Customer service public opinion monitoring method, system, medium and equipment

Abstract

The invention provides a customer service public opinion monitoring method. The method comprises the steps of obtaining customer service dialogue data; performing dialogue content understanding on the customer service dialogue data by utilizing the main model, and determining a dialogue scene which most conforms to the customer service dialogue data in the public opinion definition scene according to the dialogue content; returning a hit dialogue according to the dialogue scene output by the master model by using the slave model; calculating dialogue understanding quality of a specific dialogue in the dialogue scene and response quality of a hit dialogue; and if the corresponding set threshold value is met, outputting a public opinion monitoring result corresponding to the dialogue scene. Compared with recognition of a single model, the method has the advantages that the quality of recall results is higher, dialogue content and public opinion scenes are bidirectionally recognized through the dual model, the understanding depth of the model for customer service dialogue data can be enhanced, and the understanding accuracy and reliability are improved through a double-layer verification mechanism of the model. The invention further provides a customer service public opinion monitoring system, a computer readable storage medium and electronic equipment, which have the above beneficial effects.

Description

Technical Field

[0001] This application relates to the field of data processing, and particularly to a customer service public opinion monitoring method, system, medium and device. Background Art

[0002] In the field of public opinion analysis, currently three main methods are adopted:

[0003] One is the rule-based method, which identifies public opinion scenarios through predefined keywords or pattern matching, being relatively simple and not flexible enough.

[0004] The second is the machine learning method, such as pre-trained language models like BERT and RoBERTa. Although these models have made significant progress in processing text sentiment classification, there are certain limitations in dealing with dialogue data. The content of dialogues usually has the characteristics of non-content complexity and strong context dependence, which makes it difficult to effectively apply traditional single-sentence extraction tasks, especially in capturing complex semantic information.

[0005] The third is to utilize large-scale pre-trained language models, such as the Chatgpt series, Llama series, Qwen series, Chatglm series, etc. These models have made obvious improvements in dialogue understanding and content generation, but the effects in this vertical field of public opinion recognition are not satisfactory. The main reasons are that the models have insufficient knowledge reserves in specific fields or lack the ability to understand complex scenarios. Summary of the Invention

[0006] The purpose of this application is to provide a customer service public opinion monitoring method, system, computer-readable storage medium and electronic device, which can improve the public opinion detection ability in the customer service scenario.

[0007] To solve the above technical problems, this application provides a customer service public opinion monitoring method, and the specific technical solution is as follows:

[0008] Obtain customer service dialogue data;

[0009] Use the main model in the dual model to understand the dialogue content of the customer service dialogue data, and determine the dialogue scenario in the public opinion definition scenario that best matches the customer service dialogue data according to the dialogue content;

[0010] Use the slave model in the dual model to return the hit dialogue according to the dialogue scenario output by the main model;

[0011] Calculate the dialogue understanding quality of the specific dialogue in the dialogue scenario and the response quality of the hit dialogue;

[0012] If both the dialogue understanding quality and the response quality meet the corresponding set thresholds, output the public opinion monitoring result corresponding to the dialogue scenario.

[0013] Optionally, after obtaining the customer service conversation data, it further includes:

[0014] Filter the noise information in the customer conversation data by using regular expressions; the noise information includes modal particles and empty conversations.

[0015] Optionally, before using the main model in the dual model to understand the conversation content of the customer service conversation data and determining the conversation scenario that best matches the customer service conversation data in the public opinion definition scenario, it further includes:

[0016] Obtain the manually annotated data;

[0017] Train based on the manually annotated data to obtain the dual model.

[0018] Optionally, training based on the manually annotated data to obtain the dual model includes:

[0019] Store the manually annotated data on each distributed node, and in each training iteration, use the distributed training tool to distribute the input data to different distributed nodes;

[0020] Perform forward propagation on each distributed node to calculate the loss function value;

[0021] Use the amp.scale_loss function in the Apex library to scale the loss;

[0022] Perform backward propagation on each distributed node to calculate the gradient;

[0023] Synchronize and average the gradients on all distributed nodes;

[0024] Use the optimizer to update the model parameters;

[0025] Zero out the gradients to prepare for the next iteration, and end the training until the number of iterations is met or convergence occurs to obtain the dual model.

