An intelligent question answering method based on multi-view answer representation

Abstract

The application discloses an intelligent question and answer method based on multi-view answer representation, comprising the following steps: 1. constructing a question and answer data and preprocessing; 2. constructing a text coding module, obtaining a word vector matrix and a sentence vector of the text; 3. constructing a responder professional ability representation module, obtaining a representation of the professional ability of a responder to a question; 4. constructing an LSTM module based on an attention mechanism, obtaining information interaction of answers under the same question; 5. constructing an answer representation module based on a question description text sentence vector, weighting word vectors of an answer body by using vector representation of a question description text; 6. constructing an answer representation module based on a question and answer work graph, obtaining answer representation based on a concept graph; 7. constructing an answer representation fusion module, obtaining multi-view-based answer representation; and 8. constructing a prediction module, predicting a matching score of a question and an answer. The application can select the best answer for a question.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of community question and answer systems, and in particular to an intelligent question and answer method based on multi-view answer representation. BACKGROUND

[0002] With the rapid development of network technology, the way of information acquisition has become more diversified and convenient. As an important online knowledge sharing platform, community question and answer systems have gradually attracted widespread attention and use from users. How to select the best answer from numerous answers has become an important problem in the field of community question and answer.

[0003] Answer selection methods are mainly divided into feature engineering-based methods and deep learning-based methods. Among them, feature engineering-based methods rely on manually extracted features such as word frequency, TF-IDF, n-gram, etc., and then use traditional machine learning algorithms for classification or regression to select the best answer. However, this method is difficult to capture complex semantic information and has poor generalization ability. Deep learning-based methods can automatically learn potential features in text data through neural network models, especially through pre-trained models, which can handle complex semantic relationships in natural language.

[0004] Currently, few methods focus on the influence of entity words in the question and existing answers under the question on the quality of the answers published by the answerers when publishing answers. However, in practice, when browsing the question, the answerer will not only focus on the overall semantics of the question, but also focus on the key words in the question. The answerer will often state the answer around the question words of interest. When browsing the existing answers under the question, the answerer will understand and summarize the existing answers based on their professional ability for the question. The higher the professional ability of the answerer for the question, the higher the understanding degree of the background knowledge and terms in the high-quality answers in the existing answers. Finally, the answerer publishes the answer based on the understanding of the existing answers and their professional ability. Therefore, how to capture the fine-grained interaction between entity words in the question text and the answer text and the information interaction between answers under the same question is a difficulty in the answer selection task. SUMMARY

[0005] The present application overcomes the deficiencies in the prior art and provides an intelligent question and answer method based on multi-view answer representation to mine the interactive influence of entity words in the question text and entity words in the answer text and the influence of the first published answer under the question on the later published answer, thereby accurately recommending the answer with the highest matching degree to the question.

[0006] To achieve the above-mentioned application purposes, the present application adopts the following technical solutions:

[0007] The intelligent question-answering method based on multi-view answer representation of the present invention is characterized by the following steps:

[0008] Step 1: Construct and preprocess the question-and-answer data;

[0009] Collect question and answer data, including: questions and answers and their answerers. Questions include: question category, question topic, and question text. Answers include: answer text, posting time of each answer, and genuine evaluation tags.

[0010] Any question from the question-and-answer data After sorting all the corresponding answers by publication time, the question is obtained. The corresponding set of answer texts ;in, Indicate the problem The The sorted answer text Indicate the problem The total number of answers to the question The topic of the question is denoted as ; the problem Corresponding The set of respondents is denoted as ,in, express The respondent;

[0011] Step 2: Construct a text encoding module to encode the question topic. and the problem After concatenating the question text, we get the question. Description text Then, the BERT model was used to obtain the questions respectively. Description text word vector matrix Sentence vectors ,question The The following is the text of the arranged answers. word vector matrix Sentence vectors And the respondent Documents word vector matrix ;

[0012] Step 3: Construct a module representing the respondent's professional competence, including a TextCNN network and an attention mechanism module, and use it for... and Processing and obtaining responses On the issue The expression of professional competence ;

[0013] Step 4: Construct an LSTM module based on the attention mechanism and use it for... and Processing is performed to obtain the first... The answer text The representation of answer information interaction ;

[0014] Step 5: Construct an answer representation module based on the sentence vectors of the question description text, and use it for calculation. With the answer text No. Word vectors of each segment Match score reuse Function to answer text of The matching scores of each word segment are normalized to obtain... No. Word vectors of each segment weight value Thus, the answer text of The word vector matrix of each segment. Perform weighted aggregation to obtain Representation based on sentence vectors of the problem description text ;

