Question and answer determination method and device, equipment, medium and program product

Abstract

The invention provides a question and answer determination method which can be applied to the technical field of artificial intelligence and digital human. The question and answer determination method comprises the following steps: in response to a dialogue request of a customer, acquiring a customer utterance in a preset time window; according to the customer utterance, a question boundary of each question in the customer utterance is identified by using a pre-trained question and answer model, and all question statements are obtained according to each question boundary; according to all the question statements, determining answer statements matched with the question statements in the customer service utterances by utilizing a question and answer model; and feeding back the answer statement corresponding to each question statement to the customer. The invention further provides a question and answer determination device and equipment, a storage medium and a program product.

Description

Technical Field

[0001] The present application relates to the financial field, specifically to the field of artificial intelligence and digital human technology, and more specifically to a question-and-answer determination method, apparatus, device, medium, and program product. Background Art

[0002] In the digital age, the customer service sector is rapidly transforming toward intelligent solutions. Service organizations face a growing volume of customer inquiries and demands for personalized services. This requires customer service systems to not only provide rapid responses but also ensure the accuracy and relevance of answers. However, in customer service interactions, customers may ask multiple questions, each of which may correspond to one or more answers. Currently, it is impossible to accurately identify and locate the boundaries between these questions and their answers, resulting in low accuracy in question recognition and answer matching. Summary of the Invention

[0003] In view of the above problems, the present application provides a question and answer determination method, apparatus, device, medium and program product for improving the accuracy of question and answer matching.

[0004] According to the first aspect of the present application, a question and answer determination method is provided, including: in response to a customer's dialogue request, with the customer's consent or authorization, obtaining the customer's speech within a preset time window; based on the customer's speech, using a pre-trained question and answer model, identifying the question boundary of each question in the customer's speech, and obtaining all question statements based on each of the question boundaries; based on all the question statements, using the question and answer model, determining the answer statement in the customer service speech that matches the question statement; and feeding back the answer statement corresponding to each of the question statements to the customer.

[0005] According to an embodiment of the present application, the method of identifying the problem boundary of each question in the customer utterance based on the customer utterance using a pre-trained question-answering model includes: dividing the customer utterance into a first sentence sequence; concatenating the first sentence sequence with a task identifier, and inputting the result into a target encoder of the question-answering model to obtain a hidden representation of each sentence in the first sentence sequence; based on the hidden representation, modeling the temporal dependency between the sentences to generate a temporally enhanced contextual representation; and sequence-labeling the contextual representation to predict the boundary label of each sentence, and merging sentences corresponding to adjacent boundary labels to determine the problem boundary of each question.

[0006] According to an embodiment of the present application, the question-answering model is used to determine the answer statement in the customer service utterance that matches the question statement based on all the question statements, including: splicing each identified question statement with the second sentence sequence in the customer service utterance to obtain a question-answer pair input sequence; inputting the question-answer pair input sequence into the target encoder to obtain a joint hidden representation of the question statement and the customer service utterance; mapping all sentence representations in the joint hidden representation into sentence nodes, and constructing edges between the sentence nodes according to lexical relevance and semantic relevance, respectively, to obtain a lexical graph and a semantic graph; performing graph convolution on the lexical graph and the semantic graph respectively, and interactively fusing them through double affine transformation to obtain fused features; and based on the fused features, using the question-answering model, predicting the answer statement corresponding to the question statement.

[0007] According to an embodiment of the present application, the edges between the sentence nodes are constructed according to lexical relevance and semantic relevance respectively to obtain a lexical graph and a semantic graph, including: when the sentence node in the question statement and the sentence node in the customer service utterance contain at least one identical keyword, the edges between the sentence nodes are established to form the lexical graph; and based on the joint hidden representation, the attention matrix between the sentence nodes is calculated through a multi-head attention mechanism, and the edges in the attention matrix whose attention scores are higher than a first preset threshold are retained to form the semantic graph.

[0008] According to an embodiment of the present application, training the question-answering model includes: forming a positive sample pair with the question statement and the corresponding true answer statement, and forming a negative sample pair with the question statement and the answer statements of other questions in the same training batch; calculating the similarity score between the positive sample pair and the negative sample pair; and calculating the supervised contrast loss based on the similarity score.

[0009] According to an embodiment of the present application, training the question-answering model further includes: performing weighted summation of the sequence labeling loss, the question-answer pair matching loss and the supervised contrast loss as the total loss function of the question-answering model, wherein the sequence labeling loss is used to characterize the loss of identifying the question boundary, and the question-answer pair matching loss is used to characterize the loss of determining the answer statement that matches the question statement; based on the total loss function, adjusting the parameters of the question-answering model until the total loss function meets a preset convergence condition.

[0010] According to an embodiment of the present application, the method further includes: storing each of the question statements and the corresponding answer statements in a knowledge question and answer database; and in response to a customer inputting a new question statement, matching the corresponding answer statement in the knowledge question and answer database and feeding back to the customer.

[0011] The second aspect of the present application provides a question-and-answer determination device, including: an acquisition module for obtaining customer speech within a preset time window in response to a customer's dialogue request and with the customer's consent or authorization; a question determination module for identifying the question boundary of each question in the customer speech based on the customer speech using a pre-trained question-and-answer model, and obtaining all question statements based on each of the question boundaries; an answer determination module for determining the answer statement in the customer service speech that matches the question statement based on all the question statements using the question-and-answer model; and a feedback module for feeding back the answer statement corresponding to each of the question statements to the customer.

[0012] The third aspect of the present application provides an electronic device, comprising: one or more processors; a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the above method.

