Complaint analysis method and device, equipment, storage medium and program product

Abstract

The invention relates to the technical field of artificial intelligence, and provides a complaint analysis method and device, equipment, a storage medium and a program product. The method comprises the following steps: inputting a complaint dialogue text into a complaint analysis large model to obtain a complaint analysis result output by the complaint analysis large model; the large complaint analysis model is obtained by training an improved generative pre-training model based on a plurality of sample complaint dialogue texts and a service scene label of each sample complaint dialogue text, and the improved generative pre-training model comprises a base network, a gating network and a plurality of adaptation modules. Through the above mode, the model can perform modeling on different business scene features through the plurality of adaptation modules, so that the method has a personalized analysis capability for different business scenes, the multi-scene adaptation capability of the model is enhanced, the complaint analysis result output by the large complaint analysis model has better scene adaptability, and the complaint analysis efficiency is improved. And efficient multi-scene complaint analysis can be realized.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a complaint analysis method, device, equipment, storage medium and program product. Background Art

[0002] The annual increase in the number of network complaints has brought huge challenges to operators. In order to reduce user complaints and improve user satisfaction, it is necessary to handle complaint information in a timely and rapid manner to maintain enterprise-customer relationships and reduce user churn.

[0003] The text of online complaint dialogues contains information such as complaint phenomena, key entities, and fault scenarios that are strongly related to the content of user complaints. Reasonable extraction of this key information is helpful for complaint dispatch and processing. Some online customer service personnel are not proficient in their business and find it difficult to accurately extract complete key complaint information, which can easily have a certain negative impact on subsequent complaint processing. Therefore, cutting-edge artificial intelligence (AI) technology can be introduced to assist online customer service personnel in complaint analysis.

[0004] In some related technologies, for voice service complaint analysis scenarios, most of them use natural language models to process complaint dialogue texts to achieve complaint analysis. However, natural language models have a large demand for training samples, and often require a large number of training samples to train a model with better performance. However, the collection and annotation of training samples are time-consuming and cumbersome, making it difficult to achieve convenient model development and implementation, resulting in high model application costs. In addition, the natural language model is not adaptable enough to scenarios. For scenarios with many business types such as network complaints, the existing natural language models are still unable to achieve refined analysis and processing. The model's scenario adaptation capabilities are limited, making it difficult to achieve complaint analysis in multiple scenarios.

[0005] Therefore, the existing complaint analysis methods lack scenario adaptability and are difficult to implement complaint analysis in multiple scenarios. Summary of the invention

[0006] The embodiments of the present application provide a complaint analysis method, apparatus, device, storage medium and program product to solve the technical problems that the existing complaint analysis methods are insufficient in scene adaptability and difficult to implement complaint analysis in multiple scenes.

[0007] In a first aspect, an embodiment of the present application provides a complaint analysis method, including: obtaining a complaint conversation text; inputting the complaint conversation text into a complaint analysis big model to obtain a complaint analysis result output by the complaint analysis big model; wherein the complaint analysis big model is based on multiple sample complaint conversation texts and business scenario labels of each sample complaint conversation text, and is obtained by training an improved generative pre-trained model, wherein the improved generative pre-trained model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes multiple adaptation modules.

[0008] In one embodiment, the adaptation network also includes a weighting module, a gating network is respectively connected to each adaptation module in the adaptation network, and each adaptation module is respectively connected to the weighting module; a base network is used to perform feature encoding processing on the complaint dialogue text to obtain an intermediate vector; a gating network is used to perform feature extraction processing on the intermediate vector to obtain a probability vector, and based on the probability vector, determine at least one target adaptation module from multiple adaptation modules; each target adaptation module is used to perform feature extraction processing on the probability vector to obtain an output vector; a weighting module is used to perform weighted aggregation operations on each output vector to generate a complaint analysis result.

[0009] In one embodiment, the complaint analysis model is trained using a preset loss function; wherein the preset loss function is determined based on a first loss and a second loss, the first loss is the prediction loss of the improved generative pre-trained model during the training process, and the second loss is the output distribution sparsity loss of the gated network.

