A Customer Service Dialogue Summary Generation Method Based on a Large Model and a Readable Storage Medium

By using a large-model-based customer service dialogue summary generation method, which leverages pre-trained models and knowledge distillation techniques, the efficiency and accuracy issues of customer service dialogue summary generation are addressed, thereby improving the quality of customer service.

CN117874219BActive Publication Date: 2026-11-10JIAYIN DIGITAL INFORMATION TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410043299.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-11
Publication Date
2026-11-10
Estimated Expiration
2044-01-11

AI Technical Summary

Technical Problem

Existing technologies struggle to efficiently generate high-quality customer service dialogue summaries, leading to time-consuming and labor-intensive manual processing, and making it difficult to guarantee the accuracy and completeness of the summaries.

Method used

A customer service dialogue summary generation method based on a large model is adopted. By collecting communication records and using a pre-trained customer service dialogue summary generation model, combined with knowledge distillation and parameter fine-tuning techniques, high-quality dialogue summaries are generated.

Benefits of technology

It enables the automatic generation of useful summaries from customer service conversations, improving customer service efficiency and service quality, and enhancing the accuracy and completeness of the summaries.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117874219B_ABST
    Figure CN117874219B_ABST
Patent Text Reader

Abstract

This invention discloses a customer service dialogue summary generation method and a readable storage medium based on a large-scale model. The method includes: first, collecting communication records between a target customer and a target customer service representative; then, inputting these communication records into a pre-trained customer service dialogue summary generation model, and outputting the corresponding target customer service dialogue summary. Notably, this customer service dialogue summary generation model is trained based on a large-scale model. This method effectively enables the automatic generation of useful summaries from massive customer service dialogues, helping to improve customer service efficiency and service quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for generating customer service dialogue summaries based on a large model and a readable storage medium. Background Technology

[0002] With the development of internet technology, customer service has become a crucial channel for communication between businesses and users. However, because customer service conversations typically contain a large amount of information and are complex and diverse, extracting key information from them is extremely difficult. Traditional methods usually require manual conversation summarization, which is not only time-consuming and labor-intensive but also struggles to guarantee the accuracy and completeness of the summaries. Therefore, how to generate high-quality summaries from customer service conversations through automation is a pressing issue that needs to be addressed. Summary of the Invention

[0003] The purpose of this invention is to provide a method for generating customer service dialogue summaries based on a large model and a readable storage medium.

[0004] In a first aspect, embodiments of the present invention provide a method for generating customer service dialogue summaries based on a large model, the method comprising:

[0005] Collect communication records between target customers and target customer service representatives;

[0006] The communication record is input into a pre-trained customer service dialogue summary generation model, and the target customer service dialogue summary corresponding to the communication record is output. The customer service dialogue summary generation model is trained based on a large model.

[0007] In one possible implementation, the customer service dialogue summary generation model is trained in the following ways:

[0008] Obtain sample communication records and clean the sample communication records;

[0009] The cleaned sample communication records are input into the first initial large model for training to obtain the intermediate dialogue summary training results.

[0010] Based on the training results of the intermediate dialogue summary, knowledge distillation is performed on the second initial large model to obtain the customer service dialogue summary generation model, wherein the number of parameters of the first initial large model exceeds the number of parameters of the second initial large model.

[0011] In one possible implementation, the step of acquiring sample communication records and cleaning the sample communication records includes:

[0012] The sample communication records are converted from speech to text, and regular expressions are used to clean the sample communication records.

[0013] In one possible implementation, the first initial large model is a GPT-4 large model, and the step of inputting the cleaned sample communication records into the first initial large model for training to obtain intermediate dialogue summary training results includes:

[0014] The cleaned sample communication records were input into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model.

[0015] The basic customer service dialogue summary generation model is used as the training result of the intermediate dialogue summary.