[0026] Optionally, when zeroing out the gradients to prepare for the next iteration and ending the training until the number of iterations is met to obtain the dual model, it further includes:

[0027] Adjust the learning rate during the training process of the dual model according to the periodic characteristics of the cosine function.

[0028] Optionally, if either the conversation understanding quality or the response quality does not meet the corresponding set threshold, it further includes:

[0029] Convert the customer service conversation data into a vector representation;

[0030] Segment the vector representation and calculate the cosine similarity using the semantic vector model and the knowledge vectors in the indexing system;

[0031] Select the N pieces of knowledge and conversations with the highest cosine similarity as the input data for the self-developed model; the self-developed model is used to output the conversational implicature of the customer conversation data based on the input data.

[0032] Optionally, after switching the customer service conversation data to a vector representation, it further includes:

[0033] Construct an external knowledge base based on the vector representation; the external knowledge base is used to store the glossary related to temporary events or short-term conversation scenarios, as well as the dynamics of the conversation occurrence;

[0034] Train the self-developed model based on the external knowledge base.

[0035] This application also provides a customer service public opinion monitoring system, including:

[0036] A data acquisition module for acquiring customer service conversation data;

[0037] A bidirectional recognition module for using the main model in the dual model to understand the conversation content of the customer service conversation data, and determining the conversation scenario in the public opinion definition scenario that best matches the customer service conversation data according to the conversation content; and using the slave model in the dual model to return the hit conversation according to the conversation scenario output by the main model;

[0038] A parameter calculation module for calculating the conversation understanding quality of the specific conversation in the conversation scenario and the response quality of the hit conversation;

[0039] A public opinion detection module for outputting the public opinion monitoring result corresponding to the conversation scenario if both the conversation understanding quality and the response quality meet the corresponding set thresholds.

[0040] This application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described above are implemented.

[0041] This application also provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps of the method described above are implemented.

[0042] The present application provides a customer service public opinion monitoring method, including: obtaining customer service conversation data; using the main model in the dual model to understand the conversation content of the customer service conversation data, and determining the conversation scenario in the public opinion definition scenario that best matches the customer service conversation data according to the conversation content; using the slave model in the dual model to return the hit conversation according to the conversation scenario output by the main model; calculating the conversation understanding quality of the specific conversation in the conversation scenario and the response quality of the hit conversation; if both the conversation understanding quality and the response quality meet the corresponding set thresholds, outputting the public opinion monitoring result corresponding to the conversation scenario.

[0043] After obtaining the customer service conversation data, the present application accurately identifies and recalls the public opinion scenario through a dual model including a main model and a slave model. Compared with the identification of a single model, the recall result quality is higher, and it is also convenient to optimize for specific application scenarios. By using the dual model to bidirectionally identify the conversation content and the public opinion scenario, the understanding depth of the model for the customer service conversation data can be enhanced, and the understanding accuracy and reliability are improved through the double-layer verification mechanism of the slave model.

[0044] The present application also provides a customer service public opinion monitoring system, a computer-readable storage medium, and an electronic device, which have the above beneficial effects and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0046] Figure 1 It is a flowchart of the customer service public opinion monitoring method provided by the embodiment of the present application;

[0047] Figure 2 It is a structural schematic diagram of the customer service public opinion monitoring system provided by the embodiment of the present application;

[0048] Figure 3 It is a structural diagram of an electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Apparently, the described embodiments are some, but not all, of the embodiments of this application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts fall within the scope of protection of this application.

[0050] Several currently applied public opinion analysis methods have the following defects:

[0051] First, the data is not targeted enough: Although the existing technology can process a large amount of public opinion text, the support for data types such as customer service conversations, which are highly structured and contain a large number of informal expressions, is limited. Customer service conversations usually have more complex emotional expressions and context dependencies, and these characteristics are not fully considered.

[0052] Second, the ability to follow instructions is insufficient: In a customer service environment, accurately understanding and executing specific instructions is crucial. However, after the existing system is fine-tuned with instructions, it may still not fully meet the requirements for precise instruction following in customer service conversations, especially when dealing with complex business logics or special situations.

