Intelligent mediation assistant construction method and system based on large model

Abstract

The invention discloses an intelligent mediation assistant construction method and system based on a large model. The system comprises an acquisition module, a dialogue processing module, a verbal skill recommendation module, a case reference module, a case library storage module, a learning optimization module and a deployment configuration module. The acquisition module is used for acquiring dialogue information; the dialogue processing module is used for reading dialogue data of the terminal and performing corresponding data preprocessing; the verbal skill recommendation module is used for receiving the metadata output by the conversation processing module and giving a conversation for recommending verbal skill in combination with the actual situation of the party; according to the method, the mediation ability of the excellent mediator is solidified, different scene verbal skills are automatically and continuously learned, mediation verbal skill recommendation is provided for the new mediator, the ability of directly talking with the party is provided, then the model is continuously adjusted and optimized according to the feedback of the recommendation effect, and the mediation efficiency of the new mediator is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent mediation assistants, and specifically to a method and system for constructing an intelligent mediation assistant based on a large model. Background Art

[0002] At present, with the iteration of artificial intelligence technology, dispute mediation business has also ushered in new development. In the traditional operation process, a set of standard templates and common scenario scripts will be prepared for mediators, which need to be memorized by the mediators themselves and flexibly applied in the actual mediation process to improve their mediation ability. In this way, novice mediators often need to undergo a long period of script training and actual combat before their mediation results can meet the standards of qualified mediators.

[0003] Therefore, on the road to becoming a qualified mediator, mediators are unable to give appropriate responses to various dialogue scenarios when communicating with the parties, which has a significant negative impact on the normal mediation process.

[0004] Current technology has the following disadvantages: 1. Real conversation scenarios are often more varied, and manually summarized speech scenarios cannot cover all real situations. They rely heavily on the mediator’s own mediation ability, and manually summarizing speech is time-consuming and labor-intensive; 2. The mediation skills of some excellent mediators did not help the rest of the mediators; 3. It takes a long time for new mediators to become qualified mediators, which cannot meet the needs of business expansion; 4. The work of the mediator itself is easily affected by his or her own emotions and changes in business scenarios. Even experienced salespeople may behave inappropriately at times. To this end, we propose a method and system for building an intelligent mediation assistant based on a large model. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for constructing an intelligent mediation assistant based on a large model to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A system for building an intelligent mediation assistant based on a large model, including a dialogue processing module, a speech recommendation module, a learning optimization module, and a deployment configuration module; The dialogue processing module is used to read the dialogue data of the terminal and perform corresponding data preprocessing; The speech recommendation module is used to receive the metadata output by the dialogue processing module, and give a dialogue with recommended speech in combination with the actual situation of the parties; The learning optimization module is used to train and construct the initial mediation assistant large model and to perform subsequent continuous optimization; The deployment configuration module is used for rapid deployment and fixed script configuration after model fine-tuning.

[0007] Preferably, the dialogue processing module includes a dialogue reading unit and a dialogue processing unit; The conversation reading unit is used to read the terminal conversation data, and can obtain the current conversation information of the parties and the historical conversation information between the parties and the mediator in combination with the stored records; The dialogue processing unit is used for preprocessing the read dialogue data.

[0008] Preferably, the speech recommendation module includes a scene recognition unit, an emotion recognition unit and a speech recommendation unit; The scene recognition unit is used to judge the current conversation scene based on historical conversation data, and output the judgment result to the speech recommendation unit. The scene recognition unit adopts the large model fine-tuning technology, fine-tunes the large model according to the past actual scene conversation and scene label data, and further controls the output of the large model in the prompt. The conversation scene is summarized based on past mediation, and all scenes are preset in advance by humans. The scene recognition unit needs to identify the conversation and judge which scene it is in. Specific scenes include: questioning identity, non-cooperation with mediation, complaint tendency, questioning the source of the number, dissatisfaction with the repayment plan, repayment difficulties, etc. The emotion recognition unit is used to analyze the emotional state of the parties according to the content of the conversation. The emotion recognition unit fine-tunes the large model according to the past actual scene conversations and emotion label data, and further controls the output of the large model in the prompt. The emotional state includes hope, gratitude, anxiety, fear, frustration, anger, etc. The speech recommendation unit is used to give the best speech recommendation by receiving the dialogue emotion, dialogue scene and specific dialogue content.

