Lawyer resource dynamic adaptation and personalized recommendation method and platform
By using controlled RAG evidence gate classification and dynamic lawyer profiling, the problem of accurately mapping user-generated colloquial descriptions in existing technologies has been solved, enabling efficient and accurate lawyer resource recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-13
AI Technical Summary
Existing legal service platforms lack accurate mapping in handling users' colloquial descriptions of case causes, resulting in low matching accuracy, ignoring dynamic factors, inefficient recommendations, and a high risk of illusion.
By employing the controlled RAG evidence gate classification method, combined with a structured case knowledge base and dynamic lawyer profiles, and through multi-channel evidence weighting and case complexity self-adaptation, accurate case identification and lawyer recommendation are achieved.
It improves the accuracy and interpretability of case identification, dynamically adapts recommendation results, enhances the consistency between matching degree and actual user needs, and reduces recommendation time and the risk of illusion.
Smart Images
Figure CN121658641A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer software technology and relates to a method and platform for dynamic adaptation and personalized recommendation of lawyer resources for legal services. Background Technology
[0002] Controlled RAG (Retrieval-Augmented Generation) is a retrieval-enhanced generation method. In this invention, it refers to performing vector retrieval on user input, recalling relevant context from a structured case knowledge base, and strictly limiting the use of only the recalled content for judgment during the classification reasoning stage, prohibiting the use of external unrecalled information, in order to reduce the risk of illusion and ensure traceability.
[0003] Dynamic lawyer profiles refer to a multi-dimensional data model formed by combining static information about lawyers (years of practice, professional fields, education, law firm size, etc.) with real-time changes in business status (such as the number of cases currently being handled, capacity to handle cases, average response time in the past 30 days, etc.) to support real-time matching and ranking optimization in recommendation algorithms.
[0004] Internet legal service platforms (such as domestic platforms like Zhaofa.com, Lvtu, and Hualv, and overseas platforms like Avvo, Martindale-Hubbell, and Justia) generally offer lawyer search / sorting functions based on region, specialty, and user ratings. Mainstream technical approaches include directory-based tag-based search, keyword-matching information retrieval, simplified scoring models (star ratings, activity levels, number of reviews, etc.), and manual review / operational intervention. Some platforms support filtering by "case type / cause of action," but this largely relies on lawyers filling in tags or platform-mandated annotations.
[0005] General recommendation system technologies (collaborative filtering, content-based, learning-based ranking, and integration of multi-source behavioral features) have been explored for application in legal vertical scenarios, but there are difficulties in adaptation due to the highly professional nature of legal texts, the colloquial nature of user expressions, and various constraints (compliance, privacy, professional ethics).
[0006] The main drawbacks of existing technologies are as follows.
[0007] 1. There is a lack of effective means to map user-generated verbal descriptions to accurate case causes, making it difficult to reliably handle colloquial or incomplete case descriptions, leading to insufficient case cause identification and recall or misjudgment. There is also a lack of controlled reasoning based on authoritative definitions, resulting in weak interpretability and a high risk of hallucination.
[0008] 2. Low matching accuracy: Existing platforms mostly rely on users manually entering keywords (such as "contract dispute lawyer") to search, without conducting in-depth analysis of user needs (such as subdividing dispute types, amount involved, completeness of evidence, etc.), resulting in insufficient matching between recommended lawyers and actual needs.
[0009] 3. Recommendations that rely solely on static labels or shallow features ignore key dynamic factors such as case complexity, the amount in dispute, and the lawyer's current workload and responsiveness, resulting in insufficient matching accuracy.
[0010] 4. Inefficiency: Users need to manually browse a large amount of lawyer information (such as profiles and cases) for filtering, with an average matching time of more than 48 hours, and there is a lack of professional comparative analysis tools. Summary of the Invention
[0011] To address the problems existing in the prior art, the purpose of this invention is to provide a method and platform for dynamic adaptation and personalized recommendation of lawyer resources for legal services.
[0012] The key points of this invention are as follows.
