A question-answering text evaluation method, system and application

Abstract

The present invention belongs to the field of text evaluation technology, and specifically relates to a question-and-answer text evaluation method, system, and application. The present invention first extracts relevant text by matching investor questions with the corresponding year's MD&A through a sliding window. Secondly, a variational autoencoder and a bidirectional attention mechanism focus on relevant information in the text, mining semantic differences between questions and answers, and between answers and MD&A-related text. The bidirectional attention weighted representation and difference features are concatenated to obtain a contextual representation. The semantic differences between the question and answer text are evaluated based on the contextual representation, thereby accurately identifying key information related to stock price volatility. On this basis, the mined semantic differences are quantitatively analyzed, thereby achieving accurate prediction of stock price volatility.

Description

Technical Field

[0001] The present invention belongs to the technical field of text evaluation, and specifically relates to a question-and-answer text evaluation method, system and application. Background Art

[0002] Currently, stock price volatility prediction primarily relies on constructing predictive features from structured data, and then building predictive models. This data includes macroeconomic indicators, financial metrics, and stock market trading data. However, with the development of next-generation information technology, scholars have begun to focus on the role of sentiment embedded in textual data, such as conference call transcripts, social media posts, and Q&A texts from earnings conferences, in predicting stock volatility. Earnings conference Q&A texts, in particular, are considered to have potential for predicting stock price fluctuations, as they reflect investors' key concerns and allow for real-time, two-way interaction.

[0003] Although existing technologies have begun to exploit the utility of text data in predicting stock price volatility, they still face several challenges. First, the non-standard, emotional, and ambiguous information in spoken responses exposes text analysis methods to a high level of noise, making it difficult to accurately mine the semantic differences between questions and answers. Second, managers may selectively disclose some information and avoid negative details, making it difficult to accurately verify the authenticity and completeness of the information provided by relying solely on the response text. Furthermore, when combining management analysis and discussion content to mine for irrelevant answers, long texts present challenges such as large amounts of information, complex content, and strong subjectivity. Therefore, how to effectively mine semantic differences from question and answer texts and predict stock price volatility remains a challenge that current technologies need to address. Summary of the Invention

[0004] The purpose of the present invention is to provide a question-answer text evaluation method, system and application to solve the problems raised in the background technology.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0006] In a first aspect, the present invention proposes a method for evaluating question-answer texts, comprising:

[0007] S1. Obtain a topic text, a question text, and an answer text in a set question-and-answer scenario to form a question-and-answer dataset, wherein the question text is determined based on the topic text;

[0008] S2. Based on a preset text matching strategy, the question text and the subject text are matched for similarity, and a continuous sentence sequence with the highest similarity is generated as the matching text;

[0009] S3. Preprocess the matching text and the question-answer dataset using the BERT model to generate a text vector.

[0010] S4. Input the text vector into a variational autoencoder for decoding and encoding to obtain a denoised text vector;

[0011] S5. Inputting the denoised text vector into a bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence;

[0012] S6. Calculate the bidirectional attention weighted representation and difference features of the hidden state sequence using a bidirectional attention mechanism;

[0013] S7. Concatenate the bidirectional attention weighted representation and the difference feature to obtain a contextual representation, and evaluate the semantic difference between the question and answer text based on the contextual representation.

[0014] Furthermore, the set question-and-answer scenario includes a performance briefing of a listed company, the subject text includes a management discussion and analysis text, and the reply text includes a reply text of the highest-level management at the performance briefing of the listed company.

[0015] Furthermore, the S2 includes:

[0016] S201, deleting non-verbal symbols and numbers in the subject text;

[0017] S202: Using a sliding window-based text matching method, sliding a fixed-size window on the topic text to calculate the similarity between the question text and the topic text;

[0018] S203: Match the continuous sentence sequence with the highest similarity to the question text and use it as the matching text.

[0019] Furthermore, the preprocessing in S3 includes: performing word segmentation, removing stop words, and filtering low-frequency words on the matching text and the question-answer dataset respectively, to form a text vector containing contextual semantic information.

