Intelligent allocation and scheduling method, device and equipment for legal assistance resources and medium

Abstract

The invention relates to a legal assistance resource intelligent allocation scheduling method and device, equipment and a medium. The method comprises the following steps: acquiring consultation data; the consultation data comprises an original consultation text, a voice record and a video file; based on the consultation data, legal entities are extracted through a multi-task learning model, and three-dimensional classification labels are obtained; the three-dimensional classification label comprises a geographic position code, a case type code and an emergency degree code; matching the three-dimensional classification labels and the lawyer features through a deep forest model to obtain a matching probability matrix; monitoring the lawyer load based on the current lawyer state to obtain a resource state thermodynamic diagram; based on the matching probability matrix and the resource state thermodynamic diagram, obtaining a matching scheme set through a reinforcement learning model; the matching scheme set comprises legal assistance schemes. By adopting the method, the lawyer state can be fed back in real time, and the law assistance scheme recommendation flexibility is improved.

Description

Technical Field

[0001] The present invention belongs to the field of resource scheduling, and in particular relates to a method, device, equipment and medium for intelligent allocation and scheduling of legal aid resources. Background Art

[0002] With the increasing public demand for legal aid, traditional methods for allocating legal aid resources have emerged. These methods are typically based on structured databases and rule-based engines. These systems manually annotate lawyers' practice areas and case types with keywords and establish static mapping rules. When matching, they first filter lawyers in the same field using keyword matching, then filter their geographic locations based on administrative division codes. A fixed distance threshold is used to calculate the distance between the lawyer and the case, and finally, recommendations are ranked by the lawyer's historical win rate. However, these methods suffer from a lack of real-time feedback, long update cycles, and inflexible recommendation schemes. Summary of the Invention

[0003] Based on this, it is necessary to provide an intelligent allocation and scheduling method for legal aid resources that can provide real-time lawyer status and flexibly generate recommendation plans to address the above technical problems.

[0004] In a first aspect, the present application provides a method for intelligently allocating and scheduling legal aid resources, including:

[0005] Obtain consultation data; consultation data includes original consultation text, voice recordings and video files;

[0006] Based on the consulting data, a multi-task learning model is used to extract legal entities and obtain three-dimensional classification labels. The three-dimensional classification labels include geographic location codes, case type codes, and urgency codes.

[0007] Match the three-dimensional classification labels and lawyer features through the deep forest model to obtain the matching probability matrix;

[0008] Monitor lawyer load based on their current status and obtain a resource status heat map;

[0009] Based on the matching probability matrix and resource status heat map, a matching solution set is obtained through the reinforcement learning model; the matching solution set includes the legal aid solution.

[0010] Furthermore, based on the consulting data, a multi-task learning model is used to extract legal entities and obtain three-dimensional classification labels, including:

[0011] Generate contextual embedding vectors from consulting data through the BERT-base model;

[0012] Based on the legal knowledge graph, the context embedding vector is calculated using the graph neural network using the following formula to obtain a list of legal entities:

[0013] e ij =LeakyReLU(W g [h i ||h j ])

[0014]

[0015] P i =Softmax(W c h′ i +b c )

[0016] LegalEntityList=Node i |arg max(P i )∈ValidClasses,max(P i )≥τ

[0017] Among them, LegalEntityList is the legal entity list, ValidClasses is the predefined valid legal entity category set, τ is the confidence threshold, P i is the category probability distribution of node i, W c is the classification weight matrix, b c is the bias term of the classification layer, h i , h j is the initial feature vector of the i,jth legal entity, carrying the semantic information of the legal entity, W g is the weight matrix, LeakyReLU is the leaky rectified linear unit activation function, e ij is the attention coefficient of edge (i, j), β ij is the normalized attention weight, W v is the linear transformation matrix, N i is the neighbor set of node i, h′ i is the updated node feature vector;

[0018] Based on the context embedding vector, the probability distribution of case types is obtained;

[0019] Based on the probability distribution of case types and the list of legal entities, the case type code is obtained;

[0020] Input the context embedding vector into the BiLSTM-CRF model to obtain the geographic location encoding;

[0021] The consultation data is calculated through a hybrid model to generate urgency codes; the hybrid model includes Wav2Vec2.0 units and NLP units;

[0022] Integrate case type code, geographic location code and urgency code to obtain a three-dimensional classification label.

[0023] Furthermore, lawyer characteristics are obtained by:

[0024] Extract lawyers’ professional field collection, years of practice and historical winning rate based on lawyer practice database;

[0025] Based on the professional field set, obtain professional field vectors through the legal knowledge graph;

[0026] Based on years of practice and historical win rate, the seniority vector is calculated using the following formula:

[0027]

[0028] q=[Y norm ,W′]

[0029] Among them, q is the qualification vector, which represents the quantitative representation of the lawyer's qualifications, Y is the lawyer's years of practice, and Y norm is the normalized years of practice, W is the historical win rate, N is the total number of historical cases, and W′ is the adjusted win rate;

[0030] Based on the lawyer's current status, the lawyer's current latitude and longitude coordinates and service radius are extracted to obtain a spatial identifier;

[0031] The professional field vector, qualification vector and spatial identifier features are concatenated to obtain the lawyer features.