[0016] In one possible implementation, the step of inputting the cleaned sample communication records into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model includes:

[0017] The cleaned sample communication records were input into the GPT-4 large model, and prompt words were constructed;

[0018] Based on the ICL capability of the GPT-4 large model, multiple task examples are injected into the prompt words, and the task results are output so that the GPT-4 large model can learn the knowledge required for the task.

[0019] Based on the COT capability of the GPT-4 model, the prompt words guide the GPT-4 model to output the reasoning process of a specific task, thereby improving the model's output accuracy.

[0020] The basic customer service dialogue summary generation model is obtained by training the model by combining the ICL capability and the COT capability.

[0021] In one possible implementation, the second initial large model is a pre-defined open-source large model. The step of performing knowledge distillation on the second initial large model based on the intermediate dialogue summary training results to obtain the customer service dialogue summary generation model includes:

[0022] The basic customer service dialogue summary generation model is used as the teacher model, and the preset open-source large model is used as the student model.

[0023] Based on the model relationship between the teacher model and the student model, the summary knowledge of the basic customer service dialogue summary generation model is distilled into the preset open source large model through knowledge distillation.

[0024] Based on the pre-defined open-source large model after knowledge distillation, the customer service dialogue summary generation model is obtained.

[0025] In one possible implementation, the customer service dialogue summary generation model, derived from the pre-defined open-source large model after knowledge distillation, includes:

[0026] The parameters of the instruction training set corresponding to the preset open-source large model are fine-tuned using P-tuningV2 and lora training methods.

[0027] The preset open-source large model is trained based on the finely tuned instruction training set to obtain the customer service dialogue summary generation model.

[0028] Secondly, embodiments of the present invention provide a customer service dialogue summary generation device based on a large model, the device comprising:

[0029] The acquisition module is used to collect communication records between target customers and target customer service representatives;

[0030] The generation module is used to input the communication record into a pre-trained customer service dialogue summary generation model and output the target customer service dialogue summary corresponding to the communication record. The customer service dialogue summary generation model is trained based on a large model.

[0031] Thirdly, embodiments of the present invention provide a computer device, the computer device including a processor and a non-volatile memory storing computer instructions, wherein when the computer instructions are executed by the processor, the computer device executes the customer service dialogue summary generation method based on a large model as described in at least one possible implementation of the first aspect.

[0032] Fourthly, embodiments of the present invention provide a readable storage medium, the readable storage medium including a computer program, wherein when the computer program is executed, it controls the computer device where the readable storage medium is located to execute the customer service dialogue summary generation method based on a large model as described in at least one possible implementation of the first aspect.

[0033] Compared to existing technologies, the beneficial effects of this invention include: employing a customer service dialogue summary generation method and readable storage medium based on a large model disclosed in this invention, communication records between target customers and target customer service representatives are collected; then, these communication records are input into a pre-trained customer service dialogue summary generation model, outputting the corresponding target customer service dialogue summary. It is worth noting that this customer service dialogue summary generation model is trained based on a large model. This method effectively achieves the automatic generation of useful summaries from massive customer service dialogues, helping to improve customer service efficiency and service quality. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A flowchart illustrating the steps of a customer service dialogue summary generation method based on a large model provided in an embodiment of the present invention;

[0036] Figure 2 A schematic block diagram of the structure of a customer service dialogue summary generation device based on a large model provided in an embodiment of the present invention;

[0037] Figure 3 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0039] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0040] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0041] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0042] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0043] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0044] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the customer service dialogue summary generation method based on a large model provided in this embodiment. The following is a detailed description of the customer service dialogue summary generation method based on a large model.

[0046] Step S201: Collect communication records between the target customer and the target customer service representative;

[0047] Step S202: Input the communication record into a pre-trained customer service dialogue summary generation model, and output the target customer service dialogue summary corresponding to the communication record. The customer service dialogue summary generation model is trained based on a large model.

