Quick feedback method for user questions and related equipment

Abstract

The invention discloses a user question-oriented rapid feedback method and related equipment, belongs to the technical field of artificial intelligence, and aims at effectively improving the shopping experience by utilizing large-scale question and answer data generated by users to assist the learning of questions and related comment tasks. The rapid feedback algorithm based on the attention model has a remarkable effect in solving the problem positioning in the shopping process, and is superior to similar algorithms. According to the method, the user can be helped to quickly find the answer, the shopping efficiency is improved, and the shopping experience of the user can be further improved. In addition, the purchase demand contained in the questions of the user also provides beneficial information for commodity recommendation, so that the comprehensiveness and individuation degree of shopping experience can be further improved by performing commodity recommendation according to the questions of the user.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a quick feedback method for user questions and related equipment. Background Art

[0002] With the popularity of online shopping, consumers have become accustomed to virtual shopping, but physical store shopping provides real-life experiences that virtual shopping cannot provide, such as touching products, consulting shopping guides, and trying on clothes. However, in the process of online shopping, consumers may feel a lack of trust because they cannot directly touch the products. To make up for this shortcoming, some e-commerce websites have designed online question-and-answer systems, but this requires users to wait for others to reply. On the other hand, traditional product question-and-answer research mainly focuses on unsupervised heuristic methods, which often have difficulty handling colloquial and diverse review texts. In recent years, with the growth of the scale of question-and-answer data on e-commerce websites, some supervised product question-and-answer methods have also begun to emerge, but these methods still fail to directly identify answers from reviews. At the same time, current methods also have the problem of difficulty in obtaining large-scale training data directly through annotation. For example, the focus in the question is different from the aspects and keywords in the field of review mining, because the user's focus is more complex and there are great differences in granularity. Therefore, the existing aspect / topic extraction method is not suitable for product question-and-answer scenarios. In addition, the simple word mapping method cannot handle complex and context-dependent review questions. Experimental results show that this method is indeed effective and its performance is significantly better than similar algorithms. Summary of the invention

[0003] The present invention provides a quick feedback method for user questions and related equipment, which solves the problem that the existing aspect / topic extraction method is not suitable for commodity question-answering scenarios.

[0004] To achieve the above object, the present invention provides the following technical solutions: Quick feedback methods for user questions, including: Abstract questions, answers and comments into word sequences, and convert each word into a corresponding word vector and character vector. Then use a bidirectional recurrent neural network to obtain the character vector of each word, and concatenate the word vector and character vector and input them into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers and comments. Obtain the implicit state vector of the word sequence based on the low-level hot signature, and store the implicit state vector in the implicit state matrix. Calculate the association matrix between the hidden state matrices of the question and the answer, and use the association matrix to obtain the attention vector. Then multiply the attention vector with the corresponding hidden state matrix to obtain the attentive feature vector. At the same time, a self-attention mechanism is introduced on the comment sentence to capture the content related to the focus point. Nonlinear mapping is used to fuse the attended feature vectors to generate high-level feature vectors. For auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, the high-level feature vector and the attended feature vector are further fused to obtain a new high-level feature vector. The final output value is calculated based on the new high-level feature vector to determine whether the comment statement can answer the user's question.

[0005] Preferably, the attention mechanism includes a bidirectional attention mechanism and a self-attention mechanism, the bidirectional attention mechanism is used to calculate the association mapping between the question and the answer, and the self-attention mechanism is used to capture the content related to the focus point in the comment sentence.

[0006] Preferably, the calculation process of the bidirectional attention mechanism includes: calculating the association matrix G, performing column summation and row summation calculations on the association matrix, and then normalizing the calculation results through softmax to obtain an attention vector.

[0007] Preferably, the calculation process of the self-attention mechanism includes: calculating the attention vector and the attended feature vector of the comment sentence, and introducing a regularization term in the loss function of the model to help the self-attention mechanism better capture the focus and answer pattern.

[0008] Preferably, nonlinear mapping is used to fuse the attentive feature vectors, and the step of generating a high-level feature vector specifically comprises fusing the feature vectors of the question and the answer, and fusing the feature vectors of the question and the comment statement to obtain a high-level feature vector.

[0009] Preferably, a multi-task learning method is adopted for the auxiliary task and the main task, and the main task and the auxiliary task are processed simultaneously.

[0010] Preferably, a performance evaluation step is also included to evaluate the performance in the product question answering task through precision, recall rate, F1 value and PR curve, and compare it with the existing methods to verify the superiority of the algorithm in capturing the semantic correlation between user questions and related comment statements.