[0020] Furthermore, the S4 includes:

[0021] S401: Input the matching text and the text vector of the question-answering dataset into the encoder, perform linear transformation of the fully connected layer and apply the ReLU activation function to calculate the mean of the latent space. μ and log variance log(σ 2 ) , the latent variable z is obtained by sampling, ,in, is random noise sampled from a standard normal distribution;

[0022] S402: Obtain the denoised text vector based on the fully connected layer and activation function :

[0023] ;

[0024] ;

[0025] in, represents the output of the third hidden layer; represents the weight matrix in the third hidden layer; represents the bias vector of the third hidden layer; Represents the weight matrix in the fourth hidden layer, specifically connecting the hidden layer and text vectors The weight of Represents the bias vector of the fourth hidden layer.

[0026] Furthermore, the S5 includes: converting the denoised text vector into a deep semantic representation through a bidirectional long short-term memory network, obtaining its context information for each vocabulary unit, forming a contextual representation, and obtaining a hidden state sequence , , P= , where A corresponds to the question text, Q corresponds to the answer text, and P corresponds to the matching text.

[0027] Furthermore, the S6 includes:

[0028] S601, calculation sequence arrive ,sequence and ,sequence arrive and sequence A arrive The dot product of the transpose of is used to obtain the attention weight matrix; using Normalize the attention weight matrix so that the sum of the weights of each row is 1;

[0029] S602, the obtained normalized weight matrix and sequence Multiplying together, we get arrive Attention-weighted representation of ,get arrive Attention-weighted representation of , arrive Attention-weighted representation of arrive Attention-weighted representation of , as follows:

[0030] ;

[0031] S603, from the sequence Subtract the weighted Q(A) from the original to get the difference feature , as follows:

[0032] .

[0033] Furthermore, the S7 includes: 、 Splicing to get text representation ;

[0034] ;

[0035] Representing text Input to a fully connected layer, After converting the output into probability, we get the corresponding similarity probability between texts and the probability of difference , as follows:

[0036] ;

[0037] .

[0038] In a second aspect, the present invention provides a question-answer text evaluation system, which is applied to perform any of the above-mentioned evaluation methods, and the system includes:

[0039] The data acquisition module is used to obtain the discussion and analysis texts of listed companies' management, as well as the question and answer texts in the question and answer scenarios set by listed companies, to form a question and answer dataset;

[0040] A text matching module is used to perform similarity matching between the question text and the subject text based on a preset text matching strategy, and to generate a continuous sentence sequence with the highest similarity as the matching text;

[0041] A preprocessing module, configured to preprocess the matching text and the question-answering dataset using a BERT model to generate a text vector;

[0042] A variational autoencoder module is used to input the text vector into the variational autoencoder for decoding and encoding to obtain a denoised text vector;

[0043] A bidirectional long short-term memory network module is used to input the denoised text vector into the bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence;

[0044] a bidirectional attention mechanism module, configured to calculate a bidirectional attention weighted representation and a difference feature of the hidden state sequence using a bidirectional attention mechanism;

[0045] An evaluation module is used to concatenate the bidirectional attention weighted representation and the difference feature to obtain a context representation, and evaluate the semantic difference between the question and the answer text based on the context representation.

[0046] In a third aspect, the present invention proposes an application of a question-answer text evaluation method to stock price volatility prediction. The evaluation method described in any one of the above is used to obtain a text representation G, a similarity probability between texts, and a difference probability between texts of a listed company's performance briefing. The text representation G, the similarity probability between texts, and the difference probability between texts are then concatenated to obtain a distributed text representation S:

[0047] ;

[0048] The identity characteristics of executives of listed companies After being merged into the text representation S, it is normalized and then input into the fully connected layer of the encoder. The stock price volatility is predicted based on the output of the fully connected layer, as shown in the following formula:

[0049] ;

[0050] in, represents the weight matrix, represents the bias vector, Represents the model output results.

