Complaint data response generation method and device, equipment, medium and product
By employing multimodal data processing and intelligent response generation methods, the problem of low efficiency in manual complaint handling has been solved, achieving efficient and accurate complaint responses and improving user satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA MOBILE GRP HAINAN CO LTD
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-12
AI Technical Summary
The current technology for manually handling user complaints is inefficient, resulting in untimely responses to user complaints and affecting user satisfaction.
By acquiring multimodal complaint data, using a semantic feature embedding model to extract and fuse features, and combining intent recognition and sentiment recognition models, a response strategy is generated and a response text is generated based on a complaint processing knowledge base. This integrates complaint data from different modalities, collaboratively drives natural language processing and sentiment recognition, and optimizes human resource costs and response consistency.
It significantly improves complaint handling efficiency, generates accurate response texts, optimizes human resource costs and response consistency, and enhances user satisfaction.
Smart Images
Figure CN122197882A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to methods, apparatus, equipment, media and products for generating complaint data responses. Background Technology
[0002] In the mobile communications industry, timely and accurate responses to user complaints can effectively improve user satisfaction. Currently, complaint handling is primarily manual, involving human intervention such as receiving user calls and reviewing text messages. This manual approach not only introduces significant subjectivity and inconsistency due to individual differences, affecting the objectivity and fairness of the overall judgment, but also requires considerable processing time. In the mobile communications industry with its large user base, the low efficiency of manual complaint handling leads to delayed responses and consequently, decreased user satisfaction. Summary of the Invention
[0003] This application provides a method, apparatus, equipment, medium, and product for generating complaint data responses, in order to address the shortcomings of low efficiency in manual handling of user complaints in the prior art and to improve the efficiency of complaint handling.
[0004] This application provides a method for generating complaint data responses, including: Complaint data is acquired, which includes data from multiple modalities. The complaint data is input into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features for each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features. The semantic embedding features are input into the intent recognition model and the emotion recognition model respectively, and the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model are obtained. Based on the intent recognition result, the emotion recognition result, and the response strategy determination model, a response strategy is generated. Based on the response strategy and the complaint handling knowledge base, a response text is generated for the complaint data.
[0005] According to a complaint data response generation method provided in this application, the step of generating a response strategy based on the intent recognition result and the emotion recognition result includes: Based on the intent recognition result, a search is performed in the complaint processing knowledge base to obtain candidate knowledge fragments; The intent recognition result, the emotion recognition result, the candidate knowledge fragment, and the context data of the complaint data are input into the response strategy determination model to obtain the response strategy output by the response strategy determination model.
[0006] According to the complaint data response generation method provided in this application, the complaint data includes non-text modal data and text modal data; the step of fusing the initial features of each modality to obtain semantic embedding features includes: The initial features of each of the non-text modal data are respectively input to the semantic mapping layer of the corresponding modality to obtain the mapping features output by each of the semantic mapping layers. The semantic mapping layer is used to map the initial features of the non-text modal data to the same semantic space as the initial features of the text modal data. The semantic chimerism features are obtained by fusing each of the mapping features with the initial features of the text modality data.
[0007] According to a complaint data response generation method provided in this application, the step of fusing each of the mapping features with the initial features of the text modality data to obtain the semantic embedding features includes: Each of the mapping features and the initial features of the text modality data are input to the context alignment module. The context alignment module is used to align the features of multiple modalities in the time dimension and obtain the alignment features of each modality output by the context alignment module. The alignment features of each modality are fused to obtain the semantic chimerism features.
[0008] According to a complaint data response generation method provided in this application, after generating the response text for the complaint data, the method includes: Obtain the quality data corresponding to the response text, wherein the quality data includes at least the user's feedback data; The semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model are updated based on the quality data.
[0009] According to a complaint data response generation method provided in this application, the step of updating the semantic feature embedding model, the intent recognition model, the sentiment recognition model, and the response strategy determination model based on the quality data includes: Based on the quality data, the reward values of the models are determined for the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy, respectively. Based on the reward values corresponding to the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model, reinforcement learning is used to update the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model.