[0011] A quick feedback system for user questions, including: Text encoding module: used to abstract questions, answers and comments into word sequences, and convert each word into a corresponding word vector and character vector. Then, a bidirectional recurrent neural network is used to obtain the character vector of each word. The word vector and character vector are concatenated and input into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers and comments. The implicit state vector of the word sequence is obtained based on the low-level hot evidence, and the implicit state vector is stored in the implicit state matrix. Attention mechanism module: used to calculate the association matrix between the hidden state matrices of the question and the answer, and obtain the attention vector through the association matrix. The attention vector is then multiplied by the corresponding hidden state matrix to obtain the attentive feature vector. At the same time, a self-attention mechanism is also introduced on the comment sentence to capture the content related to the focus point. Feedback module: used to use nonlinear mapping to fuse the attended feature vectors to generate high-level feature vectors. For auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, the high-level feature vector and the attended feature vector are further fused to obtain a new high-level feature vector. The final output value is calculated based on the new high-level feature vector to determine whether the comment statement can answer the user's question.

[0012] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a fast feedback method for user questions are implemented.

[0013] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a fast feedback method for user questions.

[0014] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a fast feedback method for user questions. By using large-scale user-generated question and answer data to assist in the learning of questions and related comment tasks, the model can effectively improve the shopping experience. The fast feedback algorithm based on the attention model has a significant effect in solving problem location in the shopping process, which is better than similar algorithms. This method can not only help users find answers quickly and improve shopping efficiency, but also further enhance the user's shopping experience. In addition, the purchase needs contained in the user's questions also provide useful information for product recommendations. Therefore, product recommendations based on user questions can further enhance the comprehensiveness and personalization of the shopping experience.

[0015] Furthermore, by utilizing a large amount of user-generated question-answer data and a small amount of manually annotated questions and related comments data, a multi-task deep learning model is proposed to identify sentences related to questions from product reviews. This multi-task learning method can handle the main task and auxiliary tasks at the same time, thereby improving the generalization ability and efficiency of the model.

[0016] Furthermore, attention mechanisms, including bidirectional attention mechanisms and self-attention mechanisms, are introduced into the model design to capture the correlation between texts. These attention mechanisms can help the model better understand the focus of questions, answer patterns, and key information in comments, thereby improving the model's representation ability and performance.

[0017] Furthermore, by designing multiple sub-networks in the model and sharing part of the network structure, feature fusion and information sharing between different tasks are achieved. This design allows the model to learn useful semantic information from auxiliary tasks and transfer it to the main task, thereby improving the model's effectiveness in identifying the semantic relationship between questions and related comments.

[0018] Furthermore, a regularization term is introduced into the loss function of the model to help the self-attention mechanism better capture the focus and answer patterns. This regularization can effectively control the complexity of the model and improve the generalization ability and stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the quick feedback method for user questions of the present invention; Figure 2 This is the QAR-net model structure of the embodiment of the present invention; Figure 3 PR curve comparison of the embodiment of the present invention; Figure 4 This is a block diagram of the quick feedback system for answering user questions of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, the present invention provides a quick feedback method for user questions S101 abstracts questions, answers, and comments into word sequences, and converts each word into a corresponding word vector and character vector. It then uses a bidirectional recurrent neural network to obtain the character vector of each word, and concatenates the word vector and character vector and inputs them into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers, and comments. It then obtains the implicit state vector of the word sequence based on the low-level hot signature and stores the implicit state vector in the implicit state matrix. S102 calculates the association matrix between the implicit state matrices of the question and the answer, and obtains the attention vector through the association matrix, and then multiplies the attention vector with the corresponding implicit state matrix to obtain the attentive feature vector, and at the same time introduces a self-attention mechanism on the comment sentence to capture the content related to the focus point; S103 utilizes nonlinear mapping to fuse the attentive feature vectors to generate high-level feature vectors. For the auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, the high-level feature vectors and the attentive feature vectors are further fused to obtain new high-level feature vectors. The final output value is calculated based on the new high-level feature vectors to determine whether the comment statement can answer the user's question.

[0024] The detailed steps are: A fast feedback algorithm for user questions (QAR-net model structure is as follows Figure 2 As shown in Figure 2, the model can be divided into four parts, from bottom to top: text encoding, attention mechanism, fusion operation and output layer: Step 1: Text encoding Step 1-1: First, abstract the question, answer, and review sentence into word sequences, namely , , Since Q, A, and R all use the same text encoding operation, the superscripts are omitted in the text, and a word sequence is simply expressed as . Each word in the word sequence First, it is converted into the corresponding word vector and character vectors This paper obtains the word vector corresponding to each word based on the pre-trained GLoVE word vector list.