[0051] The beneficial effects of the present invention are:

[0052] This method combines Q&A text from listed companies' earnings conferences with relevant text from MD&A documents through data matching. Furthermore, a variational autoencoder is used to fill in missing information and reduce noise. The application of a bidirectional attention mechanism enables a more precise understanding of the context within the text, thereby accurately identifying key information related to stock price volatility. Based on this, the discovered semantic differences are quantitatively analyzed to achieve accurate predictions of stock price volatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of the evaluation method of the present invention;

[0054] Figure 2 Schematic diagram of the process of step S2 in the present invention;

[0055] Figure 3 Schematic diagram of the process of step S4 in the present invention;

[0056] Figure 4 Schematic diagram of the process of step S6 in the present invention;

[0057] Figure 5 This is a diagram of a stock price volatility prediction framework based on the question-and-answer text of a listed company's performance briefing, as proposed in Example 3 of the present invention. DETAILED DESCRIPTION

[0058] The present application is described in further detail below in conjunction with the accompanying drawings. It is necessary to point out that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technicians in this field can make some non-essential improvements and adjustments to the present application based on the above application content. Example

[0059] like Figure 1-4 As shown, this embodiment proposes a question-answer text evaluation method, which includes the following steps:

[0060] S1. Obtain a topic text, a question text, and an answer text in a set question-and-answer scenario to form a question-and-answer dataset, wherein the question text is determined based on the topic text;

[0061] As an example, we obtain the Management Discussion and Analysis (MD&A) text of a listed company, as well as the question text and answer text in a question-and-answer scenario set by the listed company, to form a question-and-answer dataset. The set question-and-answer scenario includes a performance briefing of the listed company, and the answer text is the answer text of the highest-level management at the performance briefing of the listed company.

[0062] S2. Based on a preset text matching strategy, the question text and the subject text are matched for similarity, and a continuous sentence sequence with the highest similarity is generated as the matching text;

[0063] S3. Preprocess the matching text and the question-answer dataset using the BERT model to generate a text vector.

[0064] S4. Input the text vector into a variational autoencoder for decoding and encoding to obtain a denoised text vector;

[0065] S5. Inputting the denoised text vector into a bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence;

[0066] S6. Calculate the bidirectional attention weighted representation and difference features of the hidden state sequence using a bidirectional attention mechanism;

[0067] S7. Concatenate the bidirectional attention weighted representation and the difference feature to obtain a contextual representation, and evaluate the semantic difference between the question and answer text based on the contextual representation.

[0068] It should be noted that the Management Discussion and Analysis (MD&A) of a listed company is an important part of the annual report of a listed company, requiring management to further explain and analyze important historical information in the company's current financial statements and notes, and to explain the next year's operating plan and the opportunities, challenges and various risks facing the company's future development from the perspective of the company's management.

[0069] According to the above steps S1-S7, the above evaluation method is mainly divided into three parts during its specific implementation, including data matching, data preprocessing and “irrelevant answer” scoring.

[0070] (1) In the data matching stage, this embodiment proposes a text matching method based on a sliding window, which matches the paragraph (matching text) that best matches the investor's question at the earnings conference by calculating the similarity between the investor's question and the MD&A. Preferably, S2 includes:

[0071] S201. Delete non-verbal symbols and numbers from the Management Discussion and Analysis text; and delete "Detailed Description Data" from the MD&A. This detailed description data in the MD&A may contain many non-verbal symbols and numbers. These elements may have different distribution characteristics in the vector space, resulting in vector sparseness and affecting subsequent calculations and matching.

[0072] S202: Using a sliding window-based text matching method, sliding a fixed-size window on the management discussion and analysis text to calculate the similarity between the question text and the management discussion and analysis text;

[0073] S203: Match the continuous sentence sequence with the highest similarity to the question text and use it as the matching text.

[0074] As an example, first, set the text length, that is, the window size , and segment the MD&A text and Q&A text into sentence units, i.e. and , and then calculate the similarity between each pair of sentences .

[0075] Next, we slide a fixed-size window over the MD&A text and calculate the sum of sentence similarities within each window, as shown in the following formula:

[0076] ;

[0077] Find the sequence of consecutive sentences with the highest similarity. This process is performed within the text length limit, and the text in the final selected window is the part that best matches the question and answer text.

[0078] (2) In the data preprocessing stage, the executives are first classified, and then the text is segmented, stop words and low-frequency words are removed, and the text is converted into an embedding vector using the BERT model. Preferably, the preprocessing in S3 includes: segmenting the matching text and the question-answer dataset, removing stop words, and filtering low-frequency words, respectively, to form a text vector containing contextual semantic information.