[0010] This application also provides a complaint data response generation device, including: A semantic embedding module is used to acquire complaint data, which includes data from multiple modalities. The complaint data is input into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features for each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features. The intent and emotion recognition module is used to input the semantic embedding features into the intent recognition model and the emotion recognition model respectively, and obtain the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model. The response generation module is used to generate a response strategy based on the intent recognition result, the emotion recognition result, and the response strategy determination model, and to generate a response text for the complaint data based on the response strategy and the complaint handling knowledge base.
[0011] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the complaint data response generation method as described above.
[0012] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the complaint data response generation method as described above.
[0013] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the complaint data response generation method as described above.
[0014] The complaint data response generation method, apparatus, equipment, medium, and product provided in this application acquire multimodal complaint data, extract features from each modality to obtain initial features for each modality, then fuse these initial features to obtain semantic embedding features. These semantic embedding features are then input into an intent recognition model and a sentiment recognition model to obtain intent recognition results and sentiment recognition results. Based on the intent recognition results, sentiment recognition results, and a response strategy, a model is determined, and a response strategy is generated. Finally, based on the response strategy and a complaint processing knowledge base, a response text is generated for the complaint data. This integrates complaint data from different modalities, collaboratively driving natural language processing, sentiment recognition, and intent recognition to achieve in-depth mining and multi-faceted interpretation of complaint intent, thereby generating accurate response text, significantly optimizing labor costs and response consistency, and improving complaint processing efficiency. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart illustrating the complaint data response generation method provided in this application.
[0017] Figure 2 This is a schematic diagram of the implementation process framework of the complaint data response generation method provided in this application.
[0018] Figure 3 This is a schematic diagram of the complaint data response generation device provided in this application.
[0019] Figure 4 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0022] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0023] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0024] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."
[0025] The following is combined with Figure 1 Describe the complaint data response generation method provided in this application. For example... Figure 1 As shown, the method for generating complaint data responses includes the following steps: S110. Obtain complaint data, which includes data from multiple modalities. Input the complaint data into the semantic feature embedding model to obtain the semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain the initial features of each modality. The feature fusion layer is used to map the initial features of each modality to the same semantic space and fuse them to obtain the semantic embedding features. S120. Input the semantic embedding features into the intent recognition model and the emotion recognition model respectively, and obtain the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model. S130. Based on the intent recognition results, emotion recognition results, and response strategy determination model, a response strategy is generated. Based on the response strategy and the complaint handling knowledge base, a response text is generated for the complaint data.
[0026] The complaint data response generation method provided in this application acquires multimodal complaint data, extracts features from each modality to obtain initial features for each modality, and then fuses these initial features to obtain semantic embedding features. These semantic embedding features are then input into an intent recognition model and a sentiment recognition model to obtain intent recognition results and sentiment recognition results. Based on the intent recognition results, sentiment recognition results, and a response strategy, a model is determined, and a response strategy is generated. Finally, based on the response strategy and a complaint processing knowledge base, a response text is generated for the complaint data. This method integrates complaint data from different modalities, collaboratively driving natural language processing, sentiment recognition, and intent recognition to achieve in-depth mining and multi-faceted interpretation of complaint intent, thereby generating accurate response text, significantly optimizing manpower costs and response consistency, and improving complaint processing efficiency.
[0027] In the method provided in this application, the complaint data is the data provided by the user when making a complaint. The complaint data includes data in multiple modalities, and the modalities in the complaint data include at least a text modal and a non-text modal, such as images, voice, etc.
[0028] In one possible implementation, a complaint data collection page can be provided, such as a complaint handling page for intelligent customer service, where user input is received as complaint data through input boxes. In another possible implementation, the complaint data can be obtained by aggregating data from multiple complaint channels; for example, voice data can be collected from communication data received through the complaint number, and text data can be collected from SMS messages received through the complaint number.
[0029] After obtaining the complaint data, one possible approach is to preprocess the complaint data. Preprocessing can be targeted based on the modality in the complaint data. For example, for image data, normalization and noise reduction can be performed; for voice data, frame segmentation can be performed; and for text data, word segmentation and word vector mapping can be performed.
[0030] Complaint data is input into a semantic feature embedding model, which embeds data from different modalities into the same semantic space. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer extracts features from the data of each modality in the complaint data to obtain the initial features of each modality. The feature fusion layer fuses the initial features of each modality to obtain semantic embedding features.