[0032] Furthermore, the three-dimensional classification labels and lawyer features are matched through the deep forest model to obtain a matching probability matrix, including:

[0033] Based on the geographic location code in the three-dimensional classification label and the spatial identifier in the lawyer feature, the spatial matching degree between the consulting case and the lawyer's location is obtained;

[0034] Based on the lawyer characteristics and case type codes, the professional similarity between consulting cases and lawyer fields is calculated using the following formula:

[0035]

[0036] Among them, S pro is the professional similarity; P l is the lawyer's professional field vector, t c is the case type code, γ is the professional weight adjustment coefficient, and SpecialtyWeight is the professional field weight;

[0037] Based on the seniority vector and urgency coding in the lawyer characteristics, the lawyer's timeliness baseline is calculated using the following formula:

[0038]

[0039] Among them, T base The smaller the baseline, the more efficient the lawyer is in handling cases. norm is the normalized years of practice, W′ is the adjusted win rate, and EmergencyLevel is the urgency code;

[0040] Based on the spatial matching degree, professional similarity and time-efficiency baseline, cascade forest training is performed using the following formula to obtain the matching probability matrix:

[0041]

[0042] in, is the element in the matching probability matrix, w space is the spatial matching degree, S pro is the professional similarity, T base is the time-based baseline, α, β, γ are weight coefficients, P init is the initial matching probability matrix, F (i,j) is the feature vector, CascadeForest is the cascade forest model, C is the cascade forest configuration parameter, and 1e-5 is the scientific notation, which means 1×10 -5 , minimum value, avoiding the denominator being zero.

[0043] Furthermore, based on the lawyers' current status, the lawyer load is monitored to obtain a resource status heat map, including:

[0044] Based on the current status of the lawyer, the real-time load is calculated using the following formula:

[0045]

[0046] Among them, LoadRate i is the real-time load, is the number of cases lawyer i is handling, The maximum number of cases handled by lawyer i;

[0047] Determine the lawyer's remaining time in transit based on the lawyer's current status;

[0048] Equivalent the remaining in-transit time to the case load to obtain the in-transit load;

[0049] Based on the historical case processing time series, the Prophet forecasting model is used to predict the future case volume and obtain the future load; the historical case processing time series is obtained from the lawyer practice database;

[0050] Based on the real-time load, in-transit load, and future load, the resource health is calculated using the following formula:

[0051]

[0052] Among them, Health i is the resource health, the closer to 1, the more idle the lawyers are, and negative values indicate overload; is the total predicted case volume in the future time period a, For real-time equivalent caseload, TravelLoad i is the equivalent caseload in transit, is the lawyer’s total processing capacity in time period a;

[0053] Generate a resource status heat map based on resource health and lawyers' current location.

[0054] Furthermore, based on the matching probability matrix and resource status heat map, a matching solution set is obtained through the reinforcement learning model, including:

[0055] The matching probability matrix and resource status heat map features are integrated to obtain the status matrix;

[0056] Based on the matching probability matrix and resource status heat map, the reward matrix is calculated using the following formula:

[0057]

[0058] Among them, R ij is the element in the reward matrix, P ij is the element in the matching probability matrix, is the predicted processing time for lawyer i to handle case j, α, β, θ are weight coefficients, and HealthLevel i The resource health level of lawyer i, HealthLevel i -5 means that a penalty is triggered when the monitoring degree exceeds the threshold of 5, preventing overloaded lawyers from being over-allocated;

[0059] Based on the state matrix and reward matrix, the Q value is updated using the following formula to obtain the optimized Q value matrix:

[0060]

[0061] Among them, Q new (s,a) is the updated Q value, s is the current state, a is the current action, indicating the match between lawyer i and case j, s′ is the next state, α is the learning rate, γ is the discount factor, and Q(s,a) is the current Q value;

[0062] The matching probability matrix is optimized based on the optimized Q value matrix to obtain the optimized matching probability matrix;

[0063] Based on the optimized matching probability matrix, a matching solution set is determined.

[0064] Furthermore, based on the matching probability matrix and resource status heat map, after obtaining the matching solution set through the reinforcement learning model, it also includes:

[0065] Receive case closure reports;

[0066] Compare the deviations between the case closing report and the matching solution set to generate a deviation set;

[0067] Based on the deviation set, compare the legal knowledge graph and generate the association rule set;

[0068] The weights are adjusted based on the association rule set to generate a model correction instruction set; the model correction instruction set is used to update the reinforcement learning model.

[0069] In a second aspect, the present application also provides a device for intelligently allocating and scheduling legal aid resources, including:

[0070] Consultation module, used to obtain consultation data; consultation data includes original consultation text, voice recordings and video files;

[0071] The encoding module is used to extract legal entities based on consulting data through a multi-task learning model and obtain three-dimensional classification labels; the three-dimensional classification labels include geographic location codes, case type codes, and urgency codes;

[0072] The matching module is used to match the three-dimensional classification labels and lawyer features through the deep forest model to obtain a matching probability matrix;

[0073] The status module is used to monitor lawyer load based on the lawyer's current status and obtain a resource status heat map;

[0074] The decision module is used to obtain a matching solution set based on the matching probability matrix and the resource status heat map through a reinforcement learning model; the matching solution set includes a legal aid solution.

[0075] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method provided in the first aspect of the present application when executing the computer program.

[0076] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect of the present application.

[0077] The above-mentioned intelligent allocation and scheduling method, device, equipment, and medium for legal aid resources acquire consultation data; consultation data includes original consultation text, voice recordings, and video files; based on the consultation data, legal entities are extracted through a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include geographic location codes, case type codes, and urgency codes; the three-dimensional classification labels are matched with lawyer characteristics through a deep forest model to obtain a matching probability matrix; lawyer load is monitored based on the lawyer's current status to obtain a resource status heat map; based on the matching probability matrix and resource status heat map, a reinforcement learning model is used to obtain a set of matching solutions; the matching solution set includes legal aid solutions. Through the above steps, it is possible to obtain lawyer status and flexibly provide legal aid solutions based on the case situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0079] Figure 1 This is a flow chart of the legal aid resource intelligent allocation and scheduling method of the present invention;

[0080] Figure 2 This is a diagram of the legal aid resource intelligent allocation and scheduling device of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0082] The legal aid resource intelligent allocation and scheduling method provided in the embodiment of the present application can be applied in the application environment of legal aid resource intelligent recommendation.