[0009] Preferably, the learning optimization module includes a learning unit and an automatic optimization unit; The learning unit is used to use the input excellent mediator data and the large model fine-tuning technology based on reinforcement learning to train the mediation assistant large model; The automatic optimization unit is used to self-adjust and optimize various aspects of the system.

[0010] Preferably, the learning unit includes a data processing subunit, a model training subunit and an effect testing subunit; The data processing subunit is responsible for extracting historical corpus data that meets the training conditions in the system, performing certain data cleaning on the data, dividing the cleaned data into training sets and test sets, and then processing the training set and test set data into a format that can be used for training and testing; The model training subunit encapsulates the large model fine-tuning training process based on reinforcement learning, and can be trained according to the input data to obtain the mediation assistant large model; The effect verification subunit is used to test the trained mediation assistant large model using the test set data to obtain the accuracy.

[0011] Preferably, the automatic optimization unit includes a feedback subunit and an automatic training unit; The feedback subunit is used to collect the mediator's feedback on the dialogue technique recommendation and store it in the underlying data table for subsequent tuning; The automatic training unit is used to filter out the recommended words in the feedback data that are not used by the mediator, and optimize the model according to the actual words of the mediator.

[0012] Preferably, the deployment configuration module includes a model deployment unit and a speech configuration unit; The model deployment unit is used to quickly deploy new models and manage historical versions of the mediation assistant large model; The speech configuration unit is used to manually configure standard speech for different emotions and scenarios. In particular, when the parties are emotionally agitated or prone to complaints, the model can answer with reference to preset standard speech.

[0013] A method for constructing an intelligent mediation assistant system based on a large model comprises the following steps: S1. The collection module collects relevant information from the historical conversation data between the terminal mediator and the parties, and cleans the extracted data to remove noise; S2, the dialogue processing module analyzes historical dialogue data, determines the current dialogue scene, outputs the scene classification result, and analyzes the emotional state of the parties according to the dialogue content; S3, input the scene classification results and emotional state, automatically select the most suitable case based on the relevance of the information in the case library to the current dispute, and cite the case to provide a basis for mediation speech, analyze the current conversation content in real time, and generate recommended speech; S4. Perform reinforcement learning training on the large model, test the trained large model, and continuously optimize the model; S5. Quickly deploy the optimized large model online and configure standard scripts; S6. Continuously collect conversation data during the mediation process and store it in the system's underlying database for subsequent optimization. Collect feedback from mediators on the use of dialogue technique recommendations and use the feedback data to automatically fine-tune the model.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention can realize the automatic summary of speech scenes by the model. Currently, the summary of speech scenes in the industry still relies on business experts to do it manually, which is time-consuming and labor-intensive, and is also prone to omissions. Using the model to automatically summarize can quickly realize the summary of the scene, and the probability of omissions is low.

[0015] The present invention can solidify the ability of excellent mediators into the model. The current traditional practice is for excellent mediators to manually summarize their experience and compile it into a book, which is labor-intensive. However, using a large model to solidify the ability can effectively learn the speech experience of excellent mediators and automatically and continuously update it.