[0079] (3) Using a non-answer evaluation method based on VAE (variational autoencoder) and bidirectional attention mechanism to mine the relationship between texts and calculate the degree of non-answer. Preferably, S4 includes:

[0080] S401: Input the matching text and the text vector of the question-answering dataset into the encoder, perform linear transformation of the fully connected layer and apply the ReLU activation function to calculate the mean of the latent space. μ and log variance log(σ 2 ) , the latent variable z is obtained by sampling, ,in, is random noise sampled from a standard normal distribution;

[0081] S402: Obtain the denoised text vector based on the fully connected layer and activation function :

[0082] ;

[0083] ;

[0084] in, represents the output of the third hidden layer, usually representing an intermediate layer in the network; Represents the weight matrix in the third hidden layer. In neural networks, weights are parameters in the learning process used to transform input data to the next layer; Represents the bias vector of the third hidden layer. The bias is another important learning parameter in neural networks. It is added to the product of the weight and the input to provide an offset for each neuron. Represents the weight matrix in the fourth hidden layer, specifically connecting the hidden layer and text vectors The weight of Represents the bias vector of the fourth hidden layer.

[0085] During the training process, VAE optimizes two objectives simultaneously: one is the reconstruction loss, which is used to measure the difference between the decoder output data and the original data. In this paper, it is measured by cross entropy, as shown below:

[0086] ;

[0087] in, is the reconstruction error, N is the number of data samples, D is the size of the vocabulary, The original data The first sample The value of the feature, is the first The first sample The other is a regularization term, which is measured by KL divergence in this embodiment to ensure that the distribution of the latent space is close to the prior distribution, as shown in the following formula:

[0088] ;

[0089] in, is the KL divergence, is the approximate posterior distribution of the encoder output, is a latent variable The prior distribution of is usually assumed to be a standard normal distribution. is the dimension of the latent space. is a latent variable The mean of the j-th dimension of . is a latent variable The variance of the j-th dimension of .

[0090] By continuously adjusting the parameters of the encoder and decoder to minimize the above loss function, VAE can learn an effective representation of text data and capture the structure and semantic information of the text in the latent space. In order to further explore the semantic differences between texts, the two denoised text vectors are processed forward and backward by BiLSTM on the input sequence to capture the dependencies and information in the sequence. Preferably, the S5 includes: converting the denoised text vector into a deep semantic representation through a bidirectional long short-term memory network, obtaining its contextual information for each vocabulary unit, forming a contextualized representation, and obtaining a hidden state sequence. , , P= , where A corresponds to the question text, Q corresponds to the answer text, and P corresponds to the matching text.

[0091] The bidirectional attention mechanism simultaneously considers investor questions and executive responses, capturing subtle semantic differences between the two. By building relationships between questions and responses within the attention mechanism, we can more clearly understand which question points were under-discussed or completely ignored in the executive's responses, thereby identifying unrepresented information.

[0092] Preferably, the S6 includes:

[0093] S601, calculation sequence arrive ,sequence and ,sequence arrive and sequence A arrive The dot product of the transpose of is used to obtain the attention weight matrix; using Normalize the attention weight matrix so that the sum of the weights of each row is 1;

[0094] S602, the obtained normalized weight matrix and sequence Multiplying together, we get arrive Attention-weighted representation of ,get arrive Attention-weighted representation of , arrive Attention-weighted representation of arrive Attention-weighted representation of , as follows:

[0095] ;

[0096] S603, from the sequence Subtract the weighted Q(A) from the original to get the difference feature , as follows:

[0097] .

[0098] Further preferably, the S7 includes:

[0099] will sequence 、 Splicing to get text representation ;

[0100] ;

[0101] Representing text Input to a fully connected layer, After converting the output into probability, we get the corresponding similarity probability between texts and the probability of difference , as follows: ;

[0102] .

[0103] Based on the above-mentioned embodiment of the present invention, this embodiment proposes a method for evaluating question-and-answer text from listed companies' earnings presentations. This method first uses a sliding window to match investor questions with corresponding MD&A documents to extract relevant text. Secondly, a variational autoencoder and a bidirectional attention mechanism are used to focus on relevant information within the text. This method then identifies any irrelevant responses between questions and answers, and between answers and related MD&A text, and evaluates the Q&A text based on similarity and difference probabilities.