[0010] In one embodiment, before the complaint conversation text is input into the complaint analysis big model and the complaint analysis result output by the complaint analysis big model is obtained, it also includes: obtaining sample complaint conversation texts of multiple business scenarios; inputting each sample complaint conversation text into the data sample classification model respectively, and obtaining the business scenario label of each sample complaint conversation text output by the data sample classification model respectively; wherein the data sample classification model is obtained by jointly training a first initial classification model based on the original complaint conversation texts of multiple business scenarios and the labeled complaint conversation texts of multiple business scenarios; the original complaint conversation text is an unlabeled text.

[0011] In one embodiment, before inputting each sample complaint dialogue text into the data sample classification model respectively and obtaining the business scenario label of each sample complaint dialogue text respectively output by the data sample classification model, it also includes: obtaining original complaint dialogue texts of multiple business scenarios and labeled complaint dialogue texts of multiple business scenarios; performing unsupervised training on a first initial classification model based on the multiple original complaint dialogue texts to obtain a second initial classification model; performing supervised classification task transfer learning training on the second initial classification model based on the multiple labeled complaint dialogue texts to obtain a data sample classification model.

[0012] In one embodiment, the data sample classification model uses text reconstruction loss for unsupervised training and uses multi-classification cross entropy loss for supervised classification task transfer learning training.

[0013] In a second aspect, an embodiment of the present application provides a complaint analysis device, comprising: an acquisition module for acquiring a complaint conversation text; a complaint analysis module for inputting the complaint conversation text into a complaint analysis big model to obtain a complaint analysis result output by the complaint analysis big model; wherein the complaint analysis big model is obtained by training an improved generative pre-trained model based on multiple sample complaint conversation texts and business scenario labels of each sample complaint conversation text, the improved generative pre-trained model comprises a base network, a gating network and an adaptation network connected in sequence, and the adaptation network comprises multiple adaptation modules.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, any one of the complaint analysis methods described above is implemented.

[0015] In a fourth aspect, an embodiment of the present application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the complaint analysis methods described above.

[0016] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements any of the above-mentioned complaint analysis methods when executed by a processor.

[0017] The complaint analysis method, apparatus, equipment, storage medium and program product provided in the embodiments of the present application improve the existing generative pre-trained model, and use multiple sample complaint dialogue texts and the business scenario label of each sample complaint dialogue text to train the improved generative pre-trained model to obtain a large complaint analysis model. Since the complaint analysis large model adds an adaptation network including multiple adaptation modules to the network structure of the existing generative pre-trained model, the model can model different business scenario characteristics through multiple adaptation modules, thereby having personalized analysis capabilities for different business scenarios, enhancing the multi-scenario adaptation capability of the model, and thus making the complaint analysis results output by the complaint analysis large model have better scenario adaptability, thereby realizing efficient multi-scenario complaint analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present application or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the complaint analysis method provided in the embodiment of the present application.

[0020] Figure 2 This is the second flow chart of the complaint analysis method provided in the embodiment of the present application.

[0021] Figure 3 It is a schematic diagram of the network structure of the complaint analysis model provided in the embodiment of the present application.

[0022] Figure 4 It is a network structure diagram of the data sample classification model provided in the embodiment of the present application.

[0023] Figure 5 It is a structural diagram of the complaint analysis device provided in an embodiment of the present application.

[0024] Figure 6 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0026] See also Figure 1 , Figure 1 This is one of the flowcharts of the complaint analysis method provided in the embodiment of the present application. In the embodiment of the present application, the complaint analysis method includes steps S110 to S120, and each step is as follows: S110: Obtain the complaint dialogue text.

[0027] See also Figure 2 , Figure 2 This is the second flow chart of the complaint analysis method provided in the embodiment of the present application.

[0028] like Figure 2 As shown, in the network complaint analysis scenario, the user can call the complaint hotline. After the original complaint recording of the complaint hotline is processed through voice transcription, the complaint dialogue text can be obtained. The complaint dialogue text is a natural language text.