[0048] In this embodiment of the invention, for example, suppose a company provides online customer service support and records chat logs between customers and customer service representatives. These logs may include text messages, voice recordings, or video conversations. Each time a customer contacts customer service for help, the chat log is saved. For example, a customer encounters a problem while purchasing a product and communicates with customer service through an online customer service platform. All information from this communication, including the customer's questions, the customer service representative's answers, and the interactions between the two parties, is recorded. In this step, the previously collected communication logs are input into a pre-trained customer service dialogue summary generation model. This model has been trained with a large amount of data to understand and summarize key information from the dialogue. For example, the aforementioned customer-customer dialogue log is input into the customer service dialogue summary generation model. The model analyzes the text content, tone, and context of the dialogue and attempts to extract the key points and themes. After model analysis, in this step, a target customer service dialogue summary corresponding to the dialogue log is obtained. This summary summarizes the entire dialogue to provide more concise and easily understood content. For example, the customer-customer dialogue summary generation model is used to summarize the customer-customer dialogue. The summary may include a summary of key questions, important answers, and solutions to help other customer service staff understand the conversation more quickly and answer similar questions more effectively.

[0049] This design, by collecting communication records between target customers and target customer service representatives and inputting them into a pre-trained customer service dialogue summarization model, can output dialogue summaries that can be used by other customer service personnel. This method leverages the training results of a large model, making dialogue summarization more accurate and efficient.

[0050] In one possible implementation, the aforementioned customer service dialogue summary generation model is trained in the following manner.

[0051] (1) Obtain sample communication records and clean the sample communication records;

[0052] (2) Input the cleaned sample communication records into the first initial large model for training to obtain the intermediate dialogue summary training results;

[0053] (3) Based on the training results of the intermediate dialogue summary, knowledge distillation is performed on the second initial large model to obtain the customer service dialogue summary generation model, wherein the number of parameters of the first initial large model exceeds the number of parameters of the second initial large model.

[0054] In this embodiment of the invention, exemplarily, in this step, sample communication records for training the customer service dialogue summary generation model need to be obtained and these records cleaned. The cleaning process may include removing irrelevant or redundant information, correcting erroneous or incomplete records, to ensure the quality and accuracy of the training data. For example, to train the customer service dialogue summary generation model, a company may collect a large number of chat records between customers and customer service representatives and filter and clean these records. The cleaned records should contain only valid dialogue content, removing irrelevant advertisements, greetings, or other noise information. In this step, the cleaned sample communication records are input into a first initial large-scale model for training to obtain intermediate dialogue summary training results. The first initial large-scale model may be a deep learning model that understands and summarizes dialogue content through training on a large-scale dataset. For example, cleaned customer-customer dialogue records are used as training data and input into the first initial large-scale model for training. The model learns how to extract key information from the dialogue, understand questions and answers, and generate preliminary dialogue summary results. In this step, the intermediate dialogue summary training results are used for knowledge distillation to generate the final customer service dialogue summary generation model. Knowledge distillation is a technique that transfers knowledge from a complex model to a smaller, more efficient model. For example, after a first initial large model learns rich knowledge about dialogue summarization, this knowledge is transferred to a second initial large model through knowledge distillation, allowing the second model to maintain similar performance with fewer parameters. This results in the customer service dialogue summarization model.

[0055] This design involves first acquiring and cleaning sample communication records, then inputting the cleaned records into a first initial large-scale model for training, yielding intermediate dialogue summary training results. Finally, based on the intermediate training results, knowledge distillation is performed on a second initial large-scale model to generate the final customer service dialogue summary generation model. This method fully utilizes the knowledge transfer between the two initial large-scale models to improve model performance and efficiency.

[0056] In one possible implementation, the steps of acquiring sample communication records and cleaning the sample communication records described above can be performed through the following example.

[0057] (1) Perform voice-to-text conversion on the sample communication records and clean the sample communication records using regular expressions.