[0013] The fourth aspect of the present application further provides a computer-readable storage medium having a computer program or instructions stored thereon, which implements the steps of the above method when the computer program or instructions are executed by a processor.

[0014] The fifth aspect of the present application further provides a computer program product, comprising a computer program or instructions, which implement the steps of the above method when executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The above contents and other objects, features and advantages of the present application will become more apparent through the following description of the embodiments of the present application with reference to the accompanying drawings, in which:

[0016] Figure 1 A diagram schematically illustrates an application scenario of the question-answer determination method, apparatus, device, medium, and program product according to an embodiment of the present application;

[0017] Figure 2 Schematically shows a flow chart of a question-answer determination method according to an embodiment of the present application;

[0018] Figure 3 A flowchart of determining a question-answer pair through a question-answer model according to an embodiment of the present application is schematically shown;

[0019] Figure 4 A vocabulary graph according to an embodiment of the present application is schematically shown;

[0020] Figure 5 A schematic diagram schematically shows an adjacency matrix corresponding to a vocabulary graph according to an embodiment of the present application;

[0021] Figure 6A schematic diagram of a structural block diagram of a question-answer determination device according to an embodiment of the present application is shown; and

[0022] Figure 7 A block diagram of an electronic device suitable for implementing the question-answer determination method according to an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0023] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present application. In the detailed description below, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present application. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present application.

[0024] The terms used herein are only for describing specific embodiments and are not intended to limit this application. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0026] When expressions such as "at least one of A, B, and C, etc." are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0027] In the technical solution of this application, the user information involved (including but not limited to user personal information, user image information, user device information, such as location information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0028] In the scenario of using personal information for automated decision-making, the methods, devices, and systems provided in the embodiments of the present application all provide users with corresponding operation portals for users to choose to agree or reject the automated decision-making results; if the user chooses to reject, the expert decision-making process will be entered. The expression "automated decision-making" here refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests and hobbies, or economic, health, credit status, etc. through computer programs and making decisions. The expression "expert decision-making" here refers to the activity of making decisions by people who specialize in a certain field, have specialized experience, knowledge and skills, and have reached a certain level of professionalism.

[0029] Embodiments of the present application provide a question-and-answer determination method, apparatus, device, medium, and program product.

[0030] Figure 1 The application scenario diagram of the question and answer determination method, device, equipment, medium and program product according to the embodiments of the present application is schematically shown.

[0031] like Figure 1 As shown, the application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or optical fiber cables.

[0032] A user may use a first terminal device 101, a second terminal device 102, or a third terminal device 103 to interact with a server 105 via a network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, or the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (for example only).

[0033] The first terminal device 101 , the second terminal device 102 , and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0034] The server 105 may be a server that provides various services, such as a background management server (for example only) that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process received data such as user requests, and feed back processing results (e.g., web pages, information, or data obtained or generated based on user requests) to the terminal devices.

[0035] It should be noted that the question-and-answer determination method provided in the embodiment of the present application can generally be executed by the server 105. Accordingly, the question-and-answer determination device provided in the embodiment of the present application can generally be set in the server 105. The question-and-answer determination method provided in the embodiment of the present application can also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Accordingly, the question-and-answer determination device provided in the embodiment of the present application can also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105.

[0036] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0037] The following will be based on Figure 1 The scene described by Figures 2 to 5 The question and answer determination method according to an embodiment of the present application is described in detail.

[0038] First, the technical terms described in this article are explained and described as follows.

[0039] Graph: A graph in graph theory is a mathematical model used to describe the relationship between objects. It consists of two parts: a vertex set and an edge set.

[0040] Adjacency matrix: The adjacency matrix is ​​a method used in graph theory to represent the connection relationship between vertices in a graph.

[0041] Graph Convolution: Graph Convolutional Network (GCN) is a deep learning model specifically designed for processing graph-structured data. The core idea of ​​GCN is to update the feature representation of a node by aggregating the features of the node and its neighbors.

[0042] Contrastive learning: Contrastive learning is a self-supervised learning method that trains a model to distinguish between similar and dissimilar data by constructing pairs of positive and negative examples. It uses data augmentation to create variants as positive examples and other data as negative examples. The model's goal is to narrow the gap between the feature representations of positive examples and expand the gap between the feature representations of negative examples.

[0043] Sequence labeling: Sequence labeling is mainly used to identify various parts in the text. For example, in this patent, it is used to identify which sentences in the user's speech belong to key questions and find the boundaries of the questions.

[0044] Co-occurring words: words that appear together in two texts.

[0045] Multi-head attention: This decomposes the attention mechanism into multiple "heads," allowing the model to learn information in parallel across different representation subspaces. Each head processes the input data through its own linear transformation, then computes an attention score and aggregates the information. Finally, the outputs of all heads are concatenated and passed through a linear layer to form the final output.

[0046] Conditional Random Fields (CRFs) can capture dependencies between labels and improve labeling accuracy. They optimize label predictions for the entire sequence by learning the transition probabilities of label sequences. They are commonly used in named entity recognition and part-of-speech tagging.

[0047] Longformer Encoder: An encoder suitable for long texts.

[0048] In an embodiment of the present application, training the question-answering model includes: forming a positive sample pair with the question statement and the corresponding true answer statement, and forming a negative sample pair with the question statement and the answer statements of other questions in the same training batch; calculating the similarity score between the positive sample pair and the negative sample pair; and calculating the supervised contrast loss based on the similarity score.