[0083] In one embodiment, Figure 1 As shown, a method for intelligent allocation and scheduling of legal aid resources is provided. This embodiment uses the method applied to a processor as an example for illustration. It is understandable that the method can also be applied to a server, or to a system including a processor and a server, and implemented through the interaction between the processor and the server. In this embodiment, the method includes the following steps:

[0084] Step 101, obtaining consultation data; the consultation data includes original consultation text, voice recording and video files.

[0085] Specifically, consultation data refers to the original information submitted by users during the legal aid process, including text-based conversation records, audio recordings, and video footage of body language and facial expressions. The legal services platform's front-end interface collects user-submitted consultation requests in real time, while also simultaneously receiving user-uploaded text files, call recordings, and video consultation recordings. This creates a complete initial data pool, providing multimodal data support for subsequent analysis. For example, anxious expressions in videos can help determine the level of urgency, and dialect voices can help locate geographic areas.

[0086] Step 102 : extract legal entities based on the consultation data through a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include geographic location codes, case type codes, and urgency codes.

[0087] Specifically, a multi-task learning model refers to a neural network architecture that shares underlying features but outputs multiple prediction targets. Three-dimensional classification labels are structured sets of labels. The model processes text, voice, and video in parallel. Geographic location is mapped using address entity recognition and a positioning system, case type is calculated based on legal text compatibility, and urgency is determined through a combination of semantic sentiment and biometrics. Unstructured consultation information is converted into standardized, machine-processable labels.

[0088] Step 103: Match the three-dimensional classification labels and lawyer features through a deep forest model to obtain a matching probability matrix.

[0089] Specifically, Deep Forest (gcForest) is a multi-level ensemble learning framework that achieves deep feature interaction through cascaded random forest combinations. Lawyer characteristics can include feature vectors such as years of practice, win rate, professional field matching, and historical response speed. Three-dimensional classification labels and lawyer characteristics are projected into an embedding space, and feature intersections are constructed through multi-granular scanning. Each decision tree learns the association pattern between lawyer characteristics and case labels, ultimately outputting a matching probability matrix.

[0090] Step 104: Monitor lawyer load based on the lawyer's current status to obtain a resource status heat map.

[0091] Specifically, lawyer workload is assessed using a load function constructed from indicators such as the number of cases currently handled, expected case completion time, and schedule density. A heat map is a two-dimensional spatiotemporal visualization matrix, with time slices on the horizontal axis and lawyer IDs on the vertical axis. Color depth indicates workload intensity. Real-time lawyer schedule data is obtained through the law firm's OA system interface, and workload forecasts are generated based on historical case processing times. Each lawyer's available case window within the next 72 hours is calculated, generating a dynamically updated resource distribution map.

[0092] Step 105 : Based on the matching probability matrix and the resource status heat map, a matching solution set is obtained through a reinforcement learning model; the matching solution set includes a legal aid solution.

[0093] Specifically, through the reinforcement learning method of Q-learning, based on the matching probability matrix and the resource status heat map, the two are fused through the reinforcement learning model to generate a state matrix, calculate the reward matrix, iteratively update the Q value matrix optimization strategy, output the optimized matching probability matrix, and finally generate a matching solution set.

[0094] The intelligent allocation and scheduling method for legal aid resources provided in this embodiment obtains consultation data; the consultation data includes original consultation text, voice recordings and video files; based on the consultation data, legal entities are extracted through a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include geographic location codes, case type codes and urgency codes; the three-dimensional classification labels and lawyer features are matched through a deep forest model to obtain a matching probability matrix; the lawyer load is monitored based on the lawyer's current status to obtain a resource status heat map; based on the matching probability matrix and the resource status heat map, a matching solution set is obtained through a reinforcement learning model; the matching solution set includes technical means of the legal aid solution, which can obtain the lawyer's current status in real time and flexibly provide a legal aid solution based on the case situation.

[0095] In one embodiment, a multi-task learning model is used to extract legal entities based on consulting data to obtain three-dimensional classification labels, including:

[0096] Step 201: Generate a context embedding vector for the consultation data through the BERT-base model.

[0097] Specifically, the BERT-base model is a pre-trained natural language processing model based on the Transformer architecture that generates semantic embeddings through bidirectional contextual understanding. Contextual embeddings are semantic embeddings of text that incorporate lexical context information. The pre-trained BERT-base model is used to encode consultation data, converting unstructured text into structured embeddings, providing a semantic foundation for subsequent legal entity extraction.

[0098] In step 202, based on the legal knowledge graph, the context embedding vector is calculated using a graph neural network using the following formula to obtain a list of legal entities:

[0099] e ij =LeakyReLU(W g [h i ||h j ])

[0100]

[0101] P i =Softmax(W c h′ i +b c )

[0102] LegalEntityList=Node i |arg max(P i )∈ValidClasses,max(P i )≥τ

[0103] Among them, LegalEntityList is the legal entity list, ValidClasses is the predefined valid legal entity category set, τ is the confidence threshold, P i is the category probability distribution of node i, W c is the classification weight matrix, b c is the bias term of the classification layer, h i , h j is the initial feature vector of the i,jth legal entity, carrying the semantic information of the legal entity, W g is the weight matrix, LeakyReLU is the leaky rectified linear unit activation function, e ij is the attention coefficient of edge (i, j), β ij is the normalized attention weight, W v is the linear transformation matrix, N i is the neighbor set of node i, h′ i is the updated node feature vector.

[0104] Specifically, a legal knowledge graph is a structured knowledge base containing legal entities and their relationships. A legal entity list is a collection of key legal elements extracted from legal consultation data. It can represent the specific legal provisions involved in a case, case type, legal subject, and other relevant legal concepts. A graph neural network is a deep learning model that processes graph-structured data and updates node features by aggregating information from neighboring nodes.

[0105] Step 203: Obtain case type probability distribution based on the context embedding vector.

[0106] Specifically, the case type probability distribution is the probability vector of the case belonging to a predefined type. The context embedding vector generated by BERT is input into the fully connected layer, and the probability distribution is output through the Softmax function.