[0013] 1. Controlled RAG Evidence Gate Classification Method Based on Case Type Classification and judgment are performed based on the context retrieved from the structured cause of action knowledge base, prohibiting cross-database reasoning. Multi-channel evidence weighting (definition / element matching, scene keyword matching, evidence document matching) is adopted, and the output is limited to the standard cause of action name and confidence level. It supports the upper limit of multiple cause of action outputs and domain fallback when the confidence level is low, thereby improving the accuracy and interpretability of cause of action mapping and reducing the risk of illusion.
[0014] 2. A lawyer recommendation method that integrates adaptive weighting based on case complexity and discrete scoring based on the amount in dispute. Under a unified scoring framework with a scale of 0-100 and no relation to the candidate pool, and combined with dynamic lawyer profiles (real-time workload, responsiveness, etc.), the weights of the three categories of "professional ability - activity level - experience background" are switched according to the complexity of the case; and discrete scores are added according to the five levels of target amount for the same or adjacent levels, so as to achieve stable ranking, matching of amount scale and personalized recommendation.
[0015] The technical solution of this invention is a method for dynamic adaptation and personalized recommendation of lawyer resources, the steps of which include: The preprocessing module desensitizes the input user requirement information to obtain user requirement text, and then extracts keywords from the user requirement text and standardizes them; The case identification module statistically analyzes the case causes and their frequency of occurrence in each selected data source, and determines the high-frequency case causes based on the frequency of occurrence; it constructs a refined knowledge base based on the selected high-frequency case causes, and the remaining case causes are classified into the long-tail set; it generates a standardized definition text segment for each high-frequency case cause and performs vectorization processing to obtain a high-frequency case cause vector knowledge base; Based on the standardized keywords obtained from the preprocessing module, the high-frequency case vector knowledge base is retrieved to obtain the standardized definition text segments corresponding to the K most relevant high-frequency cases to user needs; The user demand information and the standardized definition text segment are input into the large language model for classification and reasoning, and at most n candidate reasons with the highest confidence and all greater than a set threshold are output along with their confidence scores; if the confidence scores of the candidate reasons are all lower than the set threshold, the domain to which the user demand information belongs is output; the output of the large language model is sent to the recommendation module. Based on the amount of the target input by the user and the standardized keywords obtained by the preprocessing module, a structured feature is formed and sent to the recommendation module; based on the user's demand information, the case complexity is determined and sent to the recommendation module. The lawyer profiling module generates a profile of each lawyer based on their static and dynamic data. The recommendation module recommends matching lawyers to users based on each lawyer's profile and the information received.
[0016] Preferably, the standardized definition text segment includes an explanation segment, an alternative name dictionary, an applicable situation segment, an inapplicable situation segment, and an element and evidence segment.
[0017] Preferably, the user demand information includes case description, region, and preferences.
[0018] Preferably, the static data includes the lawyer's professional sub-field, years of practice, educational background, law firm size, and historical case database; the dynamic data includes the lawyer's current number of cases handled, capacity to handle cases, average response time in the past 30 days, and recent case handling activity in the target field.
[0019] Preferably, the method for classification reasoning using the large language model is as follows: combining the confidence of candidate causes of action with three types of evidence: definition / element matching, keyword / typical scene matching, and evidence word matching.
[0020] Preferably, the recommendation module recommends matching lawyers to users based on the profiles of each lawyer and the received information as follows: Taking each candidate case type c as the core, obtain the number of cases N_c handled by each lawyer for that candidate case type c, the success rate R_c of cases handled for that candidate case type c, the total workload of all types of cases in the past N years, years of practice Y, the cumulative amount of cases handled M, the quantitative value of educational level E_score, and the quantitative value of law firm size S_score; use the user's amount of cases U_amt and the lawyer's average amount of cases handled in candidate case types c L_amt_c as a reference for matching magnitude; then calculate the base total score BaseScore for each lawyer = W_prof × P + W_act × A + W_exp × E; where W_prof, W_act, and W_exp are weighting coefficients determined based on case complexity, P is the lawyer's professional competence score calculated based on N_c and R_c, A is the lawyer's activity score calculated based on the workload of all types of cases in the past N years, and E is the lawyer's experience background score calculated based on Y, M, E_score, and S_score; the lawyer's base total score BaseScore is updated based on the matching degree between the user's claim amount U_amt and the lawyer's average claim amount L_amt_c in case c; then, the updated base total score BaseScore is used to recommend matching lawyers for candidate case c; if the output of the large language model is the domain to which the user's demand information belongs, then the lawyer is recommended to match the user based on the total number of cases handled by the lawyer in that domain.