[0016] The present invention can provide real-time speech recommendation for new mediators. By providing real-time speech recommendation, the success rate of mediation by mediators can be effectively increased, and at the same time, the occurrence of illegal speech can be avoided to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a module diagram of the present invention; Figure 2 This is a traditional dialogue processing module diagram; Figure 3 This is a module diagram for recommending speech techniques of the present invention; Figure 4 Optimizing module graphs for traditional learning; Figure 5 This is a deployment configuration module diagram of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The present invention provides a technical solution: a method for constructing an intelligent mediation assistant construction system based on a large model, comprising the following steps: S1. Extract relevant information from the historical conversation data between the terminal mediator and the parties, and clean the extracted data to remove noise; S2. Analyze historical conversation data, determine the current conversation scene, output the scene classification results, and analyze the emotional state of the parties based on the conversation content; S3: Input the scene classification results and emotional state, analyze the current conversation content in real time, and generate recommended words; S4. Perform reinforcement learning training on the large model, test the trained large model, and continuously optimize the model; S5. Quickly deploy the optimized large model online and configure standard scripts; S6. Continuously collect conversation data during the mediation process and store it in the system's underlying database for subsequent optimization. Collect feedback from mediators on the use of dialogue technique recommendations and use the feedback data to automatically fine-tune the model.

[0020] See also Figure 1 and Figure 2 , a large-model-based intelligent mediation assistant construction system, including a dialogue processing module, a speech recommendation module, a learning optimization module, and a deployment configuration module; The dialogue processing module is used to read the dialogue data of the terminal and perform corresponding data preprocessing. The dialogue processing module includes a dialogue reading unit and a dialogue processing unit. The dialogue reading unit is used to read the terminal dialogue data. The sources of dialogue data include mini-program instant chat messages and traditional telephone terminal recordings. Combined with the storage records, the current dialogue of the parties and their historical dialogue information with the mediator can be obtained; the dialogue processing unit is used to preprocess the read dialogue data.

[0021] After the conversation reading unit obtains the corresponding chat data, the conversation processing unit will perform corresponding preprocessing according to the conversation data type and obtain the corresponding meta information: (1) Text data refers to ordinary text chat data, in which idiomatic expressions and incorrect words will be corrected and filtered.

[0022] (2) For voice call data, convert it into text (using ASR technology) and further verify the translation results.

[0023] (3) For video data, the audio is also converted into text using the same method as in 2) above.

[0024] (4) Emotion recognition and processing: The dialogue processing unit is responsible for translating the voice format data and checking the translation accuracy to ensure the accuracy of the content input into the speech recommendation module. In addition to text translation processing, the emotion of the dialogue processing unit needs to be recognized to judge the emotional state of the parties and the mediator, so as to facilitate some subsequent processing. The emotion recognition unit can judge the emotions of the parties in the current dialogue based on the content of the dialogue and output the judgment results to the speech recommendation unit.

[0025] See also Figure 3The speech recommendation module is used to receive the metadata output by the dialogue processing module, namely the historical dialogue information and emotion labels, and give the dialogue of the recommended speech in combination with the actual situation of the parties. The speech recommendation module includes a scene recognition unit, an emotion recognition unit and a speech recommendation unit. The scene recognition unit is used to judge the current dialogue scene based on the historical dialogue data, and output the judgment result to the speech recommendation unit. The scene recognition unit adopts the large model fine-tuning technology, fine-tunes the large model according to the past actual scene dialogue and scene label data, and further controls the output of the large model in the prompt. The dialogue scene is summarized based on the past mediation, and all scenes are preset in advance by humans. The scene recognition unit needs to recognize the dialogue and judge which scene it is in. Specific scenarios include: questioning identity, refusal to cooperate with mediation, complaint tendency, questioning the source of the number, dissatisfaction with the repayment plan, repayment difficulties, etc.; the emotion recognition unit is used to analyze the emotional state of the parties based on the content of the conversation. The emotion recognition unit fine-tunes the large model based on past actual scene conversations and emotion label data, and further controls the output of the large model in the prompt. Emotional states include hope, gratitude, anxiety, fear, frustration, anger, etc.; the speech recommendation unit is used to give the best speech recommendation by receiving conversation emotions, conversation scenarios and specific conversation content.

[0026] See also Figure 4 The learning optimization module is used for the training and construction of the initial mediation assistant large model and the subsequent continuous optimization. The learning optimization module includes a learning unit and an automatic optimization unit.