[0107] Step 204: Obtain a case type code based on the case type probability distribution and the legal entity list.

[0108] Specifically, the case type code is a unique identifier of the case type. The category with the highest probability is selected from the case type probability distribution and mapped to a predefined coding rule.

[0109] Step 205: Input the context embedding vector into the BiLSTM-CRF model to obtain the geographic location code.

[0110] Specifically, BiLSTM-CRF is a bidirectional long short-term memory network (capturing contextual dependencies) + conditional random field, which is used for sequence labeling tasks. Geographic location codes are standardized administrative division codes.

[0111] Step 206: Calculate the consultation data using a hybrid model to generate an urgency code; the hybrid model includes a Wav2Vec 2.0 unit and an NLP unit.

[0112] Specifically, the Wav2Vec 2.0 unit is a speech pre-training model that extracts acoustic features from audio. The NLP unit is a text analysis model that extracts keywords from text. The urgency code is a hierarchical code that identifies the urgency of the case.

[0113] Step 207 : Integrate the case type code, geographic location code, and urgency code to obtain a three-dimensional classification label.

[0114] Specifically, they are spliced in a fixed format to generate structured labels to uniformly represent the core attributes of the case.

[0115] This embodiment converts unstructured consultation data into three-dimensional classification labels through multimodal data processing, knowledge graph enhancement, and hybrid model decision-making, providing accurate and structured input for subsequent lawyer resource matching, and ultimately achieving efficient, fair, and explainable legal aid allocation.

[0116] In one embodiment, the lawyer feature is obtained by:

[0117] Step 301: Extract the lawyer's professional field set, years of practice and historical winning rate based on the lawyer practice database.

[0118] Specifically, the attorney practice database is a structured database that stores basic information about attorneys, including their areas of expertise, years of practice, and historical case records. The area of expertise collection is a list of legal fields in which the attorney specializes. Years of practice is the time since the attorney obtained their license to practice. The historical win rate is the percentage of cases the attorney has handled in the past in which they have won.

[0119] Step 302: Based on the professional field set, obtain the professional field vector through the legal knowledge graph.

[0120] Specifically, the legal knowledge graph is a knowledge network consisting of legal domain nodes and their relationships. The professional domain vector is an embedded representation of a lawyer's professional domain within the graph. Graph embedding methods are used to generate vector representations of nodes.

[0121] Step 303: Calculate the seniority vector based on years of practice and historical win rate using the following formula:

[0122]

[0123] q=[Y norm ,W′]

[0124] Among them, q is the qualification vector, which represents the quantitative representation of the lawyer's qualifications, Y is the lawyer's years of practice, and Y norm is the normalized years of practice, W is the historical win rate, N is the total number of historical cases, and W′ is the adjusted win rate.

[0125] Specifically, the seniority vector is a two-dimensional vector that combines years of practice and win rate. Bayesian smoothing is performed on the win rate of small sample cases to avoid extreme values.

[0126] For example, we not only look at how many cases the lawyer has fought, but also consider the change in the winning rate in the last three years. W′ can increase the weight of the winning rate in the past three years.

[0127] Step 304: Based on the lawyer's current status, the lawyer's current latitude and longitude coordinates and service radius are extracted to obtain a spatial identifier.

[0128] Specifically, the current latitude and longitude coordinates represent the lawyer's real-time geographic location. The service radius is the geographic range within which the lawyer can accept cases. The spatial identifier is a code or vector representing the lawyer's geographic service capabilities. The location is encoded into a string using geohashing.

[0129] Step 305: Concatenate the professional field vector, the qualification vector, and the spatial identifier feature to obtain the lawyer feature.

[0130] Specifically, lawyer characteristics are comprehensive vectors that integrate professional fields, qualifications, and spatial information.

[0131] This embodiment solves the multi-objective optimization problem in legal aid matching by quantifying the core attributes of lawyers, and is objective, scalable and adaptable.

[0132] In one embodiment, the three-dimensional classification labels and lawyer features are matched through a deep forest model to obtain a matching probability matrix, including:

[0133] Step 401: Based on the geographic location code in the three-dimensional classification label and the spatial identifier in the lawyer feature, the spatial matching degree between the consulting case and the lawyer's location is obtained.

[0134] Specifically, the geolocation code represents the standardized code for the location where the case occurred. The spatial identifier represents the latitude and longitude coordinates and service radius of the lawyer. The spatial matching degree represents the degree of match between the case and the lawyer's geographic location.

[0135] For example, Beijing lawyers will receive discounts when handling cases in Shanghai (taking into account the legal differences between the two places).

[0136] Step 402: Based on the lawyer characteristics and case type codes, the professional similarity between the consulting case and the lawyer's field is calculated using the following formula:

[0137]

[0138] Among them, S pro is the professional similarity; P l is the lawyer's professional field vector, t c is the case type code, γ is the professional weight adjustment coefficient, and SpecialtyWeight is the professional field weight.

[0139] Specifically, professional similarity is an indicator that quantifies the degree of match between a lawyer's professional ability and the legal needs of a case. The larger the value, the closer the match between the lawyer's professional ability and the legal needs of the case.

[0140] Step 403: Calculate the lawyer's timeliness baseline based on the seniority vector and urgency code in the lawyer's characteristics using the following formula:

[0141]

[0142] Among them, T base The smaller the baseline, the more efficient the lawyer is in handling cases. norm is the normalized years of practice, W′ is the adjusted win rate, and EmergencyLevel is the urgency code.

[0143] Specifically, the longer the years of practice, the higher the efficiency and the higher the winning rate, and the efficiency is positively enhanced. The degree of urgency includes three levels.

[0144] In step 404, based on the spatial matching degree, professional similarity, and time-based baseline, cascade forest training is performed using the following formula to obtain a matching probability matrix:

[0145]

[0146]

[0147] in, is the element in the matching probability matrix, w space is the spatial matching degree, S prois the professional similarity, T base is the time-based baseline, α, β, γ are weight coefficients, P init is the initial matching probability matrix, F (i,j) is the feature vector, CascadeForest is the cascade forest model, C is the cascade forest configuration parameter, and 1e-5 is the scientific notation, which means 1×10 -5 , minimum value, avoiding the denominator being zero.