[0058] In this embodiment of the invention, for example, in this step, it is necessary to convert the voice information in the sample communication records into text form. This can be achieved through speech recognition technology, converting the recorded voice content into text data that can be processed and analyzed by a computer. For example, a company collects telephone recordings between customers and customer service representatives as sample communication records. In order to clean and train the model, the voice information in these telephone recordings needs to be converted into text form for subsequent processing and analysis. In this step, regular expressions are used to clean the converted sample communication records. Regular expressions are tools used to match and process text patterns, and can perform operations such as filtering, replacing, or deleting text according to specific rules. For example, after conversion to text, the sample communication records may contain some unnecessary information, such as noise, repeated text, or other interfering elements. By designing appropriate regular expressions, this irrelevant information can be removed, retaining only the meaningful dialogue content.

[0059] This design first converts the speech information into text format, making it processable by a computer. Next, regular expressions are used to clean the converted text, removing unnecessary information and distracting elements. This ensures the accuracy and quality of the sample data, providing a reliable data foundation for subsequent model training.

[0060] In one possible implementation, the aforementioned first initial large model is a GPT-4 large model, and the aforementioned step of inputting the cleaned sample communication records into the first initial large model for training to obtain the intermediate dialogue summary training result can be implemented through the following example.

[0061] (1) Input the cleaned sample communication records into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model;

[0062] (2) The basic customer service dialogue summary generation model is used as the training result of the intermediate dialogue summary.

[0063] In this embodiment of the invention, exemplarily, in this step, the Wenxin-4 large-scale model is used to train the cleaned sample communication records to obtain a basic customer service dialogue summary generation model. Wenxin-4 is an advanced large-scale language model with powerful natural language processing capabilities. For example, the previously cleaned customer-customer dialogue records are input into the Wenxin-4 large-scale model for training. The model learns the semantics, context, and key information of the dialogue based on the input text data and attempts to generate a basic result of dialogue summary. In this step, the basic customer service dialogue summary generation model is used as an intermediate dialogue summary training result. The basic model is obtained by training with Wenxin-4 and contains the basic capabilities and knowledge of dialogue summarization. For example, the basic customer service dialogue summary generation model is used as an intermediate training result for further knowledge distillation or other training processes. In this way, the basic model obtained based on the Wenxin-4 large-scale model can serve as an intermediate result to provide the basic capabilities and starting point for the dialogue summary generation model.

[0064] This design involves first inputting the cleaned sample communication records into the Wenxin-4 model for training, resulting in a basic customer service dialogue summary generation model. Then, this basic model is used as an intermediate training result for further training or knowledge distillation. This approach fully leverages the powerful capabilities of the Wenxin-4 model, improving the quality and effectiveness of dialogue summary generation.

[0065] In one possible implementation, the aforementioned step of inputting the cleaned sample communication records into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model can be performed through the following example.

[0066] (1) Input the cleaned sample communication records into the GPT-4 large model and construct prompt words;

[0067] (2) Based on the ICL capability of the GPT-4 large model, multiple task examples are injected into the prompt words and the task results are output so that the GPT-4 large model can learn the knowledge required for the task.

[0068] (3) Based on the COT capability of the GPT-4 model, the prompt words guide the GPT-4 model to output the reasoning process of specific tasks, so as to improve the output accuracy of the model;

[0069] (4) The basic customer service dialogue summary generation model is obtained by training the ICL capability and the COT capability.