[0053] Third, the lack of a verification mechanism: The existing technology mainly improves the accuracy and recall rate of scenarios by optimizing the understanding of conversations, but the instability of the model itself, including hallucinations, is difficult to solve. The lack of a verification mechanism for the extraction results due to one-way (conversation -> scenario) extraction leads to unstable recognition effects.

[0054] Fourth, the real-time adaptation ability is poor: When it comes to emergencies or product updates. The existing public opinion analysis systems are not fast enough in responding to newly emerging topics or temporary events and cannot adjust the analysis strategy in a timely manner to adapt to new situations.

[0055] To address the above defects, please refer to Figure 1 , Figure 1 which is the flowchart of the customer service public opinion monitoring method provided by the embodiments of this application. The method includes:

[0056] S101: Obtain customer service conversation data;

[0057] S102: Use the main model in the dual model to understand the conversation content of the customer service conversation data, and determine the conversation scenario in the public opinion definition scenario that best matches the customer service conversation data according to the conversation content;

[0058] S103: Use the slave model in the dual model to return the hit conversation according to the conversation scenario output by the main model;

[0059] S104: Calculate the dialogue understanding quality of the specific dialogue in the dialogue scenario and the response quality of the hit dialogue;

[0060] S105: If both the dialogue understanding quality and the response quality meet the corresponding set thresholds, output the public opinion monitoring result corresponding to the dialogue scenario.

[0061] Here, there is no limitation on how to obtain customer service dialogue data. The dialogue between the incoming call and the agent can be directly obtained. To improve the detection accuracy, data filtering can be performed, that is, use regular expressions to filter the noise information in the customer dialogue data. The noise information includes modal particles and empty dialogues, etc., and can also include special symbols and punctuation marks, repeated characters, meaningless filler words, etc., so as to avoid the interference of noise information on the extraction of the dual model. Through ablation experiments, it is proved that including data filtering is 3% and 7% higher than the case without it in terms of accuracy and recall rate respectively.

[0062] Before conducting customer service public opinion monitoring, a dual model can be generated first. Specifically, manually labeled data can be obtained, and based on the manually labeled data, training is carried out to obtain the dual model.

[0063] The following is a feasible model generation method:

[0064] First step: Store the manually labeled data on each distributed node. In each training iteration, use the distributed training tool to distribute the input data to different distributed nodes;

[0065] Second step: Perform forward propagation on each distributed node to calculate the loss function value;

[0066] Third step: Use the amp.scale_loss function in the Apex library to scale the loss;

[0067] Fourth step: Perform backward propagation on each distributed node to calculate the gradient;

[0068] Fifth step: Synchronize and average the gradients on all distributed nodes;

[0069] Sixth step: Use the optimizer to update the model parameters;

[0070] Seventh step: Clear the gradient to prepare for the next iteration, and end the training until the iteration number is met or convergence occurs, and obtain the dual model.

[0071] Specifically, based on an open-source large model, tens of thousands of high-quality manually annotated data can be used to fine-tune the model for instruction, so as to standardize the inference process and output format of the model and enable it to have stronger instruction-following capabilities. During the fine-tuning process, the full-scale fine-tuning method is adopted to ensure the effect. The learning rate in the training process of the dual model can also be adjusted according to the periodic characteristics of the cosine function, realizing a fine-tuning process and ensuring that the model performs more excellently on specific tasks.

[0072] The Distributed Data Parallel (DDP) module of PyTorch is used to implement distributed training on multiple GPUs and multiple nodes. At the same time, the Apex library can be used for mixed-precision training. Using these two technologies can improve the training efficiency during training, reduce the memory requirements, and speed up the training speed on the basis of ensuring the accuracy of the model.

[0073] Specifically, first, the data is stored on each node in the distribution. Then, in each training iteration, the DDP module is used to distribute the input data to different GPUs. Forward propagation is performed on each GPU to calculate the loss function value. The amp.scale_loss function in the Apex library is used to scale the loss to ensure that the gradient does not disappear or explode due to the precision problem of FP16 during the backpropagation process. Backpropagation is performed on each GPU to calculate the gradient. The DDP module is used to synchronize and average the gradients on all GPUs. The optimizer is used to update the model parameters. The gradients are cleared to prepare for the next iteration. After each epoch, the validation set can be used to evaluate the model to monitor the training process. When the model reaches satisfactory performance, the parameters of the model are saved to the main process.