[0025] Step 1-2: Then for each word Use a bidirectional recurrent neural network (Bi-RNN) to obtain the vector corresponding to its last hidden state as the character vector of the word . Character vectors can solve the out-of-vocabulary problem of word vectors. and Spliced ​​together to represent , and then input the word vector sequence into the Bi-GRU structure to extract the low-level features of the text. The calculation formula of the Bi-GRU operation is as follows: .

[0026] Where ⊕ represents the concatenation operation of vectors, and represents the forward and reverse implicit vectors of Bi-GRU. We concatenate the two to get the word The hidden state vector The hidden state incorporates the contextual information of the current word.

[0027] Step 1-3: Store the hidden state vector corresponding to the entire word vector sequence in a hidden state matrix. The hidden state matrices corresponding to Q, A and R are , and .

[0028] Step 2: Attention Mechanism Step 2-1: and is the input and the output is a fixed-length feature vector and . Arrows in superscripts indicate the direction of attention, e.g. Represents Q's attention to A. The first step of the bidirectional attention mechanism is to calculate the correlation matrix G, which is calculated as follows: .

[0029] in, is a parameter matrix. This formula can be used to obtain the association mapping between the implicit state matrices of Q and A. Elements represent and The similarity between .

[0030] Step 2-2: Calculate the column and row sums of the correlation matrix, and then normalize the calculation results through softmax to obtain the attention vector and , the calculation formula is as follows: . . . . Note: The sum operation is used here instead of the maximum value operation. The reason is that the questions in product Q&A are highly subjective, and some focus points may correspond to multiple words related to the answer. The sum operation can better reflect the importance of the word sequence as a whole.

[0031] Step 2-3: Multiply the attention vector using Q and A with the corresponding hidden state matrix to obtain the attentive feature vector. The calculation process is as follows: .

[0032] . . Step 2-4: A self-attention mechanism is also introduced on R-subnet to capture the content related to the focus in the comment sentence R. Its attention vector and the attentive feature vector The calculation method is as follows: . . . . in, and are the parameter matrix and parameter vector of the self-attention mechanism, respectively, and k is a hyperparameter. At the same time, in order to make the self-attention mechanism better capture the relevant content of the focus in the comment R, we also designed a corresponding regularization term.

[0033] Step 3: Fusion Operation and Output Layer Step 3-1: A fusion operation will be used to generate a high-level feature vector of QA or QR. The input feature vector is fused using nonlinear mapping, because this fusion method has been shown to have good performance in natural language processing tasks. and The fusion operation is as follows: . . in, and is a 2u×2u parameter matrix. And Finally, it is used for QA tasks (auxiliary tasks). The final output value calculation method is as follows: . in, is the sigmoid function.

[0034] Step 3-2: For the QR task (main task), it is necessary to further integrate the high-level feature vectors and the attentive feature vector , which is as follows: . . The final output value calculation process is as follows: .

[0035] Another embodiment of the present invention provides a quick feedback method for user questions. Experimental dataset: (1) We selected two product categories from the public Amazon dataset: “Electronics” and “Mobile Phones and Accessories”, and collected QA data from these two categories. For each question, the answer with the most votes was selected to form the QA positive sample together with the question. We also selected two answers from other products to form the negative samples together with the question. In the end, we collected a total of 100,139 QA data.

[0036] (2) For the comment data, we used the NLTK tool to segment the comments. Then we randomly selected 10,000 questions from the two product categories and combined these questions with the comment sentences to obtain a large number of QR pairs, which were then manually annotated. However, direct annotation is not reasonable because QR positive samples are very rare. In order to improve the efficiency of annotation, a filtering strategy was also designed. This included steps such as word frequency extraction, word frequency ratio calculation, and threshold setting, and finally 50,000 QR data were obtained. Subsequently, we randomly selected a part of the QR pairs for annotation, and each data was annotated by 3 professional annotators with a consistency rate of 86%. In the end, we obtained 7,466 annotated data, with a positive and negative sample ratio of 3:7. These data covered 5,907 different questions and were divided into a training set (4,000), a validation set (977), and a test set (2,489) in the same proportion.