[0148] Specifically, Cascade Forest is an ensemble model consisting of multiple layers of cascaded random forests, where each layer enhances features using the predictions of the previous layer. Each element in the matching probability matrix represents the probability of matching the i-th case with the j-th lawyer.

[0149] This embodiment transforms the lawyer-case matching process into a quantifiable and explainable optimization problem through the fusion of spatial, professional, and time-sensitive three-dimensional features and cascade forest modeling, adapting to the diverse needs of legal services.

[0150] In one embodiment, lawyer load is monitored based on the lawyer's current status to obtain a resource status heat map, including:

[0151] Step 501: Based on the current status of the lawyer, the real-time load is calculated using the following formula:

[0152]

[0153] Among them, LoadRate i is the real-time load, is the number of cases lawyer i is handling, is the maximum number of cases handled by lawyer i.

[0154] Specifically, the real-time load is the ratio of the lawyer's current workload to the maximum processing capacity, which quantifies the lawyer's current work saturation. The closer the value is to 1, the closer it is to full load.

[0155] Step 502: Determine the remaining time the lawyer is in transit based on the lawyer's current status.

[0156] Specifically, the lawyer's remaining travel time is the time the lawyer spends on transportation when he goes out to investigate the case.

[0157] Step 503: Equivalent the remaining in-transit time to the case load to obtain the in-transit load.

[0158] For example, assuming that it takes an average of 2 hours for a lawyer to handle a single case, the remaining 8 hours of transit time is equivalent to a load of 4 cases.

[0159] Step 504 : Based on the historical case processing time series, the future case volume is predicted using the Prophet prediction model to obtain the future load; the historical case processing time series is obtained from the lawyer practice database.

[0160] Specifically, the future load is the predicted number of cases that lawyer i needs to handle in the future time period. The Prophet forecasting model is a time series forecasting model.

[0161] Step 505: Calculate resource health based on real-time load, in-transit load, and future load using the following formula:

[0162]

[0163] Among them, Health i is the resource health, the closer to 1, the more idle the lawyers are, and negative values indicate overload; is the total predicted case volume in the future time period a, For real-time equivalent caseload, TravelLoad i is the equivalent caseload in transit, is the total processing capacity of the lawyer in time period a.

[0164] Specifically, resource health is the comprehensive load status of lawyer i in the future time period a. The lawyer's health status is globally assessed by integrating real-time, in-transit, and future loads.

[0165] Step 506: Generate a resource status heat map based on the resource health and the lawyer's current location.

[0166] Specifically, the resource status heat map is a color-coded visualization of lawyer health using geographic location as coordinates. It intuitively displays the distribution and load status of regional lawyer resources, guiding task allocation.

[0167] This embodiment converts lawyer status into resource health through real-time monitoring, future prediction, and multi-dimensional load fusion, and generates a geographic heat map, which can achieve dynamic early warning and global optimization effects.

[0168] In one embodiment, based on the matching probability matrix and the resource status heat map, a matching solution set is obtained through a reinforcement learning model, including:

[0169] Step 601: The matching probability matrix and the resource status heat map features are integrated to obtain a status matrix.

[0170] Specifically, the state matrix is a feature matrix that integrates the matching probability matrix and the resource state heat map.

[0171] Step 602: Based on the matching probability matrix and the resource status heat map, the reward matrix is calculated using the following formula:

[0172]

[0173] Among them, R ij is the element in the reward matrix, P ij is the element in the matching probability matrix, is the predicted processing time for lawyer i to handle case j, α, β, θ are weight coefficients, and HealthLevel i The resource health level of lawyer i, HealthLevel i -5 means that a penalty is triggered when the monitoring degree exceeds the threshold of 5 to prevent overloaded lawyers from being over-allocated.

[0174] Specifically, the reward matrix represents the immediate payoff of assigning case i to lawyer j, guiding the optimization of reinforcement learning strategies. The predicted processing time represents the estimated time lawyer i will take to handle case j. This balances matching quality, processing efficiency, and resource health to prevent overload allocation.

[0175] Step 603: Based on the state matrix and the reward matrix, the Q value is updated using the following formula to obtain an optimized Q value matrix:

[0176]

[0177] Among them, Q new (s,a) is the updated Q value, s is the current state, a is the current action, indicating the match between lawyer i and case j, s′ is the next state, α is the learning rate, γ is the discount factor, and Q(s,a) is the current Q value.

[0178] Specifically, the Q-value matrix represents the long-term expected reward of performing action a (assigning case i to lawyer j) in state s. The learning rate controls the Q-value update step size. The discount factor balances current rewards with future gains.

[0179] Step 604: Optimize the matching probability matrix based on the optimized Q value matrix to obtain an optimized matching probability matrix.

[0180] Specifically, the optimized matching probability matrix is the case-lawyer matching probability adjusted according to the Q-value matrix. For each case i, the Q-value of lawyer j is normalized to a probability distribution, and the Softmax function is used to convert the Q-value into a probability.

[0181] Step 605: Determine a matching solution set based on the optimized matching probability matrix.

[0182] Specifically, the matching solution set is the final list of case and lawyer assignments, representing the detailed assignment strategy for a specific case. For each case i, lawyer j with the highest probability can be selected. Alternatively, random sampling can be performed based on a probability distribution.

[0183] This embodiment combines static matching probability with dynamic resource status through reinforcement learning to generate a set of sustainably optimized matching solutions, thereby achieving a decision-making method for adaptive allocation and maximizing long-term benefits.

[0184] In one embodiment, after obtaining a matching solution set based on a matching probability matrix and a resource status heat map through a reinforcement learning model, the following steps are further included:

[0185] Step 701: Receive a case closing report.