[0049] In contrastive learning, positive pairs are sentences from the same question and their corresponding question-answer pair, specifically, question statements and their true answers. For example, the question "What is a payment password?" and the answer "A payment password is a key digital credential used to secure payment transactions and verify user identity" are highly semantically related. Forming positive pairs helps the model learn this semantic matching relationship. Negative pairs are different questions and customer service utterances, specifically, question statements and answers from other questions within the same training batch. Different questions and answers within the same training batch are often semantically distinct. Combining these into negative pairs allows the model to learn to distinguish between different semantic content. For example, in a training batch containing the questions "I want to apply for a credit card installment plan. How do I do this?" and "Has the interest rate on my bank's fixed deposit been adjusted recently?", the question "I want to apply for a credit card installment plan. How do I do this?" and the answer "Has the interest rate on my bank's fixed deposit been adjusted recently?" constitute negative pairs. By learning this mismatch, the model can better understand which content is irrelevant.

[0050] The similarity score can measure the semantic similarity between sample pairs and provide a basis for the calculation of subsequent supervised contrast loss.

[0051] For example, cosine similarity can be used as the similarity metric, that is, , where h u and h v Represent the representation vectors of sentence u and sentence v respectively. Cosine similarity measures the similarity between two vectors by calculating the cosine value of the angle between them. The value range is When the directions of the two vectors are closer, the cosine value is larger, indicating that the semantic similarity between the sentences is higher; when the directions of the two vectors are more different, the cosine value is smaller, indicating that the semantic similarity between the sentences is lower.

[0052] For each pair of samples (positive and negative), the cosine similarity formula can be used to calculate their similarity score. For example, for a positive pair (a question and its true answer), the cosine similarity between their representation vectors can be calculated; for a negative pair (a question and its answer to another question in the same batch), the cosine similarity between their representation vectors can also be calculated.

[0053] For a training batch B of data, we can collect question statement representations and answer statement representations, and the supervised contrast loss can be defined as:

[0054]

[0055]

[0056] in, Represents u sentence representation vector h u Similar to v sentence representation vector h v The possibility of τ is the temperature coefficient of Softmax. u Indicates the number of samples similar to the sentences in the customer's question, Indicates that sentence u is similar to sentence v.

[0057] By calculating and minimizing the supervised contrast loss, the similarity probability between samples of the same class (i.e., semantically similar samples) can be maximized, while the similarity probability between samples of different classes (i.e., semantically dissimilar samples) can be minimized. This enables the model to learn more effective sentence representations, making semantically similar sentences closer in the representation space and semantically dissimilar sentences farther apart in the representation space, thereby improving the question-answering model's ability to recognize the semantic association between questions and answers.

[0058] It is understandable that the question-answering model's ability to recognize the semantic association between questions and answers is the key to question-answer pair extraction. By constructing positive and negative sample pairs, this application enables the model to learn to distinguish between valid and invalid question-answer pairs, thereby helping the model to accurately capture the core intent of question statements in massive data and match them to corresponding answer statements, thereby improving the accuracy and efficiency of question answering.

[0059] By using a small amount of labeled samples for in-depth training, we can efficiently build an automated question-answering model to extract and organize customer service question-answer pairs. This process not only significantly improves data organization efficiency but also ensures the accuracy and consistency of information.

[0060] In an embodiment of the present application, training the question-answering model further includes: performing weighted summation of the sequence labeling loss, the question-answer pair matching loss, and the supervised contrast loss as the total loss function of the question-answering model, wherein the sequence labeling loss is used to characterize the loss of identifying the question boundary, and the question-answer pair matching loss is used to characterize the loss of determining the answer statement that matches the question statement; based on the total loss function, adjusting the parameters of the question-answering model until the total loss function meets a preset convergence condition.

[0061] During the training process of the question-answering model, the model loss can include three parts: sequence labeling loss, question-answer pair matching loss, and supervised comparison loss.

[0062] The sequence labeling loss is used to characterize the loss of the identification problem boundary, and the calculation formula is as follows:

[0063]

[0064] Among them, Y A is the correct question tag sequence, D q Represents the input problem data. This formula indicates that given the input problem data D q Under the condition of A By minimizing L A , the model can learn more accurate problem boundary recognition capabilities.

[0065] The question-answer pair matching loss is used to characterize the loss of determining the answer statement that matches the question statement. The calculation formula is as follows:

[0066]

[0067] where Y B is the correct label sequence of the customer service utterance corresponding to the question, D r Indicates the answer data, X q Represents the problem data. This formula means that given the problem data X q and answer data D r Under the condition of B By minimizing L B , the model can better understand the semantic relationship between questions and answers and improve the accuracy of question-answer pair matching.

[0068] By summing the sequence labeling loss, question-answer pair matching loss, and supervised contrast loss, we can get the total loss function of the question-answering model, which serves as the training objective of the model, as follows:

[0069] L=L A +L B +L sup

[0070] Considering the performance requirements of the model in different aspects, the model can achieve balance and optimization in question boundary recognition, question-answer pair matching, and semantic association learning. It can also weight the sequence labeling loss, question-answer pair matching loss, and supervised contrast loss to obtain the total loss function L=αL A +βL B +γL sup , where α, β, and γ are the weights of the three losses, and α+β+γ=1.

[0071] Based on the total loss function L, optimization algorithms such as gradient descent can be used to adjust the parameters of the question-answering model. In each iteration, the gradient of the total loss function with respect to the model parameters is calculated. The parameters are then updated in the opposite direction of the gradient, gradually reducing the total loss function until convergence is reached. For example, this may occur when the total loss function falls below a certain threshold, or when the model's performance metrics (such as accuracy and recall) on the validation set reach a certain level and no longer improve significantly. Finally, model parameter adjustments can be stopped, resulting in a trained question-answering model.