[0037] 1. Experimental Details We use the Adam algorithm to train the model. The initial learning rate is set to 1e-6, and the first and second momentum factors are set to 0.9 and 0.999 respectively. We add dropout (0.5) to the output of Bi-GRU to prevent overfitting. For fair comparison experiments, we set the hyperparameters of all methods to the same value: the length of the character vector is set to 20; the length of the GRU hidden state vector u=32. The batch size (mini-batch size) is set to 128. All hyperparameter tuning experiments are conducted on the validation set. In addition, we use a random downsampling strategy for the majority class in the data. Different from the classic one-shot downsampling strategy, we downsample the training data in each round. In each round of QA / QR training, we sample the same number of negative samples as positive samples to form a 1:1 training data for training. This training strategy can weaken the negative impact caused by the information loss

[164] in the one-shot downsampling method. For fair comparison, all comparison methods use this downsampling strategy to deal with the sample imbalance problem. In addition, all multi-task learning methods use a transfer training strategy. The number of pre-training rounds is adjusted on the QR validation set, and the number of training rounds in the fine-tuning process is determined by the early stopping strategy.

[0038] 2. Results Analysis QREM: Question Review Embedding Matching (QREM). First, the question and comment sentences are converted into a word embedding matrix based on the word embedding list, and then the maximum, minimum and mean values ​​of each row of the matrix are calculated to generate a fixed-length feature vector. For a given Q and R, the cosine similarity scores of the two are calculated, and then a logistic regression classifier is trained to determine whether R can answer Q.

[0039] QAR-net: Our proposed method.

[0040] QAR-net-QR: Only use QR data to train QAR-net.

[0041] QA-net-QA: This comparison method only uses the QA-net on the left side of QAR-net, i.e. Figure 2 The network structure in the dashed box of the auxiliary task. This method is very similar to the network structure used in the best deep methods in the question answering field. We only used QA data to train this network.

[0042] QA-net-QR: This comparison method only uses QR data to train QA-net.

[0043] QA-net-MTL: This method adopts the QA-net network structure, targets QA tasks and QR tasks, and adopts a traditional parameter hard-sharing multi-task learning strategy

[165] .

[0044] CQA-hard: A multi-task learning based model for social question answering

[166] . We use a mirrored CNN network structure to process QA and QR data. The features extracted by CNN are input into a task-shared fully connected layer to perform the final prediction task.

[0045] NLP-soft: This is a soft-shared multi-task model that was originally used for cross-language dependency relationship analysis tasks.

[0046] Although there are abundant positive samples in QA data, there are few positive samples in QR data. Since product question answering mainly focuses on positive samples, we use precision, recall, F1 value and PR curve to evaluate the above methods.

[0047] Table 1 Performance comparison

[0048] The performance comparison of QAR-net and similar algorithms is shown in Table 1. QREM has the worst performance, which shows that simple pooling and concatenation operations on text word embeddings cannot handle product question answering problems well. QA-net-QA has the lowest accuracy among all deep learning methods. The reason is that this method is not trained for QR tasks, so it cannot enjoy the performance improvement brought by using QR data to train the model. However, the higher recall rate of this method shows that there are common answer patterns between the two tasks. The performance of QA-net-QR and QAR-net-QR is lower than that of QAR-net. This shows that simply using QR data with few positive examples cannot accurately capture the correlation pattern between questions and comment sentences. The performance of QA-net-MTL is better than that of QA-net-QA but the F1 value is worse than that of QA-net-QR. This means that the hard-shared multi-task learning method cannot make good use of QA data to improve the effect of QR tasks. The F1 values ​​of the other two multi-task methods (CQA-hard and NLP-soft) are the worst among all deep learning methods. The reasons are: 1) these two methods do not use the attention mechanism; 2) the parameter sharing method used in these two models is not suitable for the problem in this paper. Thanks to the sophisticated model design, the introduction of the attention mechanism and the reasonable setting of the regularization term, our proposed method (QAR-net) achieves the best performance. In order to evaluate the ability of the above methods to rank answers before non-answers, we plotted the PR curves of all methods, such as Figure 3As shown in the figure, we can see that the performance of our proposed method is significantly better than other methods.

[0049] The present invention also provides a quick feedback system for user questions, including: Text encoding module: used to abstract questions, answers and comments into word sequences, and convert each word into a corresponding word vector and character vector. Then, a bidirectional recurrent neural network is used to obtain the character vector of each word. The word vector and character vector are concatenated and input into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers and comments. The implicit state vector of the word sequence is obtained based on the low-level hot signature, and the implicit state vector is stored in the implicit state matrix. Attention mechanism module: used to calculate the association matrix between the hidden state matrices of the question and the answer, and obtain the attention vector through the association matrix. The attention vector is then multiplied by the corresponding hidden state matrix to obtain the attentive feature vector. At the same time, a self-attention mechanism is also introduced on the comment sentence to capture the content related to the focus point. Feedback module: It is used to fuse the attentive feature vectors using nonlinear mapping to generate high-level feature vectors. For auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, further integration and A new high-level feature vector is obtained, and a final output value is calculated based on the new high-level feature vector to determine whether the comment statement can answer the user's question.