[0186] Specifically, the case closing report is structured data generated after the case is handled, which may include actual processing time, winning results, changes in lawyer workload, client feedback, etc.

[0187] Step 702 : Compare the case closing report with the matching solution set to determine the deviations and generate a deviation set.

[0188] Specifically, a deviation set is a dataset that records the differences between the matching solution and the actual results. Each deviation can include the case ID, deviation type, and deviation magnitude. Deviations can include time deviation, result deviation, and load deviation.

[0189] Step 703: Based on the deviation set, compare the legal knowledge graph and generate an association rule set.

[0190] Specifically, association rules are logical rules that describe the causal relationship between deviation cases and legal knowledge entities. This involves extracting subgraphs of legal entities associated with deviation cases, analyzing frequent itemsets, verifying whether expertise mismatches lead to inefficiencies, and formalizing the rules to obtain association rules.

[0191] Step 704: Adjust the weights based on the association rule set and generate a model modification instruction set; the model modification instruction set is used to update the reinforcement learning model.

[0192] Specifically, the model modification instruction set is an operation instruction used to adjust the parameters or structure of the reinforcement learning model, modifying the coefficients in the reward function to balance multiple objectives.

[0193] This embodiment converts actual case results into model optimization instructions through a data-driven feedback mechanism and rule mining guided by domain knowledge (legal map), achieving precise parameter adjustment, continuous learning, and reducing human intervention.

[0194] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0195] Based on the same inventive concept, embodiments of the present application also provide a device for intelligently allocating and scheduling legal aid resources for implementing the aforementioned method for intelligently allocating and scheduling legal aid resources. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations in one or more embodiments of the device for intelligently allocating and scheduling legal aid resources provided below can be found in the aforementioned limitations of the method for intelligently allocating and scheduling legal aid resources, and will not be further elaborated here.

[0196] In an exemplary embodiment, Figure 2 As shown, a legal aid resource intelligent allocation and scheduling device 800 is provided, comprising:

[0197] Consultation module 801, used to obtain consultation data; consultation data includes original consultation text, voice recording and video files;

[0198] The encoding module 802 is used to extract legal entities based on the consultation data through a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include geographic location codes, case type codes, and urgency codes;

[0199] Matching module 803, used to match the three-dimensional classification labels and lawyer features through a deep forest model to obtain a matching probability matrix;

[0200] Status module 804, for monitoring lawyer load based on the lawyer's current status and obtaining a resource status heat map;

[0201] The decision module 805 is used to obtain a matching solution set based on the matching probability matrix and the resource status heat map through a reinforcement learning model; the matching solution set includes a legal aid solution.

[0202] Furthermore, the encoding module 802 is further configured to:

[0203] Generate contextual embedding vectors from consulting data through the BERT-base model;

[0204] Based on the legal knowledge graph, the context embedding vector is calculated using the graph neural network using the following formula to obtain a list of legal entities:

[0205] e ij =LeakyReLU(W g [h i ||h j ])

[0206]

[0207] P i =Softmax(W c h′ i +b c )

[0208] LegalEntityList=Node i |arg max(P i )∈ValidClasses,max(P i )≥τ

[0209] Among them, LegalEntityList is the legal entity list, ValidClasses is the predefined valid legal entity category set, τ is the confidence threshold, P i is the category probability distribution of node i, W c is the classification weight matrix, b c is the bias term of the classification layer, h i , h j is the initial feature vector of the i,jth legal entity, carrying the semantic information of the legal entity, W g is the weight matrix, LeakyReLU is the leaky rectified linear unit activation function, e ij is the attention coefficient of edge (i, j), β ij is the normalized attention weight, W v is the linear transformation matrix, N i is the neighbor set of node i, h′ i is the updated node feature vector;

[0210] Based on the context embedding vector, the probability distribution of case types is obtained;

[0211] Based on the probability distribution of case types and the list of legal entities, the case type code is obtained;

[0212] Input the context embedding vector into the BiLSTM-CRF model to obtain the geographic location encoding;

[0213] The consultation data is calculated through a hybrid model to generate urgency codes; the hybrid model includes Wav2Vec2.0 units and NLP units;

[0214] Integrate case type code, geographic location code and urgency code to obtain a three-dimensional classification label.

[0215] Furthermore, lawyer characteristics are obtained by:

[0216] Extract lawyers’ professional field collection, years of practice and historical winning rate based on lawyer practice database;

[0217] Based on the professional field set, obtain professional field vectors through the legal knowledge graph;

[0218] Based on years of practice and historical win rate, the seniority vector is calculated using the following formula:

[0219]

[0220] q=[Y norm ,W′]

[0221] Among them, q is the qualification vector, which represents the quantitative representation of the lawyer's qualifications, Y is the lawyer's years of practice, and Y norm is the normalized years of practice, W is the historical win rate, N is the total number of historical cases, and W′ is the adjusted win rate;

[0222] Based on the lawyer's current status, the lawyer's current latitude and longitude coordinates and service radius are extracted to obtain a spatial identifier;

[0223] The professional field vector, qualification vector and spatial identifier features are concatenated to obtain the lawyer features.

[0224] Furthermore, the matching module 803 is further configured to:

[0225] Based on the geographic location code in the three-dimensional classification label and the spatial identifier in the lawyer feature, the spatial matching degree between the consulting case and the lawyer's location is obtained;

[0226] Based on the lawyer characteristics and case type codes, the professional similarity between consulting cases and lawyer fields is calculated using the following formula:

[0227]

[0228] Among them, S pro is the professional similarity; P l is the lawyer's professional field vector, t c is the case type code, γ is the professional weight adjustment coefficient, and SpecialtyWeight is the professional field weight;

[0229] Based on the seniority vector and urgency coding in the lawyer characteristics, the lawyer's timeliness baseline is calculated using the following formula:

[0230]

[0231] Among them, T base The smaller the baseline, the more efficient the lawyer is in handling cases. norm is the normalized years of practice, W′ is the adjusted win rate, and EmergencyLevel is the urgency code;

[0232] Based on the spatial matching degree, professional similarity and time-efficiency baseline, cascade forest training is performed using the following formula to obtain the matching probability matrix:

[0233]

[0234] in, is the element in the matching probability matrix, w space is the spatial matching degree, S pro is the professional similarity, T base is the time-based baseline, α, β, γ are weight coefficients, P init is the initial matching probability matrix, F (i,j) is the feature vector, CascadeForest is the cascade forest model, C is the cascade forest configuration parameter, and 1e-5 is the scientific notation, which means 1×10 -5 , minimum value, avoiding the denominator being zero.