[0074] In order to ensure the accuracy and recall rate of recognition, this application designs a dual model based on a large model for bidirectional recognition. This module includes a main model LLM-M and a slave model LLM-F, and different prompt words are designed for the two respectively to perform different tasks. Among them, the main model LLM-M is responsible for understanding the conversation content and finding the scenario that best matches the current conversation from the predefined public opinion scenarios, and outputting the scenario name, the specific conversation hitting the scenario and its thinking process. The slave model LLM-F then accurately searches and returns the specific hitting conversation according to the scenario name provided by LLM-M.

[0075] Then, the Bleu-1, Bleu-2 and ROUGE-1, ROUGE-L indicators of the specific dialogue of the hit scene output by the main model LLM-M and the slave model LLM-F are calculated. If these indicators reach the set threshold, there is no specific limitation on the set threshold, for example, in a feasible implementation, the specific threshold may be that the average value of the four indicators is greater than or equal to 0.7, then the result is confirmed and the result of the main model LLM-M is output.

[0076] If any of the conversation understanding quality and the response quality does not meet the corresponding set threshold, the customer service conversation data can be switched to a vector representation, so that the vector representation is segmented, and the cosine similarity is calculated using the semantic vector model and the knowledge vector in the indexing system. Thereafter, the N knowledge and conversations with the highest cosine similarity are selected as the input data of the self-developed model. The self-developed model is used to output the conversational meaning of the customer conversation data based on the input data. N can be set by those skilled in the art and is not specifically limited here.

[0077] At the same time, an external knowledge base can be constructed based on the vector representation, and the external knowledge base is used to store the explanations of terms related to temporary events or short-term dialogue scenarios, as well as the dynamics of dialogue occurrence, so as to train a self-developed model based on the external knowledge base. In a feasible implementation, the process can be based on RAG (Retrieval-Augmented Generation) technology, and the bge model (a high-quality text embedding model that aims to convert text into low-dimensional dense vectors for efficient calculation and analysis) can be used to vectorize the text and store it in Faiss (an open source library that provides efficient and reliable similarity clustering and retrieval methods for massive data (dense vectors) in high-dimensional space, and can support searches of billions of vectors) to build an external knowledge base. The knowledge base mainly contains definitions and explanations of terms related to temporary events or short-term scenarios. The content of the knowledge base can be dynamically adjusted according to the date of the dialogue to better help the model understand the dialogue. The output of the recognition enhancement module is the final output, and the algorithm ends here, and the next round of dialogue recognition is performed.

[0078] It can be seen that in this embodiment, by applying data augmentation techniques, the dedicated dialogue data is made more diverse, enabling the model to better understand complex conversations. Through full-scale fine-tuning with high-quality manually annotated data, the present invention effectively improves the performance of the model on specific tasks, especially the instruction-following ability and the consistency of output formats, making it more in line with the actual application requirements. When the bidirectional recognition of the dual model fails to reach the expected accuracy, the enhanced recognition stage based on retrieval-augmented generation technology is initiated, achieving the effective utilization of external knowledge. In particular, the knowledge base can be dynamically updated according to the date of the conversation, containing the latest information related to temporary events or short-term scenarios, which enables the model to adapt to changing social topics and emergencies in real time and provide more accurate services.

[0079] This application can be applied to IT support, home appliance repair, software products, etc. It supports scenarios such as analyzing the cause of faults from conversations, the troubleshooting steps that have been tried, accelerating problem diagnosis, providing targeted solutions, and improving the user experience. It can also be applied to restaurant reservation and event booking scenarios: analyzing dining preferences and event types from customer service conversation data to ensure the accuracy of reservation information, improve customer satisfaction, and optimize resource allocation. And in the intelligent customer service scenario, this application can identify key information in customer service conversation data to provide more reference information for the responses of human customer service or robot customer service.

[0080] See Figure 2 , Figure 2 FIG. is a schematic structural diagram of the customer service public opinion monitoring system provided by the embodiment of the present application. The customer service public opinion monitoring system provided by the embodiment of the present application will be introduced below. The customer service public opinion monitoring system described below can be mutually corresponding and referred to the customer service public opinion monitoring method described above.