[0021] A platform for dynamic matching and personalized recommendation of lawyer resources, characterized in that it includes a preprocessing module, a case identification module, an evaluation module, a lawyer profiling module, and a recommendation module; The preprocessing module is used to desensitize the input user demand information to obtain user demand text, and then extract keywords from the user demand text and standardize them. The case identification module is used to statistically analyze the case causes and their frequency of occurrence in each selected data source, and determine the high-frequency case causes based on the frequency of occurrence; construct a refined knowledge base based on the selected high-frequency case causes, and classify the remaining case causes into a long-tail set; generate a standardized definition text segment for each high-frequency case cause and perform vectorization processing on it to obtain a high-frequency case cause vector knowledge base; The evaluation module retrieves the high-frequency case vector knowledge base based on the normalized keywords obtained from the preprocessing module, and obtains the standardized definition text segments corresponding to the K most relevant high-frequency case causes to the user's needs. Then, it inputs the user's needs information and the standardized definition text segments into a large language model for classification reasoning, outputting at most n candidate case causes with the highest confidence levels, all exceeding a set threshold, along with their confidence levels. If the confidence levels of all candidate case causes are below the set threshold, it outputs the domain to which the user's needs information belongs, and then sends the output of the large language model to the recommendation module. Furthermore, it forms structured features based on the user's input amount and the normalized keywords obtained from the preprocessing module and sends these features to the recommendation module. Finally, it determines the case complexity based on the user's needs information and sends this complexity to the recommendation module. The lawyer profiling module generates a profile of each lawyer based on their static and dynamic data. The recommendation module recommends matching lawyers to users based on each lawyer's profile and the information received.
[0022] The advantages of this invention are as follows: 1. More accurate and interpretable cause of action identification: By combining a structured cause of action knowledge base with controlled RAG reasoning, accurate classification can be made based on authoritative context, even when the user input is colloquial or incomplete; multi-channel evidence weighted output of standard cause of action names and confidence levels, thereby reducing cause of action misjudgment, improving lawyer routing accuracy, and ensuring traceability.
[0023] 2. Recommendation results dynamically adapt to case context: Introducing a weight matrix driven by case complexity and discrete bonus points for each amount in dispute, the recommendation results can be precisely adjusted according to the difficulty and amount of the case; combined with dynamic lawyer profiles (non-static features such as real-time load and response time in the past 30 days), the matching degree is improved to match the actual needs of users.
[0024] 3. Stable ranking and operable optimization: The system adopts candidate pool-independent full-database normalization and periodic parameter calibration to ensure stable recommendation results across batches and time periods; the sub-item weight matrix and scoring rules are operable and configurable, supporting canary releases and continuous iterative optimization without affecting system stability.
[0025] 4. Fast response speed, reducing user screening costs: Case identification, target amount extraction, and complexity assessment are processed in parallel, combined with vector preheating and hotspot caching, achieving an end-to-end SLA of ≤2 seconds; users can obtain the first batch of recommended lists and structured comparative analysis within seconds, reducing the time spent on manual screening. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the overall system architecture. Detailed Implementation
[0027] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0028] A. Overall Plan This invention comprises two main parts: a method for dynamically adapting lawyer resources and a personalized recommendation platform. For example... Figure 1 As shown, the system operates according to the process of "user needs analysis - feature structuring - lawyer profile matching - sorting and interpretation - output", with an online end-to-end response target of ≤2 seconds.