[0235] Furthermore, the status module 804 is further configured to:

[0236] Based on the current status of the lawyer, the real-time load is calculated using the following formula:

[0237]

[0238] Among them, LoadRate i is the real-time load, is the number of cases lawyer i is handling, The maximum number of cases handled by lawyer i;

[0239] Determine the lawyer's remaining time in transit based on the lawyer's current status;

[0240] Equivalent the remaining in-transit time to the case load to obtain the in-transit load;

[0241] Based on the historical case processing time series, the Prophet forecasting model is used to predict the future case volume and obtain the future load; the historical case processing time series is obtained from the lawyer practice database;

[0242] Based on the real-time load, in-transit load, and future load, the resource health is calculated using the following formula:

[0243]

[0244] Among them, Health i is the resource health, the closer to 1, the more idle the lawyers are, and negative values indicate overload; is the total predicted case volume in the future time period a, For real-time equivalent caseload, TravelLoad i is the equivalent caseload in transit, is the lawyer’s total processing capacity in time period a;

[0245] Generate a resource status heat map based on resource health and lawyers' current location.

[0246] Furthermore, the decision module 805 is further configured to:

[0247] The matching probability matrix and resource status heat map features are integrated to obtain the status matrix;

[0248] Based on the matching probability matrix and resource status heat map, the reward matrix is calculated using the following formula:

[0249]

[0250] Among them, R ij is the element in the reward matrix, P ij is the element in the matching probability matrix, is the predicted processing time for lawyer i to handle case j, α, β, θ are weight coefficients, and HealthLevel i The resource health level of lawyer i, HealthLevel i -5 means that a penalty is triggered when the monitoring degree exceeds the threshold of 5, preventing overloaded lawyers from being over-allocated;

[0251] Based on the state matrix and reward matrix, the Q value is updated using the following formula to obtain the optimized Q value matrix:

[0252]

[0253] Among them, Q new (s,a) is the updated Q value, s is the current state, a is the current action, indicating the match between lawyer i and case j, s′ is the next state, α is the learning rate, γ is the discount factor, and Q(s,a) is the current Q value;

[0254] The matching probability matrix is optimized based on the optimized Q value matrix to obtain the optimized matching probability matrix;

[0255] Based on the optimized matching probability matrix, a matching solution set is determined.

[0256] Furthermore, the device further includes an updating module, configured to:

[0257] Receive case closure reports;

[0258] Compare the deviations between the case closing report and the matching solution set to generate a deviation set;

[0259] Based on the deviation set, compare the legal knowledge graph and generate the association rule set;

[0260] The weights are adjusted based on the association rule set to generate a model correction instruction set; the model correction instruction set is used to update the reinforcement learning model.

[0261] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the aforementioned method for intelligent allocation and scheduling of legal aid resources when executing the computer program.

[0262] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0263] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0264] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A method for intelligent allocation and scheduling of legal aid resources, characterized in that: The method comprises: Obtain consultation data; the consultation data includes original consultation text, voice recordings and video files; Extracting legal entities based on the consultation data using a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include a geographic location code, a case type code, and an urgency code; Matching the three-dimensional classification labels and lawyer features through a deep forest model to obtain a matching probability matrix; Monitor lawyer load based on their current status and obtain a resource status heat map; Based on the matching probability matrix and the resource status heat map, a matching solution set is obtained through a reinforcement learning model; the matching solution set includes a legal aid solution.

2. The intelligent allocation and scheduling method for legal aid resources according to claim 1 is characterized in that: The extracting legal entities based on the consulting data through a multi-task learning model to obtain three-dimensional classification labels includes: Generate context embedding vectors for the consulting data through the BERT-base model; Based on the legal knowledge graph, the context embedding vector is calculated using the graph neural network using the following formula to obtain a list of legal entities: e ij =LeakyReLU(W g [h i ‖h j ]) P i =Softmax(W c h′ i +b c ) LegalEntityList=Node i |argmax(P i )∈ValidClasses,max(P i )≥τ Among them, LegalEntityList is the legal entity list, ValidClasses is the predefined valid legal entity category set, τ is the confidence threshold, P i is the category probability distribution of node i, W c is the classification weight matrix, b c is the bias term of the classification layer, h i , h j is the initial feature vector of the i,jth legal entity, carrying the semantic information of the legal entity, W g is the weight matrix, LeakyReLU is the leaky rectified linear unit activation function, e ij is the attention coefficient of edge (i, j), β ij is the normalized attention weight, W v is the linear transformation matrix, N i is the neighbor set of node i, h′ i is the updated node feature vector; Obtaining a case type probability distribution based on the context embedding vector; Obtaining a case type code based on the case type probability distribution and the legal entity list; Input the context embedding vector into the BiLSTM-CRF model to obtain the geographic location code; The consultation data is calculated using a hybrid model to generate an urgency code; the hybrid model includes a Wav2Vec2.0 unit and an NLP unit; The case type code, the geographic location code, and the urgency code are integrated to obtain a three-dimensional classification label.