[0104] Example 2

[0105] This embodiment proposes a question-answer text evaluation system, which is applied to perform the above evaluation method. The system includes:

[0106] A data acquisition module is used to acquire a topic text, a question text, and an answer text in a set question-and-answer scenario to form a question-and-answer dataset, wherein the question text is determined based on the topic text;

[0107] A text matching module is used to perform similarity matching between the question text and the subject text based on a preset text matching strategy, and to generate a continuous sentence sequence with the highest similarity as the matching text;

[0108] A preprocessing module, configured to preprocess the matching text and the question-answering dataset using a BERT model to generate a text vector;

[0109] A variational autoencoder module is used to input the text vector into the variational autoencoder for decoding and encoding to obtain a denoised text vector;

[0110] A bidirectional long short-term memory network module is used to input the denoised text vector into the bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence;

[0111] a bidirectional attention mechanism module, configured to calculate a bidirectional attention weighted representation and a difference feature of the hidden state sequence using a bidirectional attention mechanism;

[0112] An evaluation module is used to concatenate the bidirectional attention weighted representation and the difference feature to obtain a context representation, and evaluate the semantic difference between the question and the answer text based on the context representation.

[0113] It should be noted here that each module in the above evaluation system corresponds to steps S1 to S7 in implementing the above evaluation method, and the instances and application scenarios implemented by multiple modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiment 1. Example

[0114] Combine Figure 5 This embodiment proposes an application of a question-and-answer text evaluation method to stock price volatility prediction. The evaluation method described in Example 1 is used to obtain the text representation G, the similarity probability between texts, and the difference probability between texts of a listed company's performance briefing. The text representation G, the similarity probability between texts, and the difference probability between texts are then concatenated to obtain a distributed text representation S:

[0115] ;

[0116] The identity characteristics of executives of listed companies After being merged into the text representation S, it is normalized and then input into the fully connected layer of the encoder. The stock price volatility is predicted based on the output of the fully connected layer, as shown in the following formula:

[0117] ;

[0118] in, represents the weight matrix, represents the bias vector, Represents the model output results.

[0119] More specifically, based on the nonlinear characteristics of the stock market volatility, which is characterized by a peak and a thick tail, this embodiment uses the natural logarithm form to quantify stock price volatility. , as shown below

[0120] ;

[0121] in Indicates time, Indicates the The rate of change of the day (corresponding to the output Y of the above fully connected layer), It represents the average rate of change during this period, using the volatility of the week after the performance statement occurs ( ).

[0122] According to the above-mentioned embodiments of the present invention, this embodiment proposes a question-answer text evaluation method based on Example 1 to predict stock price volatility, and realizes the prediction of stock price volatility using question-answer text. Specifically, based on the variational autoencoder and the bidirectional attention mechanism, it extracts irrelevant information and calculates the difference probability, and has strong predictive ability.

[0123] In another embodiment provided by the present invention, a computer-readable storage medium is also provided, which stores a computer program. When the computer program is executed by a processor, the steps of the evaluation method of the question and answer text in the above-mentioned embodiment 1 are implemented.

[0124] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is run on a computer, the computer executes the method for evaluating the question-and-answer text in the above-mentioned embodiment 1.

[0125] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part.

[0126] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0127] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] In addition, the functional modules in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0129] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for evaluating question-answer texts, characterized in that: include: S1. Obtain a topic text, a question text, and an answer text in a set question-and-answer scenario to form a question-and-answer dataset, wherein the question text is determined based on the topic text; S2. Based on a preset text matching strategy, the question text and the subject text are matched for similarity, and a continuous sentence sequence with the highest similarity is generated as the matching text; S3. Preprocess the matching text and the question-answer dataset using the BERT model to generate a text vector. S4. Input the text vector into a variational autoencoder for decoding and encoding to obtain a denoised text vector; S5. Inputting the denoised text vector into a bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence; S6. Calculate the bidirectional attention weighted representation and difference features of the hidden state sequence using a bidirectional attention mechanism; S7. Concatenate the bidirectional attention weighted representation and the difference feature to obtain a contextual representation, and evaluate the semantic difference between the question and answer text based on the contextual representation.