[0081] The present application also provides a customer service public opinion monitoring system, which includes:

[0082] A data acquisition module for acquiring customer service conversation data;

[0083] A bidirectional recognition module for using the main model in the dual model to understand the conversation content of the customer service conversation data and determine the conversation scenario that best matches the customer service conversation data in the public opinion definition scenario; and using the slave model in the dual model to return the hit conversation according to the conversation scenario output by the main model;

[0084] A parameter calculation module for calculating the conversation understanding quality of the specific conversation in the conversation scenario and the response quality of the hit conversation;

[0085] A public opinion detection module for outputting the public opinion monitoring result corresponding to the conversation scenario if both the conversation understanding quality and the response quality meet the corresponding set thresholds.

[0086] Based on the above embodiments, as a preferred embodiment, it further includes:

[0087] A filtering module, configured to filter out noise information in the customer conversation data by using regular expressions; the noise information includes modal particles and empty conversations.

[0088] Based on the above embodiments, as a preferred embodiment, it further includes:

[0089] A model generation module, configured to obtain manually annotated data; and train based on the manually annotated data to obtain the dual model.

[0090] Based on the above embodiments, as a preferred embodiment, the model generation module is a module for performing the following steps:

[0091] Store the manually annotated data on each distributed node, and in each training iteration, use a distributed training tool to distribute the input data to different distributed nodes;

[0092] Perform forward propagation on each distributed node to calculate the loss function value;

[0093] Use the amp.scale_loss function in the Apex library to scale the loss;

[0094] Perform backward propagation on each distributed node to calculate the gradient;

[0095] Synchronize and average the gradients on all distributed nodes;

[0096] Use an optimizer to update the model parameters;

[0097] Clear the gradient to prepare for the next iteration, and end the training until the number of iterations is reached or convergence occurs, to obtain the dual model.

[0098] Based on the above embodiments, as a preferred embodiment, it further includes:

[0099] A cosine annealing module, configured to adjust the learning rate during the training process of the dual model according to the periodic characteristics of the cosine function.

[0100] Based on the above embodiments, as a preferred embodiment, it further includes:

[0101] An enhanced recognition module, configured to convert the customer service conversation data into a vector representation; segment the vector representation, and calculate the cosine similarity between the vector representation and the knowledge vectors in the index system by using a semantic vector model; select the N pieces of knowledge and conversations with the highest cosine similarity as the input data for the self-developed model; the self-developed model is used to output the conversational implicature of the customer conversation data based on the input data.

[0102] Based on the above embodiments, as a preferred embodiment, it further includes:

[0103] A self-developed model generation module, which is used to construct an external knowledge base based on the vector representation; the external knowledge base is used to store the glossary related to temporary events or short-term dialogue scenarios, as well as the dynamics of the occurrence of the dialogue; the self-developed model is trained based on the external knowledge base.

[0104] The present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed, the steps provided by the above embodiments can be implemented. The storage medium may include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0105] The present application also provides an electronic device. Refer to Figure 3 , the structural diagram of an electronic device provided by an embodiment of the present application, as shown in Figure 3 , may include a processor 1410 and a memory 1420.

[0106] Among them, the processor 1410 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1410 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor 1410 may also include a main processor and a coprocessor. The main processor is a processor used to process data in the wake state, also known as the CPU (Central Processing Unit); the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 1410 may be integrated with a GPU (Graphics Processing Unit), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1410 may also include an AI (Artificial Intelligence) processor, and the AI processor is used to process computational operations related to machine learning.

[0107] The memory 1420 may include one or more computer-readable storage media, which may be non-transitory. The memory 1420 may also include high-speed random access memory and non-volatile memory, such as one or more magnetic disk storage devices and flash storage devices. In this embodiment, the memory 1420 is at least used to store the following computer program 1421. After being loaded and executed by the processor 1410, the computer program can implement the relevant steps in the method executed by the electronic device side disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 1420 may also include an operating system 1422, data 1423, etc., and the storage method may be transient storage or permanent storage. Among them, the operating system 1422 may include Windows, Linux, Android, etc.

[0108] In some embodiments, the electronic device may further include a display screen 1430, an input / output interface 1440, a communication interface 1450, a sensor 1460, a power supply 1470, and a communication bus 1480.