[0070] In this embodiment of the invention, exemplarily, in this step, when training the basic customer service dialogue summary generation model using the Wenxin-4 model, cleaned sample communication records are used as input, and corresponding prompt words are constructed. The prompt words guide the model in generating summary content for relevant tasks. For example, cleaned customer-customer service dialogue records are input into the Wenxin-4 model as training data, and corresponding prompt words are constructed for each dialogue, such as "Problem Description:", "Solution:", etc. In this way, the model can more accurately extract and summarize the key information of the dialogue based on the guidance of the prompt words when generating summaries. In this step, the In-the-Context Learning (ICL) capability of the Wenxin-4 model is utilized to inject multiple task examples into the prompt words and output the corresponding task results. This allows the Wenxin-4 model to learn the knowledge required to perform different tasks. For example, by injecting multiple task examples into the prompt words, such as "Problem Type: Product Inquiry, Account Problem, Complaint Feedback, etc.", the Wenxin-4 model can learn the summary generation methods for different task types based on these examples and output the corresponding task results. In this way, the model can more accurately extract relevant information and generate appropriate dialogue summaries when processing specific tasks. In this step, the Chain of Thought (COT) capability of the Wenxin-4 model is utilized to guide the model's output of the reasoning process for specific tasks through the provided prompts, thereby improving the model's output accuracy. For example, by designing appropriate prompts and context, such as "problem description" and "solution," the Wenxin-4 model is guided to reason and construct logical chains when generating dialogue summaries. This improves the logical coherence and accuracy of the generated summaries. In this step, the Wenxin-4 model is trained by combining ICL and COT capabilities to obtain a basic customer service dialogue summary generation model. By combining these two capabilities, the model can better understand task examples, learn reasoning processes, and generate accurate and coherent dialogue summaries. For example, the Wenxin-4 model is trained using sample data injected with multiple task examples and guided reasoning processes. During training, both ICL and COT capabilities are considered simultaneously so that the model can gradually learn dialogue summary generation techniques and reasoning abilities for different task types. The final basic customer service dialogue summary generation model will be more accurate and reliable.

[0071] In one possible implementation, the second initial large model is a preset open-source large model. The aforementioned step of performing knowledge distillation on the second initial large model based on the intermediate dialogue summary training results to obtain the customer service dialogue summary generation model can be implemented through the following example.

[0072] (1) Use the basic customer service dialogue summary generation model as the teacher model and the preset open source big model as the student model.

[0073] (2) Based on the model relationship between the teacher model and the student model, the summary knowledge of the basic customer service dialogue summary generation model is distilled into the preset open source large model through knowledge distillation;

[0074] (3) Based on the pre-set open-source big model after knowledge distillation, the customer service dialogue summary generation model is obtained.

[0075] In this embodiment of the invention, for example, in this step, a basic customer service dialogue summary generation model is used as the teacher model, and a preset open-source large model is selected as the student model. The teacher model is used to impart knowledge and guide the learning of the student model. For example, the basic customer service dialogue summary generation model trained based on the Wenxin-4 large model is used as the teacher model, and a publicly available, preset open-source large model, such as the Wenxin-2 model, is selected as the student model. In this step, according to the model relationship between the teacher model and the student model, the summary knowledge possessed by the basic customer service dialogue summary generation model is transferred to the preset open-source large model through knowledge distillation. For example, by designing appropriate knowledge distillation algorithms and training methods, the key knowledge, skills, and predictive capabilities in the basic customer service dialogue summary generation model are transferred to the preset open-source large model. In this way, the student model can learn from the teacher model's experience and knowledge, improving its own generation capabilities and effects. In this step, the final customer service dialogue summary generation model is obtained based on the preset open-source large model that has undergone knowledge distillation.

[0076] For example, a pre-built open-source large model, after knowledge distillation, can be used as the base model, and further optimized and trained to obtain a customer service dialogue summarization generation model with superior performance. In this way, the model will possess the knowledge and capabilities of the base model based on the Wenxin-4 large model, while combining the advantages of the pre-built open-source large model to generate higher-quality dialogue summarization results.

[0077] This design first selects a base model as the teacher model and a pre-defined open-source model as the student model. Then, based on the teacher-student model relationship, the summarized knowledge from the base model is transferred to the student model through knowledge distillation. Finally, based on the pre-defined open-source model that has undergone knowledge distillation, a higher-performance customer service dialogue summary generation model is obtained.

[0078] In one possible implementation, the step of obtaining the customer service dialogue summary generation model based on the aforementioned preset open-source large model after knowledge distillation can be performed through the following example.

[0079] (1) Fine-tune the parameters of the instruction training set corresponding to the preset open-source large model using P-tuningV2 and lora training methods;

[0080] (2) The preset open-source large model is trained based on the fine-tuned instruction training set to obtain the customer service dialogue summary generation model.