3. The intelligent allocation and scheduling method for legal aid resources according to claim 1 is characterized in that: The lawyer characteristics are obtained by the following method: Extract lawyers’ professional field collection, years of practice and historical winning rate based on lawyer practice database; Based on the professional field set, obtaining professional field vectors through the legal knowledge graph; Based on the years of practice and the historical win rate, the seniority vector is calculated using the following formula: q=[Y norm ,W′] Among them, q is the qualification vector, which represents the quantitative representation of the lawyer's qualifications, Y is the lawyer's years of practice, and Y norm is the normalized years of practice, W is the historical win rate, N is the total number of historical cases, and W′ is the adjusted win rate; Based on the lawyer's current status, extract the lawyer's current latitude and longitude coordinates and service radius to obtain a spatial identifier; The professional field vector, the qualification vector and the spatial identifier feature are concatenated to obtain the lawyer feature.

4. The intelligent allocation and scheduling method for legal aid resources according to claim 1 is characterized in that: The three-dimensional classification labels and lawyer features are matched through a deep forest model to obtain a matching probability matrix, including: Based on the geographic location code in the three-dimensional classification label and the spatial identifier in the lawyer feature, obtaining a spatial matching degree between the consulting case and the lawyer's location; Based on the lawyer characteristics and the case type code, the professional similarity between the consulting case and the lawyer's field is calculated using the following formula: Among them, S pro is the professional similarity; P l is the lawyer's professional field vector, t c is the case type code, γ is the professional weight adjustment coefficient, and SpecialtyWeight is the professional field weight; Based on the seniority vector in the lawyer feature and the urgency code, the lawyer's timeliness baseline is calculated using the following formula: Among them, T base The smaller the time limit baseline is, the more efficient the lawyer is in handling cases. norm is the normalized years of practice, W′ is the adjusted win rate, and EmergencyLevel is the urgency code; Based on the spatial matching degree, the professional similarity and the timeliness baseline, cascade forest training is performed using the following formula to obtain the matching probability matrix: in, is the element in the matching probability matrix, w space is the spatial matching degree, S pro is the professional similarity, T base is the time-based baseline, α, β, γ are weight coefficients, P init is the initial matching probability matrix, F (i,j) is the feature vector, CascadeForest is the cascade forest model, C is the cascade forest configuration parameter, and 1e-5 is the scientific notation, which means 1×10 -5 , minimum value, avoiding the denominator being zero.

5. The intelligent allocation and scheduling method for legal aid resources according to claim 1 is characterized in that: The lawyer load is monitored based on the lawyer's current status to obtain a resource status heat map, including: Based on the current status of the lawyer, the real-time load is calculated by the following formula: Among them, LoadRate i is the real-time load, is the number of cases lawyer i is handling, The maximum number of cases handled by lawyer i; Determine the remaining time the lawyer is in transit based on the lawyer's current status; Equivalent the remaining in-transit time to the case load to obtain the in-transit load; Based on the historical case processing time series, the future case volume is predicted using the Prophet prediction model to obtain the future load; the historical case processing time series is obtained from the lawyer practice database; Based on the real-time load, the in-transit load, and the future load, the resource health is calculated using the following formula: Among them, Health i is the resource health, the closer to 1, the more idle the lawyers are, and negative values indicate overload; is the total predicted case volume in the future time period a, For real-time equivalent caseload, TravelLoad i is the equivalent caseload in transit, is the lawyer’s total processing capacity in time period a; Based on the resource health and the lawyer's current location, a resource status heat map is generated.

6. The method for intelligent allocation and scheduling of legal aid resources according to claim 1, characterized in that: The matching solution set is obtained based on the matching probability matrix and the resource status heat map through a reinforcement learning model, including: Fusing the matching probability matrix and the resource status heat map features to obtain a status matrix; Based on the matching probability matrix and the resource status heat map, the reward matrix is calculated using the following formula: Among them, R ij is the element in the reward matrix, P ij is the element in the matching probability matrix, is the predicted processing time for lawyer i to handle case j, α, β, θ are weight coefficients, and HealthLevel i The resource health level of lawyer i, HealthLevel i -5 means that a penalty is triggered when the monitoring degree exceeds the threshold of 5, preventing overloaded lawyers from being over-allocated; Based on the state matrix and the reward matrix, the Q value is updated using the following formula to obtain an optimized Q value matrix: Among them, Q new (s,a) is the updated Q value, s is the current state, a is the current action, indicating the match between lawyer i and case j, s′ is the next state, α is the learning rate, γ is the discount factor, and Q(s,a) is the current Q value; Optimizing the matching probability matrix based on the optimized Q value matrix to obtain an optimized matching probability matrix; Based on the optimized matching probability matrix, a matching solution set is determined.

7. The method for intelligent allocation and scheduling of legal aid resources according to any one of claims 1 to 6, characterized in that: After obtaining a matching solution set based on the matching probability matrix and the resource status heat map through a reinforcement learning model, the method further includes: Receive case closure reports; Comparing the case closing report with the matching solution set for deviations to generate a deviation set; Based on the deviation set, compare the legal knowledge graph to generate an association rule set; The weights are adjusted based on the association rule set to generate a model modification instruction set; the model modification instruction set is used to update the reinforcement learning model.

8. A device for intelligent allocation and scheduling of legal aid resources, characterized in that: The device comprises: The source module is used to obtain consultation data; the consultation data includes original consultation text, voice recordings and video files; A classification module is configured to extract legal entities based on the consultation data using a multi-task learning model to obtain three-dimensional classification labels; the three-dimensional classification labels include a geographic location code, a case type code, and an urgency code; A matching module, configured to match the three-dimensional classification labels and lawyer features through a deep forest model to obtain a matching probability matrix; The load module is used to monitor lawyer load based on the lawyer's current status and obtain a resource status heat map; A decision module is used to obtain a matching solution set based on the matching probability matrix and the resource status heat map through a reinforcement learning model; the matching solution set includes a legal aid solution.

9. 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 the method according to any one of claims 1 to 7 are implemented.

10. 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 the method according to any one of claims 1 to 7 are implemented.

Images