[0015] Step 6: Construct an answer representation module based on a question-answering working graph, and use it for construction. Corresponding question-and-answer working subgraph Question and Answer Working Diagram Thus obtain Based on question-answering working graph The representation of ;

[0016] Step 7: Construct an answer representation fusion module to process the answer representation based on the question-answering working graph. The representation of Representation based on sentence vectors of problem description text Representation of interaction with answer information By splicing and merging, we obtain Multi-view representation ;

[0017] Step 8: Construct a prediction module and use it for multi-view-based representations. , the respondent On the issue The expression of professional competence and the problem Description text sentence vector Information fusion is used to predict problems. and the The answer text Matching score ;

[0018] Step 9: Train the intelligent question-answering network, which consists of a text encoding module, a respondent's expertise representation module, an attention-based LSTM module, an answer representation module based on question description text sentence vectors, an answer representation module based on question-answering working graphs, an answer representation fusion module, and a prediction module, using gradient descent. Calculate the cross-entropy loss value based on the real evaluation label and the predicted evaluation label to update the network parameters. Stop training when the cross-entropy loss value converges or the number of training iterations reaches a predetermined threshold, thereby obtaining the optimal intelligent question-answering model, which is then used to predict the best answer for each input question.

[0019] The intelligent question-answering method based on multi-view answer representation described in this invention is characterized in that step 2 is performed as follows:

[0020] Step 2.1: Use a word segmenter to... After word segmentation, it is converted into a question. The text sequence is input into the BERT model for processing, thereby obtaining the problem. Description text word vector matrix ,in, express After word segmentation, the first The vector corresponding to each word segmentation express Length;

[0021] Will The vector corresponding to the token [CLS] is used as the problem. Description text sentence vector ;

[0022] Step 2.2: Use a word segmenter to analyze the answer text. After word segmentation, it is converted into The text sequence is input into the BERT model for processing, thereby obtaining... word vector matrix ,in, express After word segmentation, the first The vector corresponding to each word segmentation express Length;

[0023] Will The vector corresponding to the token [CLS] is used as the answer text. sentence vector ;

[0024] Step 2.3, based on the question Question categories, obtaining respondents from question-and-answer data. Answers to other questions within the same question category are concatenated to form the [number]. One answerer Documents ,in, express The first in One word, express Length;

[0025] Using a word segmenter After word segmentation, it is converted into The text sequence is input into the BERT model for processing, resulting in... word vector matrix .

[0026] Furthermore, step 3 is performed as follows:

[0027] Step 3.1: Process using TextCNN network word vector matrix ,get vector representation ,Will and After splicing, we get the first... concatenated vectors Then, the input is fed into a fully connected layer for processing to obtain the respondent's answer. Intermediate representation of professional competence Where [:] represents the concatenation operation, and D represents the dimension of the column vector;

[0028] Step 3.2: Utilize the attention mechanism module to... Processing and obtaining responses On the issue professional competence ;

[0029] Step 3.2.1, will Copy to dimension D× Professional competency matrix Thus, by using equation (1), we can obtain and fusion representation :

[0030] (1)

[0031] In equation (1), express The weight matrix, express The weight matrix, express and Interaction weight matrix, This represents the element-wise product of two matrices. It is an activation function;

[0032] Step 3.2.2, using equation (2) to... After performing normalization, the respondent was obtained. Attention score vector :

[0033] (2)

[0034] In equation (2), express The parameter vector, Indicates transpose. It is an activation function;

[0035] Step 3.2.3, Utilize right After weighting, the respondents were obtained. On the issue professional competence .

[0036] Furthermore, step 4 is performed as follows:

[0037] Step 4.1: Define the total number of time steps as... , No. The input at the nth time step is the nth time step. The answer text sentence vector , the respondent On the issue professional competence , No. The output of the nth time step is the nth time step. The answer text Answer information interaction representation ;

[0038] Step 4.2, based on the previous The answer text ,in, Indicates the first Each answer text is used to calculate the respondent's score using formula (3). Acquisition and Problems The relevant information A degree :

[0039] (3)

[0040] In equation (3), It is the first The answer text The respondent On the issue A representation of professional competence;

[0041] Step 4.3, Calculate using equation (4) The fusion vector representing the interactive representation of answer information in the answer body. :

[0042] (4)

[0043] In equation (4), It is the first The answer text The representation of interactive answer information;

[0044] Step 4.4: Using equations (5)-(10) to obtain The representation of answer information interaction :

[0045] (5)

[0046] (6)

[0047] (7)

[0048] (8)

[0049] (9)

[0050] (10)

[0051] In equations (5)-(10), , , , , They represent the first The input gate, forget gate, output gate, candidate cell state, and cell state at each time step. Indicates the first cell state at a time step, denotes an element-wise multiplication, denotes a sigmoid activation function, denotes a tanh activation function, , , denote the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the input gate , denotes the bias vector of the input gate , , , denote the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the forget gate , denotes the bias vector of the forget gate , , , denote the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the output gate , denotes the bias vector of the output gate , , , denote the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the candidate cell state , denotes the bias vector of the candidate cell state .