[0029] S120: Input the complaint dialogue text into the complaint analysis big model to obtain the complaint analysis results output by the complaint analysis big model.

[0030] Among them, the complaint analysis model is obtained by training an improved generative pre-trained model based on multiple sample complaint dialogue texts and the business scenario label of each sample complaint dialogue text. The improved generative pre-trained model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes multiple adaptation modules.

[0031] Specifically, the complaint dialogue text is input into the complaint analysis model (i.e. Figure 2 The complaint analysis model can analyze and process the complaint dialogue text and output the complaint analysis results corresponding to the complaint dialogue text.

[0032] like Figure 2 As shown, the complaint analysis results corresponding to the complaint dialogue text include the user's key complaint intentions, repeated complaint warning information, negative sentiment analysis, etc.

[0033] Optionally, before inputting the complaint dialogue text into the complaint analysis model, data preprocessing is performed on the complaint dialogue text to ensure the standardization of the model input data.

[0034] It should be noted that the complaint analysis model of this embodiment is obtained by training the improved generative pre-training model based on multiple sample complaint dialogue texts in different business scenarios and the business scenario labels of each sample complaint dialogue text. The improved generative pre-training model is a large language model. The large language model has amazing capabilities in the field of text understanding and generation, which makes it possible to be applied in the network complaint analysis scenario. Using the large language model in the network complaint analysis scenario can realize tasks such as complaint business classification and key information extraction, realize automatic processing of the entire complaint process, and help reduce customer service labor costs and improve complaint handling efficiency.

[0035] Optionally, the model side designs prompt words for specific business scenarios and uses the base model for transfer learning, which can achieve model adaptation and optimization for different business scenarios and obtain a large complaint analysis model.

[0036] Optionally, field knowledge in the fields of processing standards for online complaints, complaint interpretation, etc. can be introduced to provide professional background knowledge for the complaint analysis model, thereby helping to improve the complaint analysis effect of the complaint analysis model.

[0037] Optionally, the complaint analysis model can be deployed on a comprehensive online complaint processing platform, access the real-time data of the complaint hotline, automatically perform real-time analysis on complaint scenarios such as voice, Internet of Things, 4G, 5G, etc., obtain information such as user complaint intentions and negative emotions, and assist customer service personnel in quickly handling complaints.

[0038] The complaint analysis method provided in the embodiment of the present application improves the existing generative pre-trained model, and uses multiple sample complaint dialogue texts and the business scenario label of each sample complaint dialogue text to train the improved generative pre-trained model to obtain a complaint analysis large model. Since the complaint analysis large model adds an adaptation network including multiple adaptation modules to the network structure of the existing generative pre-trained model, the model can model different business scenario characteristics through multiple adaptation modules, thereby having personalized analysis capabilities for different business scenarios, enhancing the model's multi-scenario adaptation capabilities, and thereby making the complaint analysis results output by the complaint analysis large model have better scenario adaptability, thereby realizing efficient multi-scenario complaint analysis.

[0039] In some embodiments, the adaptation network also includes a weighting module, the gating network is respectively connected to each adaptation module in the adaptation network, and each adaptation module is respectively connected to the weighting module; the base network is used to perform feature encoding processing on the complaint dialogue text to obtain an intermediate vector; the gating network is used to perform feature extraction processing on the intermediate vector to obtain a probability vector, and based on the probability vector, determine at least one target adaptation module from multiple adaptation modules; each target adaptation module is used to perform feature extraction processing on the probability vector to obtain an output vector; the weighting module is used to perform weighted aggregation operations on each output vector to generate a complaint analysis result.

[0040] Network complaint analysis involves multiple business scenarios such as voice, IoT, 4G, 5G, and home broadband. Each business scenario has unique complaint phenomena and content, and different complaint handling methods are required for different business scenarios. The existing generative pre-trained model can be fine-tuned in a single business scenario to obtain better complaint analysis results, but in multiple business scenarios, mutual interference may occur, resulting in poor analysis results. In view of the characteristics of multiple business scenarios of network complaints, it is necessary to improve the network structure of the existing generative pre-trained model to obtain a complaint analysis model with stronger complaint analysis capabilities to adapt to the actual needs of multiple business scenarios.

