program, implements the lawyer recommendation method as described in any one of claims 1-6.

Images