[0029] The system consists of the following subsystems / modules: 1. User input and preprocessing module (text / voice, multidimensional preference acquisition, anonymization); 2. Case identification RAG and knowledge base module (structured case database, semantic retrieval, controlled reasoning); 3. Target Amount / Complexity Extraction and Evaluation Module (Parallel Computation); 4. Lawyer Profiling Center (Static + Dynamic Profiling, Data Governance and Real-time Updates); 5. Ranking and Recommendation Module (Unified Normalization, Itemized Scoring, Dynamic Weighting, Scale-Based Scoring); 6. Explanation and Display Module (item score, weight, and reasons for recommendation).
[0030] B. User Needs Analysis 1. User Requirements Collection and Preprocessing Case descriptions (text / voice), location, and preferences (gender, law firm size, years of experience, etc.) are collected through the user interface. Voice input uses ASR transcription; sensitive fields are anonymized and used for recommendations only after authorization is obtained.
[0031] Keyword extraction and normalization are performed using a rewriting proxy, enhancing robustness against colloquial and incomplete information scenarios. Both the preprocessed results and the original text are then fed into the downstream processing.
[0032] 2. A legal RAG+PE system for precise matching of case types. This system executes three core steps sequentially: 1. Knowledge base construction and governance (based on multi-source legal data, a high-frequency cause-of-case knowledge base is constructed through screening, standardization, and vectorization); 2. Retrieval and recall (inputting standardized keywords, and obtaining the top-K relevant cause-of-case definitions through vector recall and reordering); 3. Controlled classification reasoning and output (inputting the user's original text + recall results, and outputting the qualified cause of case and confidence level or domain catch-all classification through constrained reasoning and confidence scoring).
[0033] (1) Knowledge base construction and governance: Inputs: historical case data, publicly available judgments, official case reason classification tables, and professional books; sub-processes are as follows: ① Data Filtering and Classification: The frequency of occurrence of all causes of action in historical data (since 2000) was statistically analyzed. Since there are a total of 2580 causes of action in court cases, searching all of them would significantly reduce search speed and accuracy. The top 300 high-frequency causes of action cover 99.4% of litigation scenarios; therefore, a refined knowledge base was built only for these high-frequency causes of action, with the rest grouped into a long-tail set.
[0034] ② Standardized definition of high-frequency causes of action: Generate a standardized definition text segment of ≤300 words for each high-frequency cause of action, including an explanation segment (precise explanation of legal concepts), a dictionary of alternative names (colloquial expressions and regional colloquialisms), an applicable circumstances segment (conditions for application of the cause of action), an inapplicable circumstances segment (conditions for excluding easily confused causes of action), and an elements and evidence segment (factual elements required for establishment and common types of evidence); the definition must meet the dual requirements of retrieval and judgment, implement dual review, consistency verification and version release, and continuously supplement colloquialisms and high-frequency phrases based on retrieval logs.
[0035] ③ Vectorization and Knowledge Base Construction: The standardized definition text segment of each high-frequency cause of action is vectorized to establish a searchable vector knowledge base for high-frequency causes of action.
[0036] (2) Retrieval and Recall: Input: Normalized keywords extracted from user descriptions by the rewrite agent. Processing method: Normalized keywords are retrieved using a low-latency vector retrieval method with an approximate nearest neighbor index (such as HNSW / IVF), and then combined with a lightweight semantic reorderer and domain consistency score optimization results to improve Top-k accuracy and noise reduction effect.
[0037] Output: TOP-K standardized definitions of the most frequent causes of action most relevant to the user description (for use in subsequent large model inference).