[0041] The complaint analysis model of this embodiment is obtained by improving the structure of the existing generative pre-trained model. The improvement idea is to add a set of customized adaptation modules on the basis of the existing model (i.e., the base model).

[0042] See also Figure 3 , Figure 3 It is a schematic diagram of the network structure of the complaint analysis model provided in the embodiment of the present application.

[0043] like Figure 3 As shown, the complaint analysis model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes a weighted module and multiple adaptation modules (i.e. Figure 3 The top multiple block components, namely the top block _a to block _n), the gating network is respectively connected to each adaptation module in the adaptation network, and each adaptation module is respectively connected to the weighting module.

[0044] Base Network ( Figure 3 The bottom block component, i.e., the bottom block _a), is used to perform feature encoding processing on the complaint dialogue text to obtain an intermediate vector.

[0045] Among them, the block component is the decoder part of the Transformer model and is the basic unit of the large language model.

[0046] Gating Networks ( Figure 3 The Router in the example (i.e., the routing module) is used to perform feature extraction processing on the intermediate vector to obtain a probability vector, and based on the probability vector, determine at least one target adaptation module from multiple adaptation modules.

[0047] Specifically, the routing module is a gating network composed of a feedforward neural network. It can perform nonlinear transformation on the upstream input data, that is, the intermediate vector, and generate and output a probability vector after being processed by the Softmax function. A probability vector includes multiple probability values. The number of probability values ​​is consistent with the number of adaptation modules in the adaptation network. A probability value can represent the importance of an adaptation module in the adaptation network. The greater the importance of the adaptation module, the greater the probability that the adaptation module is selected as the target adaptation module. The model can select the adaptation module with the largest probability value from multiple adaptation modules as the target adaptation module according to the probability vector output by the gating network, automatically activate the target adaptation module in the adaptation network, and input the probability vector output by the gating network into each target adaptation module respectively, so as to realize the personalized processing of the model for specific scenario data and improve the overall efficiency and performance.

[0048] Among them, the expression of the gating network is as follows: ; ; ; Where, T is the temperature parameter; , and are the three weight parameters in the feedforward network; b and c are bias constants; is the parameter of the current network layer; is the parameter of the previous network layer; n is the total number of network layers; f(x) is the activation function; For input data; is the value of the i-th dimension of the input data; is the value of the j-th dimension of the input data.

[0049] The temperature parameter T is used to control the sparsity of the probability value output by the gating network to ensure that the probability value is as sparse as possible and activate fewer target adaptation modules. If the gating network activates more target adaptation modules in a single selection, although it may improve the performance of the model, it will cause more target adaptation modules to participate in the calculation, bringing additional computational complexity and time consumption. Therefore, it is necessary to set a reasonable temperature parameter according to specific application requirements and computing resource limitations to make the gating network have reasonable sparsity.

[0050] Each target adaptation module is used to perform feature extraction processing on the probability vector to obtain an output vector; the weighting module is used to perform weighted aggregation operation on each output vector to generate a complaint analysis result.

[0051] Specifically, there may be one or more target adaptation modules. Each target adaptation module can be regarded as a sub-model that can model the input data characteristics of different business scenarios. It has the customization capability for specific business scenarios. The adaptation network expands the personalization capability of the model and gives the model stronger business scenario adaptability.

[0052] If there are multiple target adaptation modules, the output vector generated by each target adaptation module can be input into the weighting module, and the weighting module can perform weighted aggregation operations on multiple output vectors to generate complaint analysis results.

[0053] If there is only one target adaptation module, the output vector generated by the target adaptation module can be input into the weighting module, and the weighting module can directly process the output vector to generate a complaint analysis result.