[0072] It's important to note that with the continuous advancement of technology, question-answering models can continuously optimize themselves through machine learning algorithms, improving the accuracy of question recognition and answer matching. This not only improves the quality of customer service but also reduces manual intervention, allowing customer service teams to focus more on resolving more complex or personalized customer issues.

[0073] It is understandable that taking the weighted sum of sequence labeling loss, question-answer pair matching loss, and supervised contrast loss as the total loss function can optimize the question-answering model from multiple dimensions, enabling the model to synergistically improve question understanding, answer selection, and semantic understanding, thereby comprehensively enhancing question-answering performance.

[0074] Figure 2 The flowchart of the question-and-answer determination method according to an embodiment of the present application is schematically shown.

[0075] like Figure 2 As shown, the question and answer determination method of this embodiment includes operations S210 to S240.

[0076] In operation S210 , in response to a conversation request from a customer, and with the customer's consent or authorization, customer speech within a preset time window is acquired.

[0077] In an embodiment of the present application, before acquiring the customer's speech within a preset time window, the user's consent or authorization may be obtained. For example, before operation S210, a request may be issued to the user to acquire the customer's speech within the preset time window. If the user agrees or authorizes the acquisition of the customer's speech within the preset time window, operation S210 is performed.

[0078] In operation S220, based on the customer utterance, a pre-trained question-answering model is used to identify the question boundary of each question in the customer utterance, and all question statements are obtained based on the question boundaries.

[0079] In operation S230, based on all the question statements, the question-answering model is used to determine an answer statement in the customer service utterance that matches the question statement.

[0080] In operation S240, the answer statements corresponding to the question statements are fed back to the customer.

[0081] In an embodiment of the present application, a corresponding operation entry can be provided for the user to choose to agree or reject the automated decision result. That is, before the answer statement corresponding to each question statement is fed back to the client for processing / decision-making, an instruction input by the user through the corresponding operation entry to agree or reject the processing / decision-making can be obtained. If the user agrees to the processing / decision-making, the answer statement corresponding to each question statement is fed back to the client for processing / decision-making, that is, step S240 is executed. If the user rejects the processing / decision-making, the expert decision-making process is entered.

[0082] Multi-turn conversations in customer service are a complex and sophisticated process. During this process, when a customer initiates a conversation request through a system interface, mobile app, or other access channel, a response mechanism can be immediately activated. The customer's input is continuously monitored and captured within a preset time window (e.g., 30 seconds or 1 minute). Sequence labeling methods can then be used to identify each specific question in the customer's utterance and accurately determine the start and end points of the question statement. First, the customer utterance can be input into a question-answering model, which learns the patterns and characteristics of questions to identify similar questions in new text. These utterances can contain multiple questions, either consecutive or scattered across a paragraph, and the boundaries between these questions must be identified. For example, for a compound sentence containing multiple questions, such as "I want to know the price of a financial product and its yield?", the question-answering model can accurately identify the boundaries between the two independent questions: "I want to know the price of the financial product" and "What is its yield?"

[0083] After all question statements have been identified, the customer service discourse segment can be entered. Using the previously identified question statements as key utterances, the question-answering model can be used to search for corresponding answer statements in the customer service discourse, quickly finding the answer that matches the question in a large amount of text. For example, when a customer asks, "What are the returns of this wealth management product?" semantic analysis can be used to understand that the customer is concerned about information such as return stability and acquisition methods. Then, answer statements that include historical return data, expected returns for different periods, and the correlation between returns and the market can be filtered out. Finally, the answer statements corresponding to each question statement can be organized and optimized, and feedback can be provided to the customer in a clear and easy-to-understand manner. The feedback format can be flexibly set according to actual needs, such as plain text, a combination of text and images, and voice broadcast, to meet the usage habits and scenario needs of different customers.

[0084] To enhance the customer experience, additional interactive features can be provided alongside feedback on responses. For example, customers can rate responses, helping the system continuously optimize response quality. If customers are dissatisfied with the response or have further questions, they can ask further questions. Furthermore, question recognition and answer matching can be re-performed based on the customer's follow-up question.

[0085] It is understandable that by accurately identifying the boundaries of each question through the question-answering model, that is, by using the sequence labeling method to identify each specific question in the customer's discourse, the start and end positions of the question statement can be accurately determined, thereby providing clear and accurate input for subsequent answer matching, and ultimately improving the accuracy and efficiency of question-answering matching.

[0086] On the basis of the above embodiments, in this embodiment, the method of identifying the problem boundary of each question in the customer utterance based on the customer utterance and using a pre-trained question-answering model includes: dividing the customer utterance into a first sentence sequence; concatenating the first sentence sequence with a task identifier, and inputting the result into a target encoder of the question-answering model to obtain a hidden representation of each sentence in the first sentence sequence; modeling the temporal dependency between the sentences based on the hidden representation to generate a temporally enhanced contextual representation; and performing sequence labeling on the contextual representation to predict the boundary label of each sentence, and merging sentences corresponding to adjacent boundary labels to determine the problem boundary of each question.

[0087] Figure 3 The flowchart of determining question-answer pairs through a question-answer model according to an embodiment of the present application is schematically shown.

[0088] In some exemplary embodiments, Figure 3 As shown, we can first split the customer utterances and customer service utterances into a first sentence sequence. This sequence is then concatenated with the task identifier to form an input sequence. This sequence is then fed into a target encoder (such as a Longformer encoder) to obtain a hidden representation of each sentence that contains information about the semantics, grammar, and other aspects of the sentence.