[0038] (3) Controlled classification reasoning and output: Input: Original user description, standardized definitions of the TOP-K high-frequency case causes retrieved; Sub-processes are as follows: Constrained reasoning: Uses a large language model for classification reasoning, and reasoning is based only on "user original text + RAG recall cause definition", prohibiting cross-database expansion and reducing illusions; Confidence score: A confidence level of [0,1] is formed by combining three types of evidence (definition / element matching, keyword / typical scenario matching, and evidence word matching). Starting from 0 points, the cumulative score is limited to the [0,1] range. Positive scores are added based on the degree of consistency between the definition / core elements, the degree of consistency between keywords / typical scenarios, and the suitability of evidence clues (e.g., high consistency of core elements +0.35 points). Negative scores are deducted for situations such as inapplicability, conflict with other more appropriate causes of action, incomplete elements, or weak relevance (e.g., triggering inapplicability -0.40 points).
[0039] Output results: After determining the highest-scoring candidate cause of action, only select the candidate outputs (maximum 3) that "meet the score standard and are not far from the highest score". The output content is the standard cause of action name and the corresponding confidence level. If all candidate scores do not meet the standard, no forced matching will be performed. Only the catch-all classification of the relevant field will be output to ensure the correctness of the routing.
[0040] (4) Safety net strategy: When all Top-k candidates are below the threshold, the cause of action is left blank; the output domain (such as labor and personnel, contracts and transactions, private lending and guarantees, personal injury and torts, etc.) is used to avoid misleading mandatory decisions.
[0041] (5) Performance and SLA: End-to-end target time ≤ 2 seconds: vector retrieval ≤ 80ms, rearrangement ≤ 120ms, RAG context assembly ≤ 50ms, controlled inference ≤ 800ms; margin is used for networking and serialization. Engineering techniques such as short templates, context deduplication, vector preheating, hotspot caching, asynchronous logging, and parallelization are adopted.
[0042] (6) Safety and Compliance: Full traceability (versioned knowledge base and prompts, output logs and feature retention), privacy minimization (only necessary fields are retained, and sensitive information is anonymized), and compliance statement.
[0043] 3. Extraction of target amount and complexity Target Amount: Guide users to fill in the amount involved in the case; combine with text automatic extraction and standardization (digitization, currency unification, range estimation) to form structured features to support cost calculation and magnitude matching.
[0044] Case complexity: Based on the number of factual elements, type of legal relationship, sufficiency of evidence, and degree of overlap of multiple causes of action in the user's information, a preliminary judgment of three levels (high / medium / low) is given, and the system is continuously optimized by combining online feedback and case handling results.
[0045] Concurrent processing: Case identification, target amount extraction, and complexity assessment are executed in parallel and uniformly aggregated into structured output at the distribution layer to ensure end-to-end latency targets.
[0046] C. Lawyer Profile Construction 1.Static portrait Collect and structure lawyer information: professional sub-fields (such as "Intellectual Property - Patent Infringement"), years of practice, educational background (PhD / Master / Bachelor), law firm size (large / medium / small), and historical case database (stored by case type, outcome, amount in dispute, etc.).
[0047] 2. Dynamic portrait Real-time status: Current number of cases being handled, available capacity, average response time over the past 30 days, and recent case-handling activity in the target area.
[0048] Data sources: a combination of reports from lawyers and data scraping from the system; abnormal data detection and manual correction mechanisms.
[0049] 3. Storage and Governance It integrates static and dynamic data into a distributed database, supporting millisecond-level queries and updates; it performs entity disambiguation, deduplication, null value completion, caliber alignment, and version management to ensure data consistency and traceability.
[0050] D. Recommendation Algorithm System 1. Definition of Indicators and Standards Centered on candidate cause of action c, the system defines N_c (the number of cases handled by each lawyer for that cause of action c), R_c (win rate [0,1], provided by an external system), A1 / A2 / A3 (workload across all types over the past three years, with higher weight for more recent cases), Y (years of practice), M (cumulative amount of damages handled), E_score (quantification of educational level), and S_score (quantification of law firm size). User damages U_amt and the average damages handled by lawyers for cause of action c L_amt_c serve as references for matching magnitudes.