[0031] Specifically, the feature extraction layer includes initial feature extraction modules corresponding to each modality. Based on the modality of the data in the complaint data, the corresponding data is input into the initial feature extraction module of the corresponding modality for processing to extract the initial features of different modalities. The initial feature extraction modules for image, voice, and text modalities are described below. It is worth noting that not all embodiments need to include initial feature extraction modules for these three modalities. Instead, the corresponding initial feature extraction module is called according to the modality included in the complaint data for appropriate processing.
[0032] The initial feature extraction modules corresponding to different modalities can adopt the structure of existing feature extraction models. Since images are a modality with strong spatial structure, Convolutional Neural Networks (CNNs) have significant advantages in image feature extraction. In one possible implementation, the initial feature extraction module corresponding to the image modality can be based on the ResNet-50 model architecture. ResNet (Residual Network) effectively alleviates the gradient vanishing problem during the training of deep networks by introducing "residual blocks." ResNet-50 is a 50-layer residual network that performs well in tasks such as image recognition and feature extraction. Traditional convolutional layers in deep networks easily lead to gradient vanishing, making it difficult to train deeper neural networks. ResNet introduces "residual connections," or skip connections, making the network easier to optimize.
[0033] The basic structure of the residual element is as follows: ; In the above formula, x is the input feature map. y represents the nonlinear transformation result after multiple convolutional layers, and y is the output feature map.
[0034] ResNet-50 consists of an initial convolutional layer, multiple residual modules (four stages in total), a global average pooling layer, and a fully connected layer. Each residual module contains multiple convolutional layers with batch normalization and ReLU activation functions. In the feature extraction stage, in some embodiments of the method provided in this application, the final fully connected layer can be removed, retaining the global average pooling layer to extract a fixed-dimensional image representation vector as the initial feature of the image modality. ; In the above formula, I represents the input image. ∈ These are the initial features of the output image modality.
[0035] The initial feature extraction module for the speech modality can also adopt the architecture of existing speech feature extraction models. Speech signals exhibit strong temporal sequence and local variation characteristics. Long Short-Term Memory (LSTM) networks can effectively model their temporal dependencies. In one possible implementation, an LSTM+Attention model architecture can be used as the initial feature extraction module for the speech modality. LSTM is an improved structure of RNN, introducing memory units (cell states) and three gating mechanisms (input gate, forget gate, and output gate) to alleviate the gradient vanishing problem. The attention mechanism further enhances the model's ability to focus on information at crucial moments, thereby improving the quality of feature representation.
[0036] The initial feature extraction module for the text modality can adopt the architecture of existing text feature extraction models. Text data also exhibits temporal sequence, and information from both preceding and following context is crucial for the semantic understanding of the current word. In one possible implementation, the initial feature extraction module for the text modality can employ a BiLSTM (Bidirectional Long Short-Term Memory) model architecture. BiLSTM effectively captures contextual information by simultaneously incorporating forward and backward LSTMs.
[0037] In one possible implementation of the method provided in this application, when fusing the initial features of each modality, the initial features of each modality can be directly concatenated and then nonlinearly transformed to generate cross-modal semantic fusion features.
[0038] Because the information carried by different modalities is heterogeneous and has different distributions, another possible implementation of the method provided in this application involves fusing the initial features of each modality to obtain semantically fused features, including: The initial features of each non-text modal data are respectively input into the semantic mapping layer of the corresponding modality to obtain the mapping features output by each semantic mapping layer. The semantic mapping layer is used to map the initial features of non-text modal data to the same semantic space as the initial features of text modal data. The semantic chimerism features are obtained by fusing the various mapping features with the initial features of the text modality data.
[0039] In this implementation, instead of directly fusing the initial features of each modality, the initial features of different modalities are mapped to the same semantic space before being fused. This generates a unified cross-modal semantic chimera vector, providing a higher quality joint representation for downstream tasks.
[0040] In some embodiments, mapping the initial features of the image modality to the same semantic space as the initial features of the text modality data can be achieved through a fully connected layer, with the mapping formula as follows: ; in, The initial features of the input image modality, and These are the parameters of the convolutional layer. and These are the parameters for the fully connected layer. These are the mapping features of the image modality.