[0089] For example, a Longformer encoder can be used to extract context information, and a special symbol [T] can be used to represent the query in the sequence labeling task, so that the question-answering model can mine all the specific questions in the customer's utterance:

[0090]

[0091] Among them I A Represents the input of Longformer in the sequence labeling task. [s] and [\s] are special symbols in Longformer. is the i-th sentence in the customer's speech. After the above sequence is fed into Longformer, the hidden representation H of each sentence can be obtained. q :

[0092]

[0093] Subsequently, a sequence tagger based on the Conditional Random Fields (CRF) algorithm can be used to identify potential questions / answers in the paragraph, i.e., H q The input is fed into a recurrent neural network (Long Short-Term Memory, LSTM). As a recurrent neural network capable of processing sequential data, LSTM can model the temporal dependencies between sentences and generate a temporally enhanced contextual representation. The contextual representation can then be fed into a CRF, taking into account the transition probabilities between labels, thereby more accurately predicting the boundary labels for each sentence. During the prediction process, CRF can assign a boundary label to each sentence, such as a question start label or a question end label, based on the contextual information and the constraints between the labels. Based on the question boundaries in the two paragraphs, i.e., the predicted boundary labels, the sentences corresponding to adjacent boundary labels are merged. Ultimately, all the question Xs in the customer's utterance can be determined based on the boundaries. q :

[0094]

[0095] in, is the i-th question in the customer's utterance and can contain multiple sentences.

[0096] It is understandable that by comprehensively considering multiple factors such as the semantics and temporal relationships of the sentence, the misjudgment rate can be effectively reduced, thereby accurately identifying the problem boundaries in the customer's speech.

[0097] On the basis of the above embodiments, in this embodiment, the question-answering model is used to determine the answer statement in the customer service utterance that matches the question statement based on all the question statements, including: concatenating each identified question statement with the second sentence sequence in the customer service utterance to obtain a question-answer pair input sequence; inputting the question-answer pair input sequence into the target encoder to obtain a joint hidden representation of the question statement and the customer service utterance; mapping all sentence representations in the joint hidden representation into sentence nodes, and constructing edges between the sentence nodes according to lexical relevance and semantic relevance, respectively, to obtain a lexical graph and a semantic graph; performing graph convolution on the lexical graph and the semantic graph respectively, and interactively fusing them through double affine transformation to obtain fused features; and based on the fused features, using the question-answering model, predicting the answer statement corresponding to the question statement.

[0098] Continue to refer Figure 3 Using the questions extracted from sequence labeling as queries, each identified question statement can be concatenated with the second statement sequence in the customer service utterance to obtain a question-answer pair input sequence. For example, suppose the identified customer question is "What is the warranty period for this product?" and the customer service utterance contains "Our product warranty period is one year, and if any quality problems occur under normal use, free repair service is available." The question statement is concatenated with each sentence in the customer service utterance (the customer service utterance can be considered a whole or a sequence of sentences) to form an input form similar to "[Question] What is the warranty period for this product? [Customer Service utterance] Our product warranty period is one year, and if any quality problems occur under normal use, free repair service is available."

[0099] Input the constructed question-answer pair into sequence I B Input the same target encoder (such as Longformer encoder) to get a new sentence representation:

[0100]

[0101]

[0102] Among them I B Represents the input of Longformer in the question-answer pair matching task, represents the jth sentence in the customer service discourse, H B The hidden layer representation of the customer question and the customer service utterance (i.e., the joint hidden representation), and Indicates the beginning and ending sentences of a question.

[0103] Subsequently, all sentences in the question and customer service answer can be regarded as nodes, and lexical graphs and semantic graphs can be constructed based on lexical relevance and semantic relevance to model the relationship between question-answer pairs.

[0104] Furthermore, based on the adjacency matrix of the above vocabulary graph and semantic graph, graph convolution can be used on the graph to interact between sentence nodes:

[0105]

[0106]

[0107] in, and Represents a trainable matrix, H B represents the joint hidden representation, and σ represents the activation function.

[0108] In order to effectively exchange the relevant features between the vocabulary graph and the semantic graph, a double affine transformation can be used for interactive fusion. Considering the bidirectional relationship between the nodes of the vocabulary graph and the semantic graph, the feature information of one graph is fused into the other graph by calculating the similarity weights between the nodes. For example, when calculating H co' When , the similarity weights between the vocabulary graph nodes and the semantic graph nodes are calculated by the softmax function, and then the feature information of the semantic graph is fused into the node representation of the vocabulary graph according to the weights, and vice versa.

[0109]

[0110]

[0111] Among them, W1 and W2 are trainable parameters.

[0112] The vocabulary graph feature H after the double affine transformation co' and semantic graph features H se' Perform splicing to obtain the fusion feature H:

[0113]

[0114] Among them, [,] means concatenating two vectors.

[0115] Finally, the fused feature H obtained above can be input into the question-answering model (a combination of Bi-LSTM (bidirectional long short-term memory network) and CRF models) for processing to obtain the answer statement corresponding to the question.

[0116] It is understandable that by considering the lexical and semantic relevance between question-answer pairs, and understanding their context and relevance in the conversation through techniques such as graph structure and biaffine transformation, the corresponding answer sentences can be matched more accurately.

[0117] Based on the above embodiments, in this embodiment, the edges between the sentence nodes are constructed according to lexical relevance and semantic relevance respectively to obtain a lexical graph and a semantic graph, including: when the sentence node in the question statement and the sentence node in the customer service utterance contain at least one identical keyword, the edges between the sentence nodes are established to form the lexical graph; and based on the joint hidden representation, the attention matrix between the sentence nodes is calculated through a multi-head attention mechanism, and the edges in the attention matrix whose attention scores are higher than a first preset threshold are retained to form the semantic graph.