[0052] Further, the step 6 is performed as follows:

[0053] Step 6.1, extracting the corresponding entity word set and and the common additional entity word set from the question and the answer text respectively by using the ConceptNet knowledge graph;

[0054] performing word segmentation, part-of-speech tagging and stemming operation on the description text of the question , and identifying the entity words matched with the ConceptNet knowledge graph from the operated description text to obtain the entity word set of the question ;

[0055] performing word segmentation, part-of-speech tagging and stemming operation on the description text The algorithm performs word segmentation, part-of-speech tagging, and stemming, and then identifies entity words matching the ConceptNet knowledge graph from the processed answer text. entity word set ;

[0056] Will The entity words in the ConceptNet knowledge graph are connected to adjacent entity word nodes, and The intersection of adjacent entity node connections between entity words in the ConceptNet knowledge graph is used as the basis for understanding the concept. The corresponding additional entity word set ;

[0057] Step 6.2: Obtain from ConceptNet knowledge graph , and The nodes corresponding to each entity word in the text are constructed. Corresponding question-and-answer working subgraph ;

[0058] Step 6.3, After linear transformation, the generation Low-dimensional vector representation And as a text node, it is added to the question-and-answer working subgraph. In, and will respectively with and Each entity word in the string is connected to obtain Corresponding Question and Answer Working Diagram ;

[0059] Step 6.4: Create a question-and-answer working subgraph. Any entity word n ​​in the string and After concatenation, we get the entity word n ​​and spliced ​​text ,Will Perform word segmentation and convert it into words. The text sequence is input into the BERT model for processing, resulting in... The vector representation of each word in the process is obtained by using a linear layer. The predicted score for each word segment;

[0060] by The integer identifiers of the text sequence are used as the actual labels, and the cross-entropy loss function is used to calculate... The difference between the predicted score and the actual label of each word segment is used to obtain the loss value of entity word n. The negative value of the loss value is taken as the score of entity word n. ;

[0061] Step 6.5, updating the embedding representation of all entity word nodes in the question-answer work graph using a graph neural network, and processing the embedding representation of all entity word nodes in the question-answer work graph using an attention mechanism to obtain an answer text based on the representation of the question-answer work graph . .

[0062] Further, the step 6.5 is performed as follows:

[0063] Step 6.5.1, for any entity word node n and its adjacent entity word node s on the question-answer work graph, respectively one-hot encoding the node type of the adjacent entity word node s, the node type of the entity word node n and the edge between the adjacent entity word node s and the entity word node n to obtain the one-hot encoding representation of the node type of the adjacent entity word node s , the one-hot encoding representation of the node type of the entity word node n , and the one-hot encoding representation of the edge between the entity word node n and the adjacent entity word node s , so as to obtain the relation type vector representation of the entity word node n and the adjacent entity word node s using formula (11) :

[0064] (11)

[0065] In formula (11), is a multi-layer perception for obtaining the relation type;

[0066] using the pre-trained vector to represent the entity word node n and the adjacent entity word node s on the question-answer work graph respectively to obtain the embedding representation of the adjacent entity word node s and the embedding representation of the entity word node n ;

[0067] linearly transforming the one-hot encoding representation of the node type of the adjacent entity word node s to obtain the vector representation of the node type of the adjacent entity word node s ;

[0068] linearly transforming the one-hot encoding representation of the node type of the entity word node n to obtain the vector representation of the node type of the entity word node n ;

[0069] Step 6.5.2, defining the current iteration number as , and initializing​​​​​​​ = 1, let denote the maximum number of iterations;

[0070] Step 6.5.3, obtain the message passed from the neighboring entity word node s to the entity word node n at the i-th iteration using formula (12):

[0071] (12)

[0072] In formula (12), denotes the embedding representation of the neighboring entity word node s at the i-th iteration, is the message linear layer; Step 6.5.4, input the score of the neighboring entity word node s and the score of the entity word node n into the multi-layer perceptron of the node score mapping respectively, and obtain the vector representation of and the vector representation of

[0073] respectively; Step 6.5.5, calculate the query vector of the neighboring entity word node s at the i-th iteration using formula (13): (13) In formula (13),

[0074] is the query vector linear layer; Step 6.5.6, calculate the key vector of the entity word node n at the i-th iteration using formula (14):

[0075] (14)

[0076] In formula (14), is the key vector linear layer;

[0077] Step 6.5.7, calculate the attention weight of using formula (15):