[0041] In some embodiments, the initial features of the speech modality are mapped to the same semantic space as the initial features of the text modality data. This can be achieved using an LSTM-based acoustic model combined with an attention mechanism, as shown in the following formula: ; ; ; in, Let T represent the initial features of the speech in frame t, where T is the total number of speech frames. This is the hidden state of the LSTM. For attention weights, , and This is the weight matrix. This is the weighted context vector, which is the mapping feature of the output speech modality.
[0042] Furthermore, based on acoustic signal modeling, attention mechanisms can be used to enhance the model's sensitivity to emotional changes, providing support for generating more empathetic responses.
[0043] In one possible implementation, after obtaining the mapping features corresponding to each non-text modality, the mapping features are fused with the initial features of the text modality data. This can be achieved by directly concatenating the mapping features with the initial features and then performing a non-linear transformation. In another possible implementation of the method provided in this application, the mapping features are fused with the initial features of the text modality data to obtain semantic embedding features, including: Each mapping feature and the initial features of the text modality data are input into the context alignment module. The context alignment module is used to align the features of multiple modalities in the time dimension and obtain the alignment features of each modality output by the context alignment module. The alignment features of each modality are fused to obtain semantic chimerism features.
[0044] The context alignment module takes as input the various mapped features and the initial features of the text modality data, and outputs a cross-modal context vector for subsequent fusion, thereby achieving alignment of multimodal information in the temporal dimension. In some embodiments, the context alignment module can adopt a Transformer architecture. Of course, it is understood that, in addition to the Transformer architecture, the context alignment module can also adopt other model architectures to achieve alignment in the temporal dimension.
[0045] After obtaining the alignment features of each modality, the alignment features of each modality are fused. Existing feature fusion methods can be used here. For example, in some embodiments, the alignment features of each modality can be evaluated first and then input into an MLP (Multi-Layer Perceptron) to obtain the final semantic embedding features.
[0046] Through a semantic mapping layer and a context alignment module, data from different modalities can be aligned in time and space, resulting in higher quality semantic chimerism features obtained through subsequent fusion, which is beneficial to the accuracy of subsequent tasks. An integrated multimodal enhanced perception mechanism constructs a multi-path processing framework consisting of image semantic understanding, speech emotion modeling, and context alignment, breaking away from the traditional single-modal-to-text mapping design concept and achieving multi-channel semantic collaborative extraction.
[0047] After obtaining the semantic embedding features, these features are input into the intent recognition model and the sentiment recognition model, respectively, to obtain the intent recognition results and sentiment recognition results output by these models. The intent recognition result reflects the intent of the source user of the complaint data, while the sentiment recognition result reflects the emotional state of the source user. The response strategy is not equivalent to the response text, but rather information used to guide the generation of the response text. The response strategy reflects whether to generate response text or what type of response text to generate. For example, response strategies can include: direct answer, counter-question, emotional reassurance, structured recommendation, ending the conversation, etc. Combining emotional state and intent to generate response strategies can produce response text that is more in line with the user's needs.
[0048] In one possible implementation, a response strategy is generated based on the intent recognition result and the emotion recognition result. This can be achieved by inputting the intent recognition result and the emotion recognition result into a response strategy determination model, which then outputs the response strategy.
[0049] In another possible implementation, a response strategy is generated based on the intent recognition result, the emotion recognition result, and the response strategy determination model, including: Based on the intent recognition results, a search is performed in the complaint handling knowledge base to obtain candidate knowledge fragments; The intent recognition results, emotion recognition results, candidate knowledge fragments, and contextual data of the complaint data are input into the response strategy determination model to obtain the response strategy output by the response strategy determination model.
[0050] In this implementation, the input to the response strategy determination model includes not only the intent recognition results and sentiment recognition results, but also candidate knowledge fragments retrieved from the complaint processing knowledge base based on the intent recognition results, as well as contextual data of the complaint data. In this way, the response strategy determination model can generate more accurate response strategies based on richer reference information.
[0051] The complaint handling knowledge base can integrate structured knowledge graphs and unstructured knowledge bases. Based on the identified intent, it can call relevant domain knowledge items, logical templates or empirical rules to generate response candidate solutions, which have a certain degree of knowledge support and contextual relevance.