[0118] In automated customer service systems, simply identifying keywords is insufficient, as the richness and diversity of language requires the system to understand the deeper meaning and semantic relationships of words. Embodiments of this application can perform deep matching of words and semantics to ensure that the match between questions and answers is not based solely on superficial lexical similarity, but on deeper semantic relevance.

[0119] A vocabulary graph A can be constructed based on the co-occurring words in the questions and customer service discourse co To establish a relationship of lexical relevance, that is, when the sentence node in the question statement and the sentence node in the customer service discourse contain at least one co-occurring word, then an edge relationship can be established in the graph to form a lexical graph, such as Figure 4 shown.

[0120] The adjacency matrix is ​​an n×n square matrix (n is the total number of sentence nodes in the vocabulary graph). Each element A(i,j) in the matrix represents the connection between node i and node j. In the vocabulary graph, if there is an edge between nodes i and j (that is, they contain at least one common keyword), then A(i,j) = 1; otherwise, A(i,j) = 0. Figure 4 The adjacency matrix shown is Figure 5 shown.

[0121] Assume that the set of sentence nodes in the question statement is Q={q1,q2,⋯,qm}, and the set of sentence nodes in the customer service utterance is R={r1,r2,⋯,rn}. For any qi∈Q and rj∈R, if they have at least one common keyword, then in the vocabulary graph A co middle, , indicating that there is an edge between qi and rj; otherwise, For example, continue to refer to Figure 3 , the question statement is "The system always prompts me that the password is wrong. I want to confirm whether my account is frozen?", and the customer service discourse is "Please provide your account and password so that I can check." and "After checking, your account status is normal. This may mean that you have remembered the wrong password." Among them, "account" and "password" are co-occurring words. In the vocabulary graph, an edge can be established between the sentence nodes containing these two words in the question sentence and the corresponding sentence nodes in the customer service discourse.

[0122] Simply considering lexical information between sentences can lead to mismatches. For example, synonyms, near-synonyms, or different expressions with the same semantics cannot be effectively captured by the lexical graph. Therefore, the semantic relevance between sentences can also be considered. Specifically, a multi-head attention mechanism can be used to calculate the attention matrix between sentence nodes between questions and customer service utterances. Edges in the attention matrix with an attention score above a first preset threshold are retained. Alternatively, a top-k strategy can be used to retain the top k edges with the highest attention scores to construct a semantic graph:

[0123]

[0124]

[0125]

[0126] Among them, W q and W r Represents the trainable matrix, d Longformer is the dimension of the sentence after Longformer encoding, and K is the number of multi-head attention.

[0127] It's clear that by constructing a lexical graph and a semantic graph, we can fully mine the connections between sentence nodes in question sentences and customer service utterances. The lexical graph captures surface connections based on keywords, while the semantic graph mines deeper semantics using multi-head attention. The combination of the two provides a more accurate and comprehensive basis for question-answer matching, helping to improve the accuracy of question-answer determination.

[0128] In an embodiment of the present application, the method further includes: storing each of the question statements and the corresponding answer statements in a knowledge question and answer database; and in response to a customer inputting a new question statement, matching the corresponding answer statement in the knowledge question and answer database and feeding it back to the customer.

[0129] The question-and-answer determination method provided in the embodiments of the present application can accurately identify customer questions and find corresponding solutions in customer service responses. Based on this, past customer questions and their solutions can be converted into structured question-and-answer pairs. Based on valid question-and-answer pairs in historical conversation data between customers and customer service representatives, an efficient, accurate, and dynamically updated knowledge question-and-answer database can be constructed. This knowledge database contains customer questions and customer service representatives' responses. As the question-and-answer database continues to improve, customer service representatives can directly query this knowledge database when handling new customer questions, quickly match answers to similar questions in the knowledge question-and-answer database, and provide them to customers, thereby providing more timely and accurate customer support. At the same time, because the system relies on existing data accumulation, these question-and-answer pairs will gradually accumulate over time, and the intelligence level of the knowledge database will continue to improve as the amount of data increases. The matching effect of the question-and-answer database will also become more accurate, ultimately forming a comprehensive and rich knowledge database covering common industry questions and answers. In addition, the system can continuously optimize itself through deep learning technology, ensuring that customer service representatives can quickly find corresponding solutions when faced with new customer questions. In addition, the knowledge question-and-answer database can also be used as training materials for new employees, helping them to familiarize themselves with the business and answer customer questions more quickly. Ultimately, this automated knowledge base will become a key tool for improving customer satisfaction and business operational efficiency.

[0130] It's understandable that matching answers to the knowledge Q&A database significantly improves customer service response speed and accuracy. This not only reduces the workload of human customer service representatives but also ensures standardized and consistent service quality. Furthermore, this approach to building a Q&A database based on historical conversation data can save companies significant time and costs associated with manual compilation.

[0131] Based on the above question-answer determination method, this application also provides a question-answer determination device. Figure 6 The device is described in detail.

[0132] Figure 6 The structural block diagram of the question and answer determination device according to an embodiment of the present application is schematically shown.

[0133] like Figure 6 As shown, the question and answer determination device 600 of this embodiment includes an acquisition module 610 , a question determination module 620 , an answer determination module 630 and a feedback module 640 .

[0134] The acquisition module 610 is used to respond to the customer's dialogue request and, with the customer's consent or authorization, acquire the customer's speech within a preset time window. In one embodiment, the acquisition module 610 can be used to perform the operation S210 described above, which will not be repeated here.

[0135] The question determination module 620 is configured to identify the question boundary of each question in the customer's utterance using a pre-trained question-answering model based on the customer's utterance, and to obtain all question statements based on each question boundary. In one embodiment, the question determination module 620 can be configured to perform operation S220 described above, and will not be further described here.