[0081] In this embodiment of the invention, exemplarily, in this step, a training set of instructions is first prepared, which contains specific instructions and data required for the customer service dialogue summary generation task. Then, the parameters of a pre-defined open-source large model are fine-tuned using P-tuningV2 and LoRa training methods to better adapt it to these instructions and data. For example, suppose there is an instruction training set containing instructions related to customer dialogue summary generation, such as "generate a summary of a dialogue with a customer" or "answer customer questions." P-tuningV2 and LoRa training methods can be used to fine-tune the parameters of the pre-defined open-source large model (such as Wenxin-2) by training on this training set, making it more focused on performing operations related to the customer service dialogue summary generation task. In this step, the parameter-fine-tuned instruction training set is applied to the pre-defined open-source large model and further trained to obtain the final customer service dialogue summary generation model. Specifically, the fine-tuned instruction training set is used to train the pre-defined open-source large model. This means that samples from this training set are used to iteratively train and optimize the model parameters, continuously adjusting the model's weights and structure to better adapt it to the customer service dialogue summary generation task. Ultimately, this will result in a higher-performing and higher-quality customer service dialogue summarization model. For example, a finely tuned instruction training set can be used to update and optimize the parameters of a pre-defined open-source model through multiple iterations of training. In this way, the model will gradually learn how to generate accurate and concise dialogue summaries based on given instructions and context. After training, this customer service dialogue summarization model will be better able to meet user needs and provide higher-quality and more accurate dialogue summaries.

[0082] This design, during the parameter fine-tuning phase, uses an instruction training set and P-tuningV2 and LoRa training methods to adjust the parameters of the pre-set open-source large model to better suit the customer service dialogue summarization task. Then, further training is performed on the fine-tuned instruction training set, enabling the pre-set open-source large model to generate higher-quality dialogue summarization results. In this way, the final customer service dialogue summarization generation model possesses the capabilities of fine-tuning and training, better meeting user needs and generating accurate dialogue summaries.

[0083] The following provides an overall implementation process for an embodiment of the present invention.

[0084] (1) Data preprocessing: Since the text comes from the ASR system (Automatic Speech-to-Text), it often contains many reduplicated words and colloquialisms such as "hey," "um," and "ah." The dialogue text data is cleaned using rules (regular expressions) to obtain a higher quality text result.

[0085] (2) Fine-tuning Data Generation: Obtaining a certain amount of summary instructions for fine-tuning data is time-consuming and costly. To alleviate the cost of manual annotation and improve data acquisition efficiency, large-scale models such as GPT-4, which have high generation costs, were used to output summaries of the dialogue content. The main steps in data generation are the design of prompt words and the use of the large-scale model's ICL (In-the-Context Learning) and COT (Chain of Thought) capabilities to guide GPT-4 to output high-quality summaries.

[0086] ICL: Contextual learning capability of large models. By injecting multiple task examples into the prompt words, the model can quickly learn the knowledge required for the task and output the task results.

[0087] COT: The thought process of a large model. Guiding the large model through prompts to output the reasoning process for a specific task can significantly improve the model's output accuracy.

[0088] (3) Data Filtering: Due to the instability of the output content of large models, it is necessary to review and filter the output results. Filtering methods include machine filtering and human supervision. Machine filtering mainly reviews the content format to prevent the large model from generating obviously erroneous results such as duplicate prompts or blank results. Human supervision relies on data analysts to filter the data; since it does not require actively writing summaries but only judging the quality of the summaries, efficiency is greatly improved.

[0089] (4) Fine-tuning of the open-source large model: The open-source ChatGLM-6B dialogue pre-training model is adopted. The parameters are fine-tuned on the constructed instruction training set using fine-tuning training methods such as P-tuningV2 and lora.

[0090] ChatGLM-6B: A commercially viable open-source dialogue model with 6 billion records released by Tsinghua Zhipu. It performs excellently on various large model test sets.