[0109] Of course, Figure 3 The structure of the shown electronic device does not constitute a limitation on the electronic device in the embodiments of the present application. In practical applications, the electronic device may include more or fewer components than Figure 3 those shown, or combine certain components.

[0110] The various embodiments in the specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the system provided by the embodiment, since it corresponds to the method provided by the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0111] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. It should be noted that for those of ordinary skill in the art in the technical field of the present application, without departing from the principle of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the present application.

[0112] It should also be noted that in this specification, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

Claims

1. A customer service public opinion monitoring method, characterized in that: include: Get customer service conversation data; Using the main model in the dual model to understand the conversation content of the customer service conversation data, and determining the conversation scenario that best matches the customer service conversation data in the public opinion definition scenario according to the conversation content; Using the slave model in the dual model to return the hit dialogue according to the dialogue scene output by the master model; Calculating the dialogue understanding quality of a specific dialogue in the dialogue scenario and the response quality of the hit dialogue; If the dialogue understanding quality and the response quality both meet the corresponding set thresholds, the public opinion monitoring result corresponding to the dialogue scenario is output.

2. The customer service public opinion monitoring method according to claim 1 is characterized in that: After obtaining customer service conversation data, it also includes: Regular expressions are used to filter noise information in the customer conversation data; the noise information includes modal particles and empty conversations.

3. The customer service public opinion monitoring method according to claim 1 is characterized in that: Before using the main model in the dual model to understand the conversation content of the customer service conversation data and determining the conversation scenario that best matches the customer service conversation data in the public opinion definition scenario according to the conversation content, the method further includes: Obtain manually annotated data; The dual model is obtained by training based on the manually labeled data.

4. The customer service public opinion monitoring method according to claim 3 is characterized in that: The dual model is obtained by training based on the manually labeled data, including: The manually annotated data is stored on each distributed node, and in each training iteration, the input data is distributed to different distributed nodes using a distributed training tool; Perform forward propagation on each distributed node to calculate the loss function value; Use the amp.scale_loss function in the Apex library to scale the loss; Perform back propagation on each distributed node and calculate the gradient; Synchronize and average the gradients on all distributed nodes; Update model parameters using optimizer; The gradient is cleared to prepare for the next iteration, and the training is terminated when the number of iterations is met or convergence occurs to obtain the dual model.

5. The customer service public opinion monitoring method according to claim 4 is characterized in that: The gradient is cleared to prepare for the next iteration until the training is terminated when the number of iterations is met. When the dual model is obtained, the following is also included: The learning rate during the dual model training process is adjusted according to the periodic characteristics of the cosine function.

6. The customer service public opinion monitoring method according to claim 1 is characterized in that: If any one of the dialogue understanding quality and the response quality does not meet the corresponding set threshold, the method further includes: Switching the customer service conversation data into a vector representation; Segmenting the vector representation and calculating cosine similarity using a semantic vector model and knowledge vectors in an indexing system; N pieces of knowledge and dialogues with the highest cosine similarity are selected as input data for the self-developed model; the self-developed model is used to output the conversational meaning of the customer dialogue data based on the input data.

7. The customer service public opinion monitoring method according to claim 6 is characterized in that: After the customer service conversation data is converted into a vector representation, the method further includes: An external knowledge base is constructed based on the vector representation; the external knowledge base is used to store explanations of terms related to temporary events or short-term dialogue scenarios, as well as dialogue dynamics; The self-developed model is obtained by training based on the external knowledge base.

8. A customer service public opinion monitoring system, characterized in that: include: Data acquisition module, used to obtain customer service conversation data; A bidirectional recognition module, for understanding the conversation content of the customer service conversation data by using the main model in the dual model, and determining the conversation scene that best matches the customer service conversation data in the public opinion definition scene according to the conversation content; and returning the hit conversation according to the conversation scene output by the main model by using the slave model in the dual model; A parameter calculation module, used to calculate the dialogue understanding quality of a specific dialogue in the dialogue scenario and the response quality of the hit dialogue; The public opinion detection module is used to output the public opinion monitoring result corresponding to the dialogue scenario if the dialogue understanding quality and the response quality both meet the corresponding set thresholds.

9. 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.

10. An electronic device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps of the method according to any one of claims 1 to 7 are implemented.

Images