[0136] The answer determination module 630 is used to determine the answer statement in the customer service speech that matches the question statement based on all the question statements using the question-answer model. In one embodiment, the answer determination module 630 can be used to perform the operation S230 described above, which will not be repeated here.

[0137] The feedback module 640 is used to feed back the answer statements corresponding to each of the question statements to the customer. In one embodiment, the feedback module 640 can be used to perform the operation S240 described above, which will not be repeated here.

[0138] According to an embodiment of the present application, the question determination module 620 is specifically used to: divide the customer's speech into a first sentence sequence; concatenate the first sentence sequence with the task identifier, and input the result into the target encoder of the question-answering model to obtain a hidden representation of each sentence in the first sentence sequence; based on the hidden representation, model the temporal dependency relationship between each of the sentences to generate a temporal enhanced context representation; and perform sequence labeling on the context representation, predict the boundary label of each of the sentences, and merge the sentences corresponding to adjacent boundary labels to determine the problem boundary of each of the questions.

[0139] According to an embodiment of the present application, the answer determination module 630 is specifically used to: splice each identified question statement with the second sentence sequence in the customer service utterance to obtain a question-answer pair input sequence; input the question-answer pair input sequence into the target encoder to obtain a joint hidden representation of the question statement and the customer service utterance; map all sentence representations in the joint hidden representation into sentence nodes, and construct edges between the sentence nodes according to lexical relevance and semantic relevance, respectively, to obtain a lexical graph and a semantic graph; perform graph convolution on the lexical graph and the semantic graph respectively, and interactively fuse them through double affine transformation to obtain fused features; and based on the fused features, use the question-answer model to predict the answer statement corresponding to the question statement.

[0140] According to an embodiment of the present application, the answer determination module 630 can also be used to: when a sentence node in the question statement and a sentence node in the customer service discourse contain at least one identical keyword, establish an edge between the sentence nodes to form the vocabulary graph; and based on the joint hidden representation, calculate the attention matrix between the sentence nodes through a multi-head attention mechanism, and retain the edges in the attention matrix whose attention scores are higher than a first preset threshold to form the semantic graph.

[0141] According to an embodiment of the present application, the question determination module 620 can also train the question-answering model, including: forming a positive sample pair with the question statement and the corresponding true answer statement, and forming a negative sample pair with the question statement and the answer statements of other questions in the same training batch; calculating the similarity score between the positive sample pair and the negative sample pair; and calculating the supervised contrast loss based on the similarity score.

[0142] According to an embodiment of the present application, the question determination module 620 can also be used to train the question-answering model, and also includes: performing weighted summation of the sequence labeling loss, the question-answer pair matching loss and the supervised contrast loss as the total loss function of the question-answering model, wherein the sequence labeling loss is used to characterize the loss of identifying the question boundary, and the question-answer pair matching loss is used to characterize the loss of determining the answer statement that matches the question statement; based on the total loss function, adjusting the parameters of the question-answering model until the total loss function meets the preset convergence condition.

[0143] According to an embodiment of the present application, the feedback module 640 can also be used to: store each of the question statements and the corresponding answer statements in the knowledge question and answer base; and in response to a customer inputting a new question statement, match the corresponding answer statement in the knowledge question and answer base and feedback it to the customer.

[0144] According to an embodiment of the present application, any multiple modules among the acquisition module 610, the question determination module 620, the answer determination module 630, and the feedback module 640 can be combined into a single module, or any one of them can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in a single module. According to an embodiment of the present application, at least one of the acquisition module 610, the question determination module 620, the answer determination module 630, and the feedback module 640 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented in hardware or firmware by any other reasonable means of integrating or packaging circuits, or can be implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of them. Alternatively, at least one of the acquisition module 610 , the question determination module 620 , the answer determination module 630 , and the feedback module 640 may be at least partially implemented as a computer program module, which may perform corresponding functions when executed.

[0145] Figure 7 A block diagram of an electronic device suitable for implementing the question-answer determination method according to an embodiment of the present application is schematically shown.

[0146] like Figure 7 As shown, an electronic device 700 according to an embodiment of the present application includes a processor 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage unit 708 into a random access memory (RAM) 703. The processor 701 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 701 may also include onboard memory for caching purposes. The processor 701 may include a single processing unit or multiple processing units for performing different actions of the method flow according to the embodiment of the present application.

[0147] Various programs and data required for the operation of the electronic device 700 are stored in the RAM 703. The processor 701, ROM 702, and RAM 703 are connected to each other via a bus 704. The processor 701 performs various operations of the method flow according to the embodiment of the present application by executing the programs in the ROM 702 and / or RAM 703. It should be noted that the programs may also be stored in one or more memories other than the ROM 702 and the RAM 703. The processor 701 may also perform various operations of the method flow according to the embodiment of the present application by executing the programs stored in the one or more memories.

[0148] According to an embodiment of the present application, electronic device 700 may further include an input / output (I / O) interface 705, which is also connected to bus 704. Electronic device 700 may also include one or more of the following components connected to I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including devices such as a cathode ray tube (CRT), liquid crystal display (LCD), and speakers; a storage section 708 including a hard disk; and a communication section 709 including a network interface card such as a LAN card or modem. Communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to I / O interface 705 as needed. Removable media 711, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed in drive 710 as needed, so that computer programs read from the removable media can be installed into storage section 708 as needed.

[0149] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments, or may exist independently and not be incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, and when the one or more programs are executed, the method according to the embodiments of this application is implemented.

[0150] According to an embodiment of the present application, a computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present application, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present application, a computer-readable storage medium may include the ROM 702 and / or RAM 703 described above and / or one or more memories other than ROM 702 and RAM 703.