[0051] 2. Unified Normalization and Basic Total Score All sub-items are standardized to dimensions 0-100; instead of using candidate pool extreme values, long-term statistics from the national full-volume database (such as P95 / P99 or business upper limit) are used for monthly calibration to ensure stability across batches. Base total score (before adding dynamic items): BaseScore = W_prof × P + W_act × A + W_exp × E, default weights 0.45 / 0.30 / 0.25 (configurable).
[0052] 3. Itemized scoring Professional competence P (related to the case): The quantity N_c is logarithmically compressed and normalized, and R_c is linearly normalized. The combined weight is 0.6 / 0.4, reflecting "experience + quality in this case". Active score A (workload of all types): Construct W = 2×A1+3×A2+5×A3, log normalized to 0-100; Experience background E: Y (capped at 30 years), M (log normalized), E_score / S_score linearly scaled, in combinations of 0.35 / 0.25 / 0.20 / 0.20.
[0053] 4. Dynamic weighting and monetary value matching for bonus points The complexity is weighted in three tiers: low complexity (W_prof = 0.45 / W_act = 0.35 / W_exp = 0.20), medium complexity (W_prof = 0.45 / W_act = 0.30 / W_exp = 0.25), and high complexity (W_prof = 0.45 / W_act = 0.20 / W_exp = 0.35). This means that the weights of the lawyer's professional ability, activity score, and experience background in the total score are dynamically adjusted according to the complexity of the user's case. For complex cases, senior lawyers score higher, making the ranking self-adaptive to complexity. Amount matching degree: The target amount is divided into 5 levels; 1-5 points are added for the same level as the user, 1 point is subtracted from the adjacent level to no less than 1 point, and no points are added for levels further away; the levels are discrete and the upper limit is controlled to avoid excessive disturbance.
[0054] If the output domain of the large language model does not have a cause of action, the amount and the number of cases for a specific cause of action are not calculated. Instead, the total number of cases handled by the lawyer in that domain is directly calculated. Then, a matching lawyer is recommended to the user based on the total number of cases handled by the lawyer in that domain.
[0055] 5. Cold start and missing values For key features, robust defaults and prior shrinkage are adopted, and hard zeroing is avoided to prevent systematic underestimation of new lawyers; for example, if M is missing, the median of the entire database is used, and if E_score / S_score is missing, the median grade is used, etc.
[0056] 6. Sorting, Explanation, and Ranking Calculate P / A / E, replace the weights according to complexity, add points based on the amount, truncate to 100, sort, and output Top 10-15.
[0057] Show the three main categories, complexity weight, and amount bonus, and provide 2-3 explanatory reasons (such as "active in the target case with a high success rate in the past year", "matching the target amount", "faster response and currently available for handling").
[0058] E. System Deployment and Interfaces Architecture: Microservice deployment; decoupling of RAG retrieval service, inference service, profiling service, ranking service, and monitoring service; unified authentication and auditing; caching and message queues to support high concurrency; Access and Output: Requests include case details (text / speech transcription), location, and preferences; responses include the case cause and confidence level, catch-all area, structured features (amount in dispute, complexity), a top list, and explanatory fields. Output fields are all structured key-value pairs; no code or script is returned.
[0059] The following are alternative solutions to the above embodiments.
[0060] On the retrieval side: it can be replaced with a fusion retrieval system based on BM25+ dense retrieval; different vector indexes can be used (such as FaissIVF-PQ / ScaNN); the rearranger can be replaced with a lightweight model based on interactive representation.
[0061] On the inference side: a restricted multi-label classifier (trained directly with case label) can be used to replace / supplement the controlled RAG, or a joint discriminative and generative voting method can be used; the confidence level can be replaced with temperature scaling / Platt calibration.
[0062] Knowledge base: A knowledge graph structure (cause of action - elements - evidence - legal provisions) can be introduced to enhance discrimination; the alias dictionary can be updated using automatic terminology mining and proactive learning.
[0063] Feature extraction: The target amount can be categorized using a logarithmic scale with adaptive distribution; the complexity assessment can introduce a graph structure (relationship between the parties involved) indicator.