[0052] In some embodiments, when retrieving information from the complaint handling knowledge base based on intent recognition results, entity recognition results (such as the time, location, product name, etc. mentioned by the user) in the complaint data can also be combined to retrieve knowledge fragments more suitable for the current user. Retrieving information from the complaint handling knowledge base can be achieved by extracting triples from the knowledge graph using SPARQL or semantic retrieval mechanisms. This process can be represented as: ; In the above formula, For a knowledge graph set, h, r, and t represent head entity, relation, and tail entity, respectively.
[0053] For complex questions that cannot be matched using structured knowledge, relevant answer paragraphs are matched from unstructured knowledge bases using semantic retrieval methods as knowledge fragments.
[0054] Based on the response strategy determination model, the response strategy output can be used to generate response text. Specifically, a corresponding response text generation method can be pre-set for each possible response strategy. Response text generation methods can include two types: template filling and language generation model generation. For certain response strategies, corresponding templates are set. For example, for a complaint intent + anger emotion, the response strategy determination model determines the response strategy as "appeasement + responsibility confirmation," and calls the corresponding template. .
[0055] For certain response strategies targeting complex question-and-answer tasks, such as "asking a question in return" or "directly answering," a language generation model can be invoked. Based on sentiment recognition and intent recognition results, the extracted knowledge fragments are semantically compressed and natural language generated to obtain the response text. This process can be represented as: ; in, This represents the generated response text. Represents language generation model decoding processing, This represents the extracted knowledge fragments. These are the results of intent recognition and emotion recognition, respectively.
[0056] After generating the response text, the response text can be presented to the user through speech synthesis or a multimodal interface. In one possible implementation, to enhance the diversity and personalization of responses, a historical conversation memory mechanism can also be supported. That is, a memory enhancement module is introduced to improve the consistency and style control of long-term dialogues. This memory enhancement module can be embedded in the response strategy generation model or the semantic embedding model. That is, before the input data of the response strategy generation model or the semantic embedding model mentioned above is input into the response strategy generation model or the semantic embedding model, the input data is subjected to memory enhancement processing based on historical conversation data.
[0057] In the method provided in this application, the aforementioned semantic embedding model, intent recognition model, sentiment recognition model, and response strategy determination model can be trained together. Training can be divided into two stages: offline training before application and online training during application. Alternatively, in some embodiments, only the online training portion may be included. Offline training refers to using existing sample datasets to perform preliminary training on the semantic embedding model, intent recognition model, sentiment recognition model, and response strategy determination model. Online training refers to receiving complaint data from actual users, and based on the complaint data, using the method provided in this application, generating a response strategy using the aforementioned semantic embedding model, intent recognition model, sentiment recognition model, and response strategy determination model, generating response text based on the response strategy, and then feeding the response text back to the customer. User satisfaction is monitored in real time, and the semantic embedding model, intent recognition model, sentiment recognition model, and response strategy determination model are updated based on customer feedback data. Figure 2 As shown. That is, after generating the response text for the complaint data, it includes: Obtain the quality data corresponding to the response text; the quality data shall include at least the user's feedback data. The semantic feature embedding model, intent recognition model, sentiment recognition model, and response strategy determination model are updated based on quality data.
[0058] User feedback data can be extracted from user interaction logs, including user click-through rates, response times, and emotional responses.
[0059] In one possible implementation, updating the semantic feature embedding model, intent recognition model, sentiment recognition model, and response strategy determination model can be based on reinforcement learning. That is, updating these models based on quality data corresponding to the response text, including: Based on quality data, the reward values for the models are determined for the semantic feature embedding model, the intent recognition model, the sentiment recognition model, and the response strategy, respectively. Based on the reward values corresponding to the semantic feature embedding model, intent recognition model, sentiment recognition model, and response strategy determination model, reinforcement learning is used to update these models.
[0060] In reinforcement learning architectures, the response policy generation model can employ a policy network, which treats "generating a suitable response" as the process of selecting an action in the state space to obtain the maximum long-term reward. The policy function of the policy network can be defined as: Output the probability of choosing each action given the given state. State Based on the current dialogue context User intent Emotional state and candidate knowledge fragments Composition. Action This is an optional response strategy. The strategy network can be built using a lightweight MLP or Transformer module and shares some parameters with the response content generation module.