[0151] The embodiments of the present application also include a computer program product, which includes a computer program containing program code for executing the method shown in the flowchart. When the computer program product is executed in a computer system, the program code is used to cause the computer system to implement the question-answer determination method provided in the embodiments of the present application.

[0152] The computer program executes the above functions defined in the system / device of the embodiment of the present application when the computer program is executed by the processor 701. According to the embodiment of the present application, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0153] In one embodiment, the computer program may be stored on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may be transmitted and distributed in the form of a signal on a network medium, downloaded and installed via the communication portion 709, and / or installed from a removable medium 711. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0154] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 709, and / or installed from a removable medium 711. When the computer program is executed by the processor 701, the above-described functions defined in the system of the embodiment of the present application are performed. According to the embodiment of the present application, the systems, devices, means, modules, units, etc. described above can be implemented by computer program modules.

[0155] According to an embodiment of the present application, the program code for executing the computer program provided by the embodiment of the present application can be written in any combination of one or more programming languages. Specifically, these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C" or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect via the Internet).

[0156] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of the boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0157] Those skilled in the art will appreciate that the features described in the various embodiments of this application may be combined and / or coupled in various ways, even if such combinations or couplings are not explicitly described in this application. In particular, the features described in the various embodiments of this application may be combined and / or coupled in various ways without departing from the spirit and teachings of this application. All such combinations and / or couplings fall within the scope of this application.

Claims

1. A question-answer determination method, characterized in that: The method comprises: In response to a customer's request for a conversation, and with the customer's consent or authorization, obtaining the customer's speech within a preset time window; Based on the customer's utterance, using a pre-trained question-answering model, the question boundary of each question in the customer's utterance is identified, and all question statements are obtained based on each of the question boundaries; Based on all the question statements, using the question-answering model, determining an answer statement in the customer service utterance that matches the question statement; and The answer statements corresponding to the question statements are fed back to the customer.

2. The method according to claim 1, characterized in that The step of identifying the question boundary of each question in the customer's utterance using a pre-trained question-answering model based on the customer's utterance includes: Dividing the customer's speech into a first sentence sequence; After concatenating the first sentence sequence with the task identifier, the concatenated sentence is input into the target encoder of the question-answering model to obtain a hidden representation of each sentence in the first sentence sequence; Based on the hidden representation, modeling the temporal dependency between the sentences to generate a temporally enhanced contextual representation; and The context representation is sequence labeled, a boundary label of each of the sentences is predicted, and sentences corresponding to adjacent boundary labels are merged to determine the problem boundary of each of the questions.

3. The method according to claim 2, characterized in that The step of determining, based on all the question statements, an answer statement in the customer service utterance that matches the question statement using the question-answering model includes: Concatenate each identified question statement with the second statement sequence in the customer service utterance to obtain a question-answer pair input sequence; Inputting the question-answer pair input sequence into the target encoder to obtain a joint hidden representation of the question sentence and the customer service utterance; Mapping all sentence representations in the joint hidden representation into sentence nodes, and constructing edges between the sentence nodes according to lexical relevance and semantic relevance, respectively, to obtain a lexical graph and a semantic graph; Performing graph convolution on the vocabulary graph and the semantic graph respectively, and interactively fusing them through double affine transformation to obtain fused features; and Based on the fusion features, the question-answering model is used to predict an answer statement corresponding to the question statement.

4. The method according to claim 3, characterized in that The steps of constructing edges between the sentence nodes according to lexical relevance and semantic relevance to obtain a lexical graph and a semantic graph include: When a sentence node in the question statement and a sentence node in the customer service utterance contain at least one identical keyword, establishing an edge between the sentence nodes to form the vocabulary graph; and Based on the joint hidden representation, the attention matrix between the sentence nodes is calculated through a multi-head attention mechanism, and the edges in the attention matrix with attention scores higher than a first preset threshold are retained to form the semantic graph.

5. The method according to claim 1, wherein Training the question-answering model, including: The question statement and the corresponding true answer statement form a positive sample pair, and the question statement and the answer statement of other questions in the same training batch form a negative sample pair; Calculating a similarity score between the positive sample pair and the negative sample pair; Based on the similarity scores, a supervised contrastive loss is calculated.

6. The method according to claim 5, characterized in that Training the question-answering model further includes: A weighted sum of the sequence labeling loss, the question-answer pair matching loss, and the supervised contrast loss is used as the total loss function of the question-answering model, wherein the sequence labeling loss is used to represent the loss of identifying the question boundary, and the question-answer pair matching loss is used to represent the loss of determining the answer statement that matches the question statement; Based on the total loss function, the parameters of the question-answering model are adjusted until the total loss function meets a preset convergence condition.

7. The method according to claim 1 or 2, characterized in that The method further comprises: Storing each of the question statements and the corresponding answer statements in a knowledge question and answer database; and In response to a new question statement input by a customer, a corresponding answer statement is matched in the knowledge question and answer database and fed back to the customer.

8. A question-answer determination device, characterized in that: The device comprises: An acquisition module, configured to respond to a conversation request from a customer and, with the customer's consent or authorization, acquire the customer's speech within a preset time window; A question determination module is used to identify the question boundary of each question in the customer's speech using a pre-trained question-answering model based on the customer's speech, and obtain all question statements based on each question boundary; An answer determination module is used to determine, based on all the question statements, an answer statement in the customer service utterance that matches the question statement using the question-answer model; and The feedback module is used to feed back the answer statements corresponding to each of the question statements to the customer.

9. An electronic device comprising: one or more processors; a memory for storing one or more computer programs, It is characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images