[0054] The complaint analysis method provided in the embodiment of the present application performs network complaint analysis based on large language model technology, which can improve the automation and intelligence of complaint handling. In view of the numerous network complaint scenarios, a large complaint analysis model is constructed based on a multi-business scenario adaptation mechanism, multiple adaptation modules are added and a gated network is used for selection control, which can realize specific analysis of specific scenarios and the integration of multiple complaint scenarios.

[0055] In some embodiments, the complaint analysis model is trained using a preset loss function; wherein the preset loss function is determined based on a first loss and a second loss, the first loss is the prediction loss of the improved generative pre-trained model during the training process, and the second loss is the output distribution sparsity loss of the gated network.

[0056] Understandably, during the training of the large complaint analysis model, several specific adaptation modules may be selected by the gating network at high frequency, while other adaptation modules are not selected, which will cause some adaptation modules to be deeply trained, while other adaptation modules are not trained enough. As the training process progresses, the adaptation modules selected by the gating network at high frequency become more and more powerful, which will increase their probability of being selected again, causing the overall model to lose the ability to customize the output of the scene. To avoid this situation, a penalty loss term can be added to the original loss function to control the sparsity of the output distribution of the gating network, avoiding the problem of over-training of some adaptation modules and under-training of other adaptation modules.

[0057] Specifically, the complaint analysis model is trained using a preset loss function, which can be determined based on a first loss and a second loss. The first loss is the prediction loss of the improved generative pre-training model during the training process, and the second loss is the output distribution sparsity loss of the gated network.

[0058] In this embodiment, the output distribution sparsity loss of the gated network can be represented by the gated network loss function, and the gated network loss function is used as a penalty loss term. The main design idea of ​​the gated network loss function is to statistically select the selection results of each sample by the gated network, so as to make it select different adaptation modules as evenly as possible to fully train the entire adaptation network.

[0059] Among them, the gated network loss function The calculation formula is as follows: ; ; Among them, vector I is a vector used to measure the distribution of the adaptation modules selected by the gating network in the training samples of multiple training batches (for example, k training batches are used for model training, and each training batch uses multiple training samples). The smaller the standard deviation, the more uniform the distribution. The function of vector I is to count the discrete type of the sparse vector y; x is the probability vector output by the gating network, and a probability vector includes multiple probability values; is the i-th probability value in the probability vector; is the sparse vector after the i-th probability value is transformed (the vector index value where 1 is located represents the position of the selected adaptation module). The sparse vectors after all probability values ​​of a probability vector x are transformed can be represented as a sparse vector y; Y is a set of sparse vectors y in multiple training batches; , All are hyperparameters; represents the standard deviation of vector I; represents the mean of vector I; is a counting function.

[0060] It should be noted that during the model training process, a large number of training samples are required for multi-batch training. Multiple training batches can ensure that the model can learn enough samples to ensure the distribution credibility. This embodiment introduces an exponential function in the gated network loss function to enhance the penalty for distribution concentration. It should be noted that during the training process, the Transformer model block component performs a gradient update in each training batch, while the gated network performs a gradient update every k training batches. At this time, the preset loss function The calculation formula can be expressed as: ; in, is the first loss, i.e., the loss function of the existing generative pre-training model; C is the sample set in the current training batch; k is the update interval training batch number of the gated network. After multiple rounds of training iterations, the gated network can output a relatively sparse probability distribution, fully train each adaptation module, and give the complaint analysis large model the ability to analyze multiple scenarios; The second loss.

[0061] The complaint analysis method provided in the embodiment of the present application introduces the sparsity loss of the output distribution of the gated network as a penalty loss term on the basis of the existing loss function to avoid the situation where the overall performance of the model is poor due to the uneven training of different adaptation modules. The sparsity loss is used to control the sparsity of the output distribution of the gated network, improve the convergence speed and generalization of model training, and ensure the model performance of the large complaint analysis model to ensure its ability to adapt to multiple scenarios.