2. The method for evaluating a question-answer text according to claim 1, wherein: The set question-and-answer scenario includes a performance briefing of a listed company, the subject text includes a management discussion and analysis text, and the reply text includes a reply text of the highest-level management at the performance briefing of the listed company.

3. The method for evaluating a question-answer text according to claim 2, wherein: The S2 includes: S201, deleting non-verbal symbols and numbers in the subject text; S202: Using a sliding window-based text matching method, sliding a fixed-size window on the topic text to calculate the similarity between the question text and the topic text; S203: Match the continuous sentence sequence with the highest similarity to the question text and use it as the matching text.

4. The method for evaluating a question-answer text according to claim 2, wherein: The preprocessing in S3 includes: performing word segmentation, removing stop words, and filtering low-frequency words on the matching text and the question-answer dataset, respectively, to form a text vector containing contextual semantic information.

5. The method for evaluating a question-answer text according to claim 3, wherein: The S4 includes: S401: Input the matching text and the text vector of the question-answering dataset into the encoder, perform linear transformation of the fully connected layer and apply the ReLU activation function to calculate the mean of the latent space. μ and log variance , the latent variables are obtained by sampling , in, is random noise sampled from a standard normal distribution; S402: Obtain the denoised text vector based on the fully connected layer and activation function : ; in, represents the output of the third hidden layer; represents the weight matrix in the third hidden layer; represents the bias vector of the third hidden layer; Represents the weight matrix in the fourth hidden layer, specifically connecting the hidden layer and text vectors The weight of Represents the bias vector of the fourth hidden layer.

6. The method for evaluating a question-answer text according to claim 5, wherein: The S5 includes: converting the denoised text vector into a deep semantic representation through a bidirectional long short-term memory network, obtaining its context information for each vocabulary unit, forming a contextual representation, and obtaining a hidden state sequence , where A corresponds to the question text, Q corresponds to the answer text, and P corresponds to the matching text.

7. The method for evaluating a question-answer text according to claim 6, wherein: The S6 includes: S601, calculation sequence arrive ,sequence arrive ,sequence arrive and sequence arrive The dot product of the transpose of is used to obtain the attention weight matrix; using Normalize the attention weight matrix so that the sum of the weights of each row is 1; S602, the obtained normalized weight matrix and sequence Multiplying together, we get arrive Attention-weighted representation of ,get arrive Attention-weighted representation of , arrive Attention-weighted representation of , arrive Attention-weighted representation of , as follows: ; S603, from the sequence After subtracting the weighted , get the difference characteristics , as follows: 。 8. The method for evaluating a question-answer text according to claim 7, wherein: The S7 includes: will sequence Splicing to get text representation ; ； Representing text Input to a fully connected layer, After converting the output into probability, we get the corresponding similarity probability between texts and the probability of difference , as follows: ； 。 9. A question-answer text evaluation system, characterized in that: Applied to executing the evaluation method according to any one of claims 1 to 8, the system comprises: The data acquisition module is used to obtain the discussion and analysis texts of listed companies' management, as well as the question and answer texts in the question and answer scenarios set by listed companies, to form a question and answer dataset; A text matching module is used to perform similarity matching between the question text and the subject text based on a preset text matching strategy, and to generate a continuous sentence sequence with the highest similarity as the matching text; A preprocessing module, configured to preprocess the matching text and the question-answering dataset using a BERT model to generate a text vector; A variational autoencoder module is used to input the text vector into the variational autoencoder for decoding and encoding to obtain a denoised text vector; A bidirectional long short-term memory network module is used to input the denoised text vector into the bidirectional long short-term memory network for forward and backward processing to obtain a hidden state sequence; a bidirectional attention mechanism module, configured to calculate a bidirectional attention weighted representation and a difference feature of the hidden state sequence using a bidirectional attention mechanism; An evaluation module is used to concatenate the bidirectional attention weighted representation and the difference feature to obtain a context representation, and evaluate the semantic difference between the question and the answer text based on the context representation.

Images