[0027] The learning unit is used to use the input data of excellent mediators and the large model fine-tuning technology based on reinforcement learning to train the mediation assistant large model. The learning unit includes a data processing subunit, a model training subunit and an effect verification subunit. The data processing subunit is responsible for extracting the historical corpus data that meets the training conditions in the system, and performing certain data cleaning on it, dividing the cleaned data into training sets and test sets, and then processing the training set and test set data into a format that can be used for training and testing; the model training subunit encapsulates the large model fine-tuning training process based on reinforcement learning, and can be trained according to the input data to obtain the mediation assistant large model; the effect verification subunit is used to test the trained mediation assistant large model using the test set data to obtain the accuracy; The large model fine-tuning technology of reinforcement learning is to continuously optimize the model through the reinforcement learning algorithm to improve its performance in practical applications. Some key components of reinforcement learning include: state( ): represents the environment of the agent at a certain moment. In the intelligent mediation assistant, the state can include the content of the current conversation, the scene classification results, the emotions of the parties involved, etc. action( ): Indicates the decision made by the agent in a certain state. In this scenario, the action can be a recommended wording; award( ): An evaluation of the environment’s feedback after the agent takes an action. In a mediation assistant system, rewards can be defined based on the mediator’s feedback on the recommended speech (e.g., whether the mediator thinks the speech is effective or ineffective); Strategy (π ( / )): Defines the rules for how the agent chooses actions, generating the optimal speech based on historical data and current conversation content; Value function (V( )): Indicates that the state The expected total reward of the agent; Q value function (Q( , )): Indicates that the state Take action After that, the long-term returns obtained; The goal of reinforcement learning is to learn an optimal strategy so that the agent can obtain the maximum cumulative reward throughout the process. The goal of reinforcement learning is to maximize the expected total reward (i.e., discounted return). The Q-learning algorithm formula is: It can approximate the optimal strategy by continuously updating the Q value function. , where Q( , ) is in the state Take action Q value; is the immediate reward obtained at time t+1; is the discount factor, which controls the weight of future rewards; α is the learning rate, which controls the step size of each update; For the next state All possible actions The maximum Q value.

[0028] The automatic optimization unit is used to self-adjust and optimize various aspects of the system. The automatic optimization unit includes a feedback subunit and an automatic training unit. The feedback subunit is used to collect the mediator's feedback on the recommended dialogue and store it in the underlying data table for subsequent tuning; the automatic training unit is used to filter out the recommended dialogue that is not used by the mediator in the feedback data, and optimize the model according to the mediator's actual dialogue.

[0029] See also Figure 5The deployment configuration module is used for rapid deployment and fixed script configuration after fine-tuning the model. The deployment configuration module includes the model deployment unit and the script configuration unit. The model deployment unit is used to quickly deploy new models and manage historical versions of the mediation assistant model; the script configuration unit is used to manually configure standard scripts for different emotions and scenarios. In particular, when the parties are emotional or prone to complaints, the model can refer to the preset standard scripts for answers.

[0030] In summary: the present invention solidifies the mediation ability of excellent mediators, automatically and continuously learns the dialogue in different scenarios, provides mediation dialogue recommendations for new mediators, and provides the ability to directly communicate with the parties. Then, based on the feedback on the recommended effects, the model is continuously tuned to improve the mediation efficiency of new mediators.

[0031] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for constructing an intelligent mediation assistant based on a large model, characterized in that: It includes collection module, dialogue processing module, speech recommendation module, case reference module, case library storage module, learning optimization module and deployment configuration module; The collection module is used to collect conversation information; The dialogue processing module is used to read the dialogue data of the terminal and perform corresponding data preprocessing; The speech recommendation module is used to receive the metadata output by the dialogue processing module, and give a dialogue with recommended speech in combination with the actual situation of the parties; The case reference module is used to reference the latest case to generate corresponding mediation speech; The case library storage module is used to store the latest news, case libraries, mediation cases and court record cases; The learning optimization module is used to train and construct the initial mediation assistant large model and to perform subsequent continuous optimization; The deployment configuration module is used for rapid deployment and fixed script configuration after model fine-tuning.