[0064] Ranking and scoring: The complexity weight matrix can be extended to a continuous function or a learning-gated network, while still maintaining interpretable output; the amount matching can be replaced with a penalty for the difference in KL distance or Wasserstein distance, which is equivalent to achieving "adjacent priority".
[0065] "Location / Court Level / Trial Level Matching Degree" could be introduced as a bonus factor, or "User Preferences (Gender / Communication Style)" could be included with a lighter weight.
[0066] Profiles and Data: Dynamic load can be replaced by manual reporting through API integration with the law firm's scheduling / internal management system; response time can be replaced by self-reporting through message platform receipt statistics.
[0067] Data storage can be replaced with a columnar OLAP+KV hybrid architecture; the online vector library can be replaced with a self-developed or commercial engine.
[0068] Monitoring and SLA: Stability indicators can be replaced or supplemented with Kendall / Tau coefficient and NDCG@K change rate; drift detection can be replaced with distribution divergence (JS divergence / PSI).
[0069] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.
Claims
1. A method for dynamic adaptation and personalized recommendation of lawyer resources, comprising the following steps: The preprocessing module desensitizes the input user requirement information to obtain user requirement text, and then extracts keywords from the user requirement text and standardizes them; The case identification module statistically analyzes the case causes and their frequency of occurrence in each selected data source, and determines the high-frequency case causes based on the frequency of occurrence; it constructs a refined knowledge base based on the selected high-frequency case causes; it generates a standardized definition text segment for each high-frequency case cause and performs vectorization processing to obtain a high-frequency case cause vector knowledge base; Based on the standardized keywords obtained from the preprocessing module, the high-frequency case vector knowledge base is retrieved to obtain the standardized definition text segments corresponding to the K most relevant high-frequency cases to user needs; The user demand information and the standardized definition text segment are input into the large language model for classification and reasoning, and at most n candidate reasons with the highest confidence and all greater than a set threshold are output along with their confidence scores; if the confidence scores of the candidate reasons are all lower than the set threshold, the domain to which the user demand information belongs is output; the output of the large language model is sent to the recommendation module. Based on the amount of the target input by the user and the standardized keywords obtained by the preprocessing module, a structured feature is formed and sent to the recommendation module; based on the user's demand information, the case complexity is determined and sent to the recommendation module. The lawyer profiling module generates a profile of each lawyer based on their static and dynamic data. The recommendation module recommends matching lawyers to users based on each lawyer's profile and the information received.
2. The method according to claim 1, characterized in that, The standardized definition text segment includes an explanation segment, an alternative name dictionary, an applicable situation segment, an inapplicable situation segment, and an element and evidence segment.
3. The method according to claim 1, characterized in that, The user demand information includes case description, region, and preferences.
4. The method according to claim 1, 2, or 3, characterized in that, The static data includes the lawyer's professional sub-field, years of practice, educational background, law firm size, and historical case database; the dynamic data includes the lawyer's current number of cases handled, capacity to handle cases, average response time in the past 30 days, and recent case handling activity in the target field.
5. The method according to claim 1, 2, or 3, characterized in that, The method for classification reasoning using the large language model is as follows: combining the confidence of candidate causes of action with three types of evidence: definition / element matching, keyword / typical scene matching, and evidence word matching.