[0061] The quality of the generated response text is difficult to evaluate using a single metric. In some embodiments of the method provided in this application, a multi-dimensional reward function is used to calculate the reward value for the response strategy generation model. The specific formula is as follows: in Let t be the total reward value of the action at time t, which is the total reward value of the response policy generated by the policy generation model. To answer the consistency score between text content and knowledge base, User satisfaction (explicit rating). Negative penalty for inefficient dialogue rounds with users. To determine the match between the sentiment of the response text and the user's emotional state, γ1-γ4 are the weight coefficients for each dimension.
[0062] The goal of reinforcement learning is to maximize cumulative reward: ; Where J(π) is the performance objective function of policy π. Let be the expected value under strategy π, γ be the discount factor (0 < γ ≤ 1) controlling the importance of long-term rewards, t be the dialogue round time step, and T be the maximum dialogue round or termination step. This is the immediate reward for step t.
[0063] To maximize the cumulative reward, the REINFORCE policy gradient method can be used to iteratively update the response policy function: Where θ represents the policy network parameters (variables to be optimized and updated), ∇θ is the gradient operator with respect to θ, πθ(at|st) is the policy network parameterized with respect to θ, and the output state is... Next action The probability is logπθ(at|st), which is the logarithmic probability of the action (used for gradient ascent). For cumulative reward returns (as a weighted gradient).
[0064] The process of updating semantic feature embedding models, intent recognition models, and sentiment recognition models using a reinforcement learning framework is similar to that of updating response policy generation models. These models can all be viewed as policy functions, with actions as outputs and states as inputs. The reward value corresponding to the policy is calculated, and the model is updated based on this reward value to maximize the cumulative reward. The reward values for semantic feature embedding models, intent recognition models, and sentiment recognition models are explained below.
[0065] In one possible implementation, the reward function of the semantic embedding model is defined as follows: ; Where S represents user satisfaction rating, C represents response consistency index, and P represents cost of manual intervention. , , These are the weighting coefficients.
[0066] In one possible implementation, the reward R value of the intent recognition model and the emotion recognition model can be calculated based on the user's satisfaction rating (such as a 5-star rating) or a binary label indicating whether the problem is resolved after the session ends, which is a session-level delayed reward.
[0067] It is worth noting that in some embodiments, the reward value of each model in the method provided in this application can also be calculated based on other indicators that can reflect the quality of the response text, such as the length of the response text, the fluency of the response text, etc. This application does not make specific limitations on this.
[0068] By updating the semantic feature embedding model, intent recognition model, sentiment recognition model, and response strategy determination model using the quality data of the response text, the execution status of the scheme can be detected in real time, and the model can be continuously optimized based on the feedback of the business processing effect.
[0069] The complaint data response generation apparatus provided in this application is described below. The complaint data response generation apparatus described below corresponds to and can be referred to in relation to the complaint data response generation method described above. For example... Figure 3 As shown, the complaint data response generation device provided in this application includes: The semantic embedding module 310 is used to acquire complaint data, which includes data from multiple modalities. The complaint data is input into the semantic feature embedding model to obtain the semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain the initial features of each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain the semantic embedding features. The intent and emotion recognition module 320 is used to input semantic embedding features into the intent recognition model and the emotion recognition model respectively, and obtain the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model. The response generation module 330 is used to generate a response strategy based on the intent recognition result, the emotion recognition result, and the response strategy determination model, and to generate a response text for the complaint data based on the response strategy and the complaint handling knowledge base.
[0070] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logical instructions in the memory 430 to execute a complaint data response generation method. This method includes: acquiring complaint data, which includes data from multiple modalities; inputting the complaint data into a semantic feature embedding model to obtain semantic embedding features output by the model; the semantic feature embedding model includes a feature extraction layer and a feature fusion layer; the feature extraction layer extracts features from the data of each modality in the complaint data to obtain initial features for each modality; the feature fusion layer fuses the initial features of each modality to obtain semantic embedding features; inputting the semantic embedding features into an intent recognition model and an emotion recognition model respectively to obtain the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model; determining a model based on the intent recognition result, the emotion recognition result, and a response strategy; generating a response strategy; and generating a response text for the complaint data based on the response strategy and a complaint processing knowledge base.