2. According to claim 1, a large model-based intelligent mediation assistant construction system is characterized by: The dialogue processing module includes a dialogue reading unit and a dialogue processing unit; The conversation reading unit is used to read the terminal conversation data, and can obtain the current conversation information of the parties and the historical conversation information between the parties and the mediator in combination with the stored records; The dialogue processing unit is used for preprocessing the read dialogue data.

3. According to claim 2, a large model-based intelligent mediation assistant construction system is characterized by: The speech recommendation module includes a scene recognition unit, an emotion recognition unit and a speech recommendation unit; The scene recognition unit is used to determine the current conversation scene based on the historical conversation data, and output the determination result to the speech recommendation unit; The emotion recognition unit is used to analyze the emotional state of the parties according to the conversation content; The speech recommendation unit is used to give the best speech recommendation by receiving the dialogue emotion, dialogue scene and specific dialogue content.

4. According to claim 3, a large model-based intelligent mediation assistant construction system is characterized by: The learning optimization module includes a learning unit and an automatic optimization unit; The learning unit is used to use the input excellent mediator data and the large model fine-tuning technology based on reinforcement learning to train the mediation assistant large model; The automatic optimization unit is used to self-adjust and optimize various aspects of the system.

5. According to claim 4, a system for constructing an intelligent mediation assistant based on a large model is characterized in that: The learning unit includes a data processing subunit, a model training subunit and an effect testing subunit; The data processing subunit is responsible for extracting historical corpus data that meets the training conditions in the system, performing certain data cleaning on the data, dividing the cleaned data into training sets and test sets, and then processing the training set and test set data into a format that can be used for training and testing; The model training subunit encapsulates the large model fine-tuning training process based on reinforcement learning, and can be trained according to the input data to obtain the mediation assistant large model; The effect verification subunit is used to test the trained mediation assistant large model using the test set data to obtain the accuracy.

6. The system for constructing an intelligent mediation assistant based on a large model according to claim 5, characterized in that: The automatic optimization unit includes a feedback subunit and an automatic training unit; The feedback subunit is used to collect the mediator's feedback on the dialogue technique recommendation and store it in the underlying data table for subsequent tuning; The automatic training unit is used to filter out the recommended words in the feedback data that are not used by the mediator, and optimize the model according to the actual words of the mediator.

7. The system for constructing an intelligent mediation assistant based on a large model according to claim 6, characterized in that: The deployment configuration module includes a model deployment unit and a speech configuration unit; The model deployment unit is used to quickly deploy new models and manage historical versions of the mediation assistant large model; The speech configuration unit is used to manually configure standard speech for different emotions and scenarios. In particular, when the parties are emotionally agitated or prone to complaints, the model can answer with reference to preset standard speech.

8. The method for constructing a large model-based intelligent mediation assistant construction system according to claim 7, characterized in that: The steps include: S1. The collection module collects relevant information from the historical conversation data between the terminal mediator and the parties, and cleans the extracted data to remove noise; S2, the dialogue processing module analyzes historical dialogue data, determines the current dialogue scene, outputs the scene classification result, and analyzes the emotional state of the parties according to the dialogue content; S3, input the scene classification results and emotional state, automatically select the most suitable case based on the relevance of the information in the case library to the current dispute, and cite the case to provide a basis for mediation speech, analyze the current conversation content in real time, and generate recommended speech; S4. Perform reinforcement learning training on the large model, test the trained large model, and continuously optimize the model; S5. Quickly deploy the optimized large model online and configure standard scripts; S6. Continuously collect conversation data during the mediation process and store it in the system's underlying database for subsequent optimization. Collect feedback from mediators on the use of dialogue technique recommendations and use the feedback data to automatically fine-tune the model.

Images