[0078] (15)

[0079] In formula (15), is the attention weight linear layer;

[0080] Step 6.5.8, obtain the score of the entity word node n at the i-th iteration using formula (16):

[0081] (16)

[0082] In formula (16), ​​​​​​​​​is the neighborhood of the neighboring entity word node s; denotes the transpose, is the key vector of the entity word node n at the th iteration;

[0083] Step 6.5.8, obtain the embedding representation of the entity word node n at the th iteration using formula (16) , so as to obtain the embedding representation of the entity word node n after the th iteration :

[0084] (16)

[0085] In formula (16), denotes the neighborhood of the entity word node n, is the message fusion multi-layer perception; denotes the attention weight of the message transmitted from the neighboring entity word node s' to the entity word node n at the th iteration;

[0086] Step 6.5.9, calculate the weight value of the entity word node n after the th iteration using formula (17) ;

[0087] (17)

[0088] Step 6.5.10, after calculating the weight value of each entity word node in the question and answer working subgraph , perform weighted summation on all entity word nodes in the question and answer working subgraph , to obtain the representation based on the question and answer working graph .

[0089] Further, the step 7 is performed as follows:

[0090] Step 7.1, obtain the final representation based on the question using formula (18) :

[0091] (18)

[0092] In formula (18), is the weight matrix of the concatenated vector and , is the weight matrix of the concatenated vector and ​​​​bias vector of the is an activation function;

[0093] Step 7.2, obtaining by using formula (19) based on multi-view representation :

[0094] (19)

[0095] In formula (19), is and the concatenated vector weight matrix of the is bias vector of the

[0096] Further, the step 8 is performed as follows:

[0097] Step 8.1, obtaining by using formula (20) based on multi-view representation and the professional ability of the answerer to the question fusion representation of the :

[0098] (20)

[0099] In formula (20), is and the concatenated vector weight matrix of the is bias vector of the

[0100] Step 8.2, obtaining by using formula (21) matching score of the question , so as to obtain the predicted evaluation label of the :

[0101] (21)

[0102] In formula (21), is and the concatenated vector weight matrix of the is bias vector of the is and parameter vector after information fusion through the full connection layer​ is and The bias term after information fusion through the full connection layer, and the sigmoid is an activation function.

[0103] The electronic device comprises a memory and a processor, and the memory is used for storing a program supporting the processor to execute the intelligent question and answer method, and the processor is configured to execute the program stored in the memory.

[0104] The computer readable storage medium stores a computer program, and when the computer program is run by a processor, the steps of the intelligent question and answer method are executed.

[0105] Compared with the existing technology, the present application has the following advantages:

[0106] 1. The present application splices the question topic and the question text as the description text of the question, so that the semantic information of the question can be obtained more completely;

[0107] 2. The present application fuses the answer information published by the answerer under the current question and the answer information published under other questions of the same question category, better describes the breadth of the answerer's understanding of the current question category, and then uses the attention mechanism to obtain the representation of the answerer's professional ability for the current question, which reduces the noise compared with modeling the professional ability of the answerer by using all the answers published by the answerer, and can dynamically adjust the professional ability representation of the answerer for different questions under the same question category;

[0108] 3. Before publishing the answer, the answerer will be influenced by the existing answers under the question, the present application adds an attention mechanism to the existing LSTM model, based on the professional ability of the answerer for the current question, uses the attention mechanism to obtain the comprehensive influence of all the previously published answers under the question on the later answers, thereby enhancing the interpretability of the model;

[0109] 4. The present application introduces a knowledge graph, constructs a question and answer work graph combining a pre-trained language model and a knowledge graph, uses a graph neural network to capture the influence of the question entity word and the overall semantic of the question on the answer entity word, and has stronger interpretability and generalization ability compared with the traditional method using the attention mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0110] Figure 1 is the overall model diagram of the present application;

[0111] Figure 2 is the single cell state structure diagram of the LSTM of the present application with an attention mechanism;

[0112] Figure 3 This is the question-and-answer working diagram of the present invention. Detailed Implementation

[0113] In this embodiment, an intelligent question-answering method based on multi-view answer representation utilizes a graph neural network to capture fine-grained interactions between question and answer texts based on entity words. It also leverages the overall semantics of the question description text to obtain the semantic representation of the answer. In modeling the information interaction between answers to the same question, it comprehensively considers the influence of the overall question semantics on the answer and the interaction between entity words and question text information in the question and answer texts to obtain rich semantic information about the answer. Furthermore, it employs an attention-based LSTM module to assess the influence of earlier published answers on later published answers based on the respondent's expertise in the question, thereby selecting the optimal answer for the question. Specifically, as... Figure 1 As shown, the method is performed according to the following steps:

[0114] Step 1