[0071] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0072] On the other hand, this application also provides a computer program product, which includes a computer program that 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 data response generation method provided by the above methods. The method includes: acquiring complaint data, which includes data of multiple modalities; inputting the complaint data into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features of each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features; inputting the semantic embedding features into an intent recognition model and an emotion recognition model respectively to obtain the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model; determining a model based on the intent recognition result, the emotion recognition result, and the response strategy to generate a response strategy; and generating a response text for the complaint data based on the response strategy and a complaint processing knowledge base.
[0073] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for generating a complaint data response provided by the methods described above. This method includes: acquiring complaint data, which includes data from multiple modalities; inputting the complaint data into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features for each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features; inputting the semantic embedding features into an intent recognition model and an emotion recognition model respectively to obtain intent recognition results output by the intent recognition model and emotion recognition results output by the emotion recognition model; determining a model based on the intent recognition results, emotion recognition results, and a response strategy to generate a response strategy; and generating a response text for the complaint data based on the response strategy and a complaint processing knowledge base.
[0074] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0075] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for generating complaint data responses, characterized in that, include: Complaint data is acquired, which includes data from multiple modalities. The complaint data is input into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features for each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features. The semantic embedding features are input into the intent recognition model and the emotion recognition model respectively, and the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model are obtained. Based on the intent recognition result, the emotion recognition result, and the response strategy determination model, a response strategy is generated. Based on the response strategy and the complaint handling knowledge base, a response text is generated for the complaint data.
2. The complaint data response generation method according to claim 1, characterized in that, The step of generating a response strategy based on the intent recognition result and the emotion recognition result includes: Based on the intent recognition result, a search is performed in the complaint processing knowledge base to obtain candidate knowledge fragments; The intent recognition result, the emotion recognition result, the candidate knowledge fragment, and the context data of the complaint data are input into the response strategy determination model to obtain the response strategy output by the response strategy determination model.
3. The complaint data response generation method according to claim 1, characterized in that, The complaint data includes non-text modal data and text modal data; the fusion of the initial features of each modality to obtain semantic embedding features includes: The initial features of each of the non-text modal data are respectively input to the semantic mapping layer of the corresponding modality to obtain the mapping features output by each of the semantic mapping layers. The semantic mapping layer is used to map the initial features of the non-text modal data to the same semantic space as the initial features of the text modal data. The semantic chimerism features are obtained by fusing each of the mapping features with the initial features of the text modality data.
4. The complaint data response generation method according to claim 3, characterized in that, The step of fusing each of the mapping features with the initial features of the text modality data to obtain the semantic embedding features includes: Each of the mapping features and the initial features of the text modality data are input to the context alignment module. The context alignment module is used to align the features of multiple modalities in the time dimension and obtain the alignment features of each modality output by the context alignment module. The alignment features of each modality are fused to obtain the semantic chimerism feature.
5. The complaint data response generation method according to claim 1, characterized in that, After generating the response text for the complaint data, the process includes: Obtain the quality data corresponding to the response text, wherein the quality data includes at least the user's feedback data; The semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model are updated based on the quality data.
6. The complaint data response generation method according to claim 5, characterized in that, The step of updating the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model based on the quality data includes: Based on the quality data, the reward values of the models are determined for the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy, respectively. Based on the reward values corresponding to the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model, reinforcement learning is used to update the semantic feature embedding model, the intent recognition model, the emotion recognition model, and the response strategy determination model.
7. A complaint data response generation device, characterized in that, include: A semantic embedding module is used to acquire complaint data, which includes data from multiple modalities. The complaint data is input into a semantic feature embedding model to obtain semantic embedding features output by the semantic feature embedding model. The semantic feature embedding model includes a feature extraction layer and a feature fusion layer. The feature extraction layer is used to extract features from the data of each modality in the complaint data to obtain initial features for each modality. The feature fusion layer is used to fuse the initial features of each modality to obtain semantic embedding features. intention The semantic embedding features are input into the intent recognition model and the emotion recognition model respectively, and the intent recognition result output by the intent recognition model and the emotion recognition result output by the emotion recognition model are obtained. The response generation module is used to generate a response strategy based on the intent recognition result, the emotion recognition result, and the response strategy determination model, and to generate a response text for the complaint data based on the response strategy and the complaint handling knowledge base.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the complaint data response generation method as described in any one of claims 1 to 6.
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, it implements the complaint data response generation method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the complaint data response generation method as described in any one of claims 1 to 6.