[0091] P-tuningV2: Inserts additional parameters into the input and intermediate layers of a large model, freezes the already trained parameters, and updates the new weights during training.

[0092] Lora: Additional parameter matrices are added as a bypass to large models; the model itself does not participate in training, but only debugs the added matrices.

[0093] (5) Model testing: The fine-tuned model is tested using the ROGUE method.

[0094] (6) Model deployment: In order to cope with real-time scenarios, the model is quantized to reduce the memory and computational load of deployment.

[0095] Please refer to the following: Figure 2 , Figure 2 A customer service dialogue summary generation device 110 based on a large model is provided for embodiments of the present invention. The device includes:

[0096] The acquisition module 1101 is used to collect communication records between target customers and target customer service representatives;

[0097] The generation module 1102 is used to input the communication record into a pre-trained customer service dialogue summary generation model and output the target customer service dialogue summary corresponding to the communication record. The customer service dialogue summary generation model is trained based on a large model.

[0098] It should be noted that the implementation principle of the aforementioned customer service dialogue summary generation device 110 based on a large model can be referred to in the above description of the implementation principle of the customer service dialogue summary generation device 110 based on a large model, and will not be repeated here. It should be understood that the division of the various modules in the above device is merely a logical functional division; in actual implementation, they can be fully or partially integrated into a single physical entity, or physically separated. Furthermore, these modules can all be implemented in software through processing element calls; they can all be implemented in hardware; or some modules can be implemented by processing element calls to software, and some modules can be implemented in hardware. For example, the customer service dialogue summary generation device 110 based on a large model can be a separately established processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and called and executed by a processing element of the above device. The implementation of other modules is similar. Furthermore, these modules can be fully or partially integrated together, or implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step or module of the above method can be completed by the integrated logic circuit in the hardware of the processor element or by instructions in the form of software.

[0099] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together to implement a system-on-a-chip (SOC).

[0100] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned customer service dialogue summary generation device 110 based on a large model. Figure 3 As shown, Figure 3 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a customer service dialogue summary generation device 110 based on a large model, a memory 111, a processor 112, and a communication unit 113.

[0101] To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The large-model-based customer service dialogue summary generation device 110 includes at least one software function module that can be stored in the memory 111 or embedded in the operating system (OS) of the computer device 100 in the form of software or firmware. The processor 112 is used to execute the large-model-based customer service dialogue summary generation device 110 stored in the memory 111, such as the software function modules and computer programs included in the large-model-based customer service dialogue summary generation device 110.

[0102] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to employ various embodiments with different modifications to suit a particular intended application. For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for generating customer service dialogue summaries based on a large model, characterized in that, The method includes: Collect communication records between target customers and target customer service representatives; The communication record is input into a pre-trained customer service dialogue summary generation model, and the target customer service dialogue summary corresponding to the communication record is output. The customer service dialogue summary generation model is trained based on a large model. The customer service dialogue summary generation model is trained in the following ways: Obtain sample communication records and clean the sample communication records; The cleaned sample communication records are input into the first initial large model for training to obtain the intermediate dialogue summary training results. Based on the training results of the intermediate dialogue summary, knowledge distillation is performed on the second initial large model to obtain the customer service dialogue summary generation model, wherein the number of parameters of the first initial large model exceeds the number of parameters of the second initial large model. The first initial large model is a GPT-4 large model. The step of inputting the cleaned sample communication records into the first initial large model for training, and obtaining the intermediate dialogue summary training results, includes: The cleaned sample communication records were input into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model. The basic customer service dialogue summary generation model is used as the training result of the intermediate dialogue summary. The process involves inputting the cleaned sample communication records into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model, including: The cleaned sample communication records were input into the GPT-4 large model, and prompt words were constructed; Based on the ICL capability of the GPT-4 large model, multiple task examples are injected into the prompt words, and the task results are output so that the GPT-4 large model can learn the knowledge required for the task. Based on the COT capability of the GPT-4 model, the prompt words guide the GPT-4 model to output the reasoning process of a specific task, thereby improving the model's output accuracy. The basic customer service dialogue summary generation model is obtained by training the ICL and COT capabilities together. The second initial large model is a preset open-source large model. The step of performing knowledge distillation on the second initial large model based on the intermediate dialogue summary training results to obtain the customer service dialogue summary generation model includes: The basic customer service dialogue summary generation model is used as the teacher model, and the preset open-source large model is used as the student model. Based on the model relationship between the teacher model and the student model, the summary knowledge of the basic customer service dialogue summary generation model is distilled into the preset open source large model through knowledge distillation. Based on the pre-defined open-source large model after knowledge distillation, the customer service dialogue summary generation model is obtained.