[0062] In some embodiments, before the complaint conversation text is input into the complaint analysis big model and the complaint analysis results output by the complaint analysis big model are obtained, it also includes: obtaining sample complaint conversation texts of multiple business scenarios; inputting each sample complaint conversation text into the data sample classification model respectively, and obtaining the business scenario label of each sample complaint conversation text output by the data sample classification model respectively; wherein the data sample classification model is based on the original complaint conversation texts of multiple business scenarios and the labeled complaint conversation texts of multiple business scenarios, and a first initial classification model is jointly trained; the original complaint conversation text is an unlabeled text.

[0063] A large number of recordings collected by hotline complaints can be converted into complaint dialogue texts after voice recording. The complaint dialogue texts contain data samples of multiple business scenarios such as voice calls, text messages, 4G, 5G, traffic, and home broadband. In order to achieve multi-scenario adaptation of the model, the collected historical complaint dialogue text data can be classified according to the business scenario type to meet the needs of subsequent complaint analysis large model training. This embodiment uses a data sample classification model to achieve the classification of complaint samples.

[0064] See also Figure 4 , Figure 4 It is a network structure diagram of the data sample classification model provided in the embodiment of the present application.

[0065] like Figure 4 As shown, the basic unit of the data sample classification model is the Transformer encoder (i.e. Figure 4The data sample classification model combines the advantages of sequence reconstruction and classification. It uses the encoder and decoder (Trans-AE) to calculate the loss based on text reconstruction, and the sequence model (Trans-Encoder) to calculate the error based on prediction. The combination of the two can provide a more comprehensive sample classification effect and accurately identify the business scenario type to which the complaint dialogue text belongs.

[0066] The main process of the text sequence reconstruction task is: first, the sample complaint dialogue texts of multiple business scenarios are input into the data sample classification model for vector serialization representation; each sample complaint dialogue text (i.e. Figure 4 x in the image) can be input into the encoder component (Encoder) for encoding to obtain an intermediate vector; then the intermediate vector (i.e. Figure 4 z in the decoder to get the decoded vector x 1 ; Finally calculate the decoded vector x 1 The error between the original input x and the original input x is used to obtain the text reconstruction loss value and update the encoder and decoder parameters. The classification task is further trained on the basis of the encoder based on the transfer learning method. A layer of feedforward neural network (FNN) can be added after the encoder to adapt to the classification task and convert the intermediate vector output by the encoder into the probability distribution of the business scenario category.

[0067] After completing the training of the data sample classification model, for each sample complaint dialogue text of an unlabeled business scenario type, the sample complaint dialogue text of the unlabeled business scenario type can be input into the trained data sample classification model to obtain the business scenario label of the sample complaint dialogue text of the unlabeled business scenario type output by the data sample classification model.

[0068] In some embodiments, before inputting each sample complaint conversation text into the data sample classification model respectively and obtaining the business scenario label of each sample complaint conversation text output by the data sample classification model respectively, it also includes: obtaining original complaint conversation texts of multiple business scenarios and labeled complaint conversation texts of multiple business scenarios; performing unsupervised training on a first initial classification model based on the multiple original complaint conversation texts to obtain a second initial classification model; performing supervised classification task transfer learning training on the second initial classification model based on the multiple labeled complaint conversation texts to obtain a data sample classification model.

[0069] Specifically, in the training process of the data sample classification model, it is necessary to combine the two tasks of text reconstruction and text classification for joint training: first obtain the original complaint dialogue texts of multiple business scenarios and the labeled complaint dialogue texts of multiple business scenarios. The original complaint dialogue texts are unlabeled texts.

[0070] Furthermore, based on multiple original complaint dialogue texts, the first initial classification model is trained unsupervised to obtain a second initial classification model.

[0071] During the unsupervised training process, the model can calculate the text reconstruction loss and update some parameters of the encoder and decoder based on the text reconstruction loss. The text reconstruction task uses an unsupervised training method, which does not require labeled data and can use a large amount of unlabeled data to adjust the encoder module parameters, effectively reducing the training cost of the model.

[0072] Furthermore, based on multiple annotated complaint dialogue texts, supervised classification task transfer learning training is performed on the second initial classification model to obtain a data sample classification model.