[0050] A terminal device is provided in one embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0051] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to accomplish the present invention.

[0052] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0053] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0054] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0055] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

Claims

1. A quick feedback method for user questions, characterized in that: include: Abstract questions, answers and comments into word sequences, and convert each word into a corresponding word vector and character vector. Then use a bidirectional recurrent neural network to obtain the character vector of each word, and concatenate the word vector and character vector and input them into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers and comments. Obtain the implicit state vector of the word sequence based on the low-level hot signature, and store the implicit state vector in the implicit state matrix. Calculate the association matrix between the hidden state matrices of the question and the answer, and use the association matrix to obtain the attention vector. Then multiply the attention vector with the corresponding hidden state matrix to obtain the attentive feature vector. At the same time, a self-attention mechanism is introduced on the comment sentence to capture the content related to the focus point. Nonlinear mapping is used to fuse the attended feature vectors to generate high-level feature vectors. For auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, the high-level feature vector and the attended feature vector are further fused to obtain a new high-level feature vector. The final output value is calculated based on the new high-level feature vector to determine whether the comment statement can answer the user's question.

2. The rapid feedback method for user questions according to claim 1 is characterized in that: The attention mechanism includes a bidirectional attention mechanism and a self-attention mechanism. The bidirectional attention mechanism is used to calculate the association mapping between questions and answers, and the self-attention mechanism is used to capture content related to the point of interest in the comment sentence.

3. The rapid feedback method for user questions according to claim 2 is characterized in that: The calculation process of the bidirectional attention mechanism includes: calculating the association matrix G, performing column sum and row sum calculations on the association matrix, and then normalizing the calculation results through softmax to obtain an attention vector.

4. The rapid feedback method for user questions according to claim 2, characterized in that: The calculation process of the self-attention mechanism includes: calculating the attention vector and the attentive feature vector of the comment sentence, and introducing a regularization term in the loss function of the model to help the self-attention mechanism better capture the focus and answer pattern.

5. The rapid feedback method for user questions according to claim 1, characterized in that: The steps of using nonlinear mapping to fuse the attentive feature vectors and generate high-level feature vectors are specifically to fuse the feature vectors of questions and answers, and to fuse the feature vectors of questions and comment sentences to obtain high-level feature vectors.

6. The rapid feedback method for user questions according to claim 1, characterized in that: A multi-task learning method is adopted for auxiliary tasks and main tasks, processing the main tasks and auxiliary tasks at the same time.

7. The rapid feedback method for user questions according to claim 1, characterized in that: It also includes a performance evaluation step, which evaluates the performance in the product question answering task through precision, recall, F1 value and PR curve, and compares it with existing methods to verify the superiority of this algorithm in capturing the semantic correlation between user questions and related comment statements.

8. A quick feedback system for user questions, characterized by: include: Text encoding module: used to abstract questions, answers and comments into word sequences, and convert each word into a corresponding word vector and character vector. Then, a bidirectional recurrent neural network is used to obtain the character vector of each word. The word vector and character vector are concatenated and input into a bidirectional gated recurrent unit structure to extract low-level features of questions, answers and comments. The implicit state vector of the word sequence is obtained based on the low-level hot signature, and the implicit state vector is stored in the implicit state matrix. Attention mechanism module: used to calculate the association matrix between the hidden state matrices of the question and the answer, and obtain the attention vector through the association matrix. The attention vector is then multiplied by the corresponding hidden state matrix to obtain the attentive feature vector. At the same time, a self-attention mechanism is also introduced on the comment sentence to capture the content related to the focus point. Feedback module: used to use nonlinear mapping to fuse the attended feature vectors to generate high-level feature vectors. For auxiliary tasks, the final output value is calculated based on the high-level feature vectors to determine whether the comment statement can answer the user's question. For the main task, the high-level feature vector and the attended feature vector are further fused to obtain a new high-level feature vector. The final output value is calculated based on the new high-level feature vector to determine whether the comment statement can answer the user's question.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for rapid feedback to user questions as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for rapid feedback to user questions as claimed in any one of claims 1 to 7 are implemented.

Images