6. The method according to claim 1, characterized in that, The recommendation module recommends matching lawyers to users based on each lawyer's profile and received information as follows: Taking each candidate case type c as the core, it obtains each lawyer's case volume N_c for that candidate case type c, their case success rate R_c for that candidate case type c, their workload across all types of cases in the past N years, their years of practice Y, their cumulative case value M, their educational level E_score, and their law firm size S_score; it uses the user's case value U_amt and the lawyer's average case value L_amt_c for candidate case types c as a reference for matching magnitudes; then it calculates each lawyer's base score BaseScore = W_prof × P + W_act × A + W_exp × E; where W_prof, W_act, and W_exp are weighting coefficients determined based on case complexity, P is the lawyer's professional competence score calculated based on N_c and R_c, A is the lawyer's activity score calculated based on the workload of all types of cases in the past N years, and E is the lawyer's experience background score calculated based on Y, M, E_score, and S_score; the lawyer's base total score BaseScore is updated based on the matching degree between the user's claim amount U_amt and the lawyer's average claim amount L_amt_c in case c; then, the updated base total score BaseScore is used to recommend matching lawyers for candidate case c; if the output of the large language model is the domain to which the user's demand information belongs, then the lawyer is recommended to match the user based on the total number of cases handled by the lawyer in that domain.
7. A platform for dynamic matching and personalized recommendation of lawyer resources, characterized in that, It includes a preprocessing module, a case identification module, an evaluation module, a lawyer profiling module, and a recommendation module; The preprocessing module is used to desensitize the input user demand information to obtain user demand text, and then extract keywords from the user demand text and standardize them. The case identification module is used to statistically analyze the case causes and their frequency of occurrence in each selected data source, determine the high-frequency case causes based on the frequency of occurrence, construct a refined knowledge base based on the selected high-frequency case causes, generate a standardized definition text segment for each high-frequency case cause and perform vectorization processing to obtain a high-frequency case cause vector knowledge base; The evaluation module retrieves the high-frequency case vector knowledge base based on the normalized keywords obtained from the preprocessing module, and obtains the standardized definition text segments corresponding to the K most relevant high-frequency case causes to the user's needs. Then, it inputs the user's needs information and the standardized definition text segments into a large language model for classification reasoning, outputting at most n candidate case causes with the highest confidence levels, all exceeding a set threshold, along with their confidence levels. If the confidence levels of all candidate case causes are below the set threshold, it outputs the domain to which the user's needs information belongs, and then sends the output of the large language model to the recommendation module. Furthermore, it forms structured features based on the user's input amount and the normalized keywords obtained from the preprocessing module and sends these features to the recommendation module. Finally, it determines the case complexity based on the user's needs information and sends this complexity to the recommendation module. The lawyer profiling module generates a profile of each lawyer based on their static and dynamic data. The recommendation module recommends matching lawyers to users based on each lawyer's profile and the information received.
8. The platform according to claim 7, characterized in that, The recommendation module recommends matching lawyers to users based on each lawyer's profile and received information as follows: Taking each candidate case type c as the core, it obtains each lawyer's case volume N_c for that candidate case type c, their case success rate R_c for that candidate case type c, their workload across all types of cases in the past N years, their years of practice Y, their cumulative case value M, their educational level E_score, and their law firm size S_score; it uses the user's case value U_amt and the lawyer's average case value L_amt_c for candidate case types c as a reference for matching magnitudes; then it calculates each lawyer's base score BaseScore = W_prof × P + W_act × A + W_exp × E; where W_prof, W_act, and W_exp are weighting coefficients determined based on case complexity, P is the lawyer's professional competence score calculated based on N_c and R_c, A is the lawyer's activity score calculated based on the workload of all types of cases in the past N years, and E is the lawyer's experience background score calculated based on Y, M, E_score, and S_score; the lawyer's base total score BaseScore is updated based on the matching degree between the user's claim amount U_amt and the lawyer's average claim amount L_amt_c in case c; then, the updated base total score BaseScore is used to recommend matching lawyers for candidate case c; if the output of the large language model is the domain to which the user's demand information belongs, then the lawyer is recommended to match the user based on the total number of cases handled by the lawyer in that domain.
9. The platform according to claim 7, characterized in that, The standardized definition text segment includes an explanation segment, an alternative name dictionary, an applicable situation segment, an inapplicable situation segment, and an element and evidence segment.
10. The platform according to claim 7, characterized in that, The method for classification reasoning using the large language model is as follows: combining the confidence of candidate causes of action with three types of evidence: definition / element matching, keyword / typical scene matching, and evidence word matching.