[0073] Understandably, unsupervised training alone cannot guarantee the classification accuracy of the model. Therefore, supervised training can be used to fine-tune the parameters of the unsupervised trained model to improve the model's directional analysis capabilities for different business scenarios.

[0074] Specifically, based on a small amount of annotated complaint dialogue text, a supervised algorithm is used to perform classification task transfer learning on the second initial classification model to obtain the final data sample classification model; during the supervised training process, the model can calculate the multi-classification cross entropy loss, and update the encoder and the feedforward neural network responsible for the classification task based on the multi-classification cross entropy loss, thereby achieving efficient joint training that combines the advantages of supervised and unsupervised methods. The joint training mode can combine the advantages of both unsupervised and supervised methods, overcome the problem of difficulty in obtaining a large amount of annotated data, and achieve ideal model classification effects.

[0075] The complaint analysis method provided in the embodiment of the present application adopts a complaint sample classification method based on a sample reconstruction mechanism, uses a large amount of original complaint dialogue texts for unsupervised pre-training, uses a supervised learning algorithm for classification task transfer learning, and combines the advantages of unsupervised learning methods and supervised learning methods to perform efficient model training. It can achieve efficient text data classification under the condition of fewer labeled samples, ensuring that the data sample classification model has a good text classification effect.

[0076] In some embodiments, the data sample classification model uses text reconstruction loss for unsupervised training and uses multi-classification cross entropy loss for supervised classification task transfer learning training.

[0077] The present application also provides a complaint analysis device. Figure 5 , Figure 5 5 is a schematic diagram of the structure of the complaint analysis device provided in the embodiment of the present application. In the embodiment of the present application, the complaint analysis device includes an acquisition module 510 and a complaint analysis module 520.

[0078] The acquisition module 510 is used to acquire the complaint dialogue text.

[0079] The complaint analysis module 520 is used to input the complaint dialogue text into the complaint analysis big model to obtain the complaint analysis results output by the complaint analysis big model.

[0080] Among them, the complaint analysis model is obtained by training an improved generative pre-trained model based on multiple sample complaint dialogue texts and the business scenario label of each sample complaint dialogue text. The improved generative pre-trained model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes multiple adaptation modules.

[0081] In some embodiments, the adaptation network also includes a weighting module, the gating network is respectively connected to each adaptation module in the adaptation network, and each adaptation module is respectively connected to the weighting module; the base network is used to perform feature encoding processing on the complaint dialogue text to obtain an intermediate vector; the gating network is used to perform feature extraction processing on the intermediate vector to obtain a probability vector, and based on the probability vector, determine at least one target adaptation module from multiple adaptation modules; each target adaptation module is used to perform feature extraction processing on the probability vector to obtain an output vector; the weighting module is used to perform weighted aggregation operations on each output vector to generate a complaint analysis result.

[0082] In some embodiments, the complaint analysis model is trained using a preset loss function; wherein the preset loss function is determined based on a first loss and a second loss, the first loss is the prediction loss of the improved generative pre-trained model during the training process, and the second loss is the output distribution sparsity loss of the gated network.

[0083] In some embodiments, the complaint analysis device further includes a sample classification module.

[0084] A sample classification module is used to obtain sample complaint dialogue texts of multiple business scenarios; each sample complaint dialogue text is input into a data sample classification model to obtain a business scenario label for each sample complaint dialogue text output by the data sample classification model; wherein the data sample classification model is obtained by jointly training a first initial classification model based on original complaint dialogue texts of multiple business scenarios and labeled complaint dialogue texts of multiple business scenarios; the original complaint dialogue text is an unlabeled text.

[0085] In some embodiments, a sample classification module is used to obtain original complaint conversation texts of multiple business scenarios and labeled complaint conversation texts of multiple business scenarios; based on the multiple original complaint conversation texts, unsupervised training is performed on a first initial classification model to obtain a second initial classification model; based on the multiple labeled complaint conversation texts, supervised classification task transfer learning training is performed on the second initial classification model to obtain a data sample classification model.