2. The method according to claim 1, characterized in that, The process of obtaining and cleaning sample communication records includes: The sample communication records are converted from speech to text, and regular expressions are used to clean the sample communication records.

3. The method according to claim 1, characterized in that, The customer service dialogue summary generation model, obtained from the pre-set open-source large model after knowledge distillation, includes: The parameters of the instruction training set corresponding to the preset open-source large model are fine-tuned using P-tuningV2 and lora training methods. The preset open-source large model is trained based on the finely tuned instruction training set to obtain the customer service dialogue summary generation model.

4. A customer service dialogue summary generation device based on a large model, characterized in that, The device includes: The acquisition module is used to collect communication records between target customers and target customer service representatives; The generation module is used to input the communication record into a pre-trained customer service dialogue summary generation model and output the target customer service dialogue summary corresponding to the communication record. The customer service dialogue summary generation model is trained based on a large model. The customer service dialogue summary generation model is trained in the following ways: Obtain sample communication records and clean the sample communication records; The cleaned sample communication records are input into the first initial large model for training to obtain the intermediate dialogue summary training results. Based on the training results of the intermediate dialogue summary, knowledge distillation is performed on the second initial large model to obtain the customer service dialogue summary generation model, wherein the number of parameters of the first initial large model exceeds the number of parameters of the second initial large model. The first initial large model is a GPT-4 large model. The step of inputting the cleaned sample communication records into the first initial large model for training, and obtaining the intermediate dialogue summary training results, includes: The cleaned sample communication records were input into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model. The basic customer service dialogue summary generation model is used as the training result of the intermediate dialogue summary. The process involves inputting the cleaned sample communication records into the GPT-4 large model for training to obtain a basic customer service dialogue summary generation model, including: The cleaned sample communication records were input into the GPT-4 large model, and prompt words were constructed; Based on the ICL capability of the GPT-4 large model, multiple task examples are injected into the prompt words, and the task results are output so that the GPT-4 large model can learn the knowledge required for the task. Based on the COT capability of the GPT-4 model, the prompt words guide the GPT-4 model to output the reasoning process of a specific task, thereby improving the model's output accuracy. The basic customer service dialogue summary generation model is obtained by training the ICL and COT capabilities together. The second initial large model is a preset open-source large model. The step of performing knowledge distillation on the second initial large model based on the intermediate dialogue summary training results to obtain the customer service dialogue summary generation model includes: The basic customer service dialogue summary generation model is used as the teacher model, and the preset open-source large model is used as the student model. Based on the model relationship between the teacher model and the student model, the summary knowledge of the basic customer service dialogue summary generation model is distilled into the preset open source large model through knowledge distillation. Based on the pre-defined open-source large model after knowledge distillation, the customer service dialogue summary generation model is obtained.

5. A computer device, characterized in that, The computer device includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device performs the customer service dialogue summary generation method based on any one of claims 1-3.

6. A readable storage medium, characterized in that, The readable storage medium includes a computer program, which, when executed, controls the computer device on which the readable storage medium is located to perform the customer service dialogue summary generation method based on any one of claims 1-3.

Citation Information

Patent Citations

  • Corpus expansion method and device under multiple tasks based on small language model

    CN116595130A

  • Artificial intelligence-based intention recognition method and device, equipment and storage medium

    CN117057362A