[0086] In some embodiments, the data sample classification model uses text reconstruction loss for unsupervised training and uses multi-classification cross entropy loss for supervised classification task transfer learning training.

[0087] The present application also provides an electronic device, Figure 6 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, such as Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the complaint analysis method.

[0088] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0089] An embodiment of the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the complaint analysis method provided by the above methods is implemented.

[0090] An embodiment of the present application also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the complaint analysis methods provided by the above methods.

[0091] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0092] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A complaint analysis method, characterized in that: include: Get the complaint dialogue text; Inputting the complaint dialogue text into the complaint analysis big model to obtain the complaint analysis result output by the complaint analysis big model; Among them, the complaint analysis model is obtained by training an improved generative pre-trained model based on multiple sample complaint dialogue texts and business scenario labels of each sample complaint dialogue text. The improved generative pre-trained model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes multiple adaptation modules.

2. The complaint analysis method according to claim 1, characterized in that: The adaptation network further includes a weighting module, the gating network is respectively connected to each of the adaptation modules in the adaptation network, and each of the adaptation modules is respectively connected to the weighting module; The base network is used to perform feature encoding processing on the complaint dialogue text to obtain an intermediate vector; The gating network is used to perform feature extraction processing on the intermediate vector to obtain a probability vector, and determine at least one target adaptation module from the plurality of adaptation modules based on the probability vector; Each of the target adaptation modules is used to perform feature extraction processing on the probability vector to obtain an output vector; The weighting module is used to perform weighted aggregation operation on each of the output vectors to generate the complaint analysis result.

3. The complaint analysis method according to claim 1, characterized in that: The complaint analysis model is trained using a preset loss function; Among them, the preset loss function is determined based on a first loss and a second loss, the first loss is the prediction loss of the improved generative pre-training model during the training process, and the second loss is the output distribution sparsity loss of the gating network.

4. The complaint analysis method according to claim 1, characterized in that: Before inputting the complaint dialogue text into the complaint analysis model and obtaining the complaint analysis result output by the complaint analysis model, the method further includes: Obtain sample complaint dialogue texts for multiple business scenarios; Input each of the sample complaint dialogue texts into a data sample classification model, respectively, to obtain a business scenario label for each of the sample complaint dialogue texts output by the data sample classification model; Among them, the data sample classification model is obtained by jointly training a first initial classification model based on original complaint dialogue texts of multiple business scenarios and labeled complaint dialogue texts of multiple business scenarios; the original complaint dialogue text is an unlabeled text.

5. The complaint analysis method according to claim 4, characterized in that: Before inputting each sample complaint dialogue text into the data sample classification model and obtaining the business scenario label of each sample complaint dialogue text output by the data sample classification model, the method further includes: Obtaining the original complaint dialogue texts of multiple business scenarios and the annotated complaint dialogue texts of multiple business scenarios; Based on the multiple original complaint dialogue texts, performing unsupervised training on the first initial classification model to obtain a second initial classification model; Based on the multiple labeled complaint dialogue texts, supervised classification task transfer learning training is performed on the second initial classification model to obtain the data sample classification model.

6. The complaint analysis method according to claim 5, characterized in that: The data sample classification model uses text reconstruction loss for unsupervised training and uses multi-classification cross entropy loss for supervised classification task transfer learning training.

7. A complaint analysis device, characterized in that: include: The acquisition module is used to obtain the complaint dialogue text; A complaint analysis module, used to input the complaint dialogue text into the complaint analysis model, and obtain the complaint analysis results output by the complaint analysis model; Among them, the complaint analysis model is obtained by training an improved generative pre-trained model based on multiple sample complaint dialogue texts and business scenario labels of each sample complaint dialogue text. The improved generative pre-trained model includes a base network, a gating network and an adaptation network connected in sequence, and the adaptation network includes multiple adaptation modules.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the complaint analysis method according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the complaint analysis method according to any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the complaint analysis method according to any one of claims 1 to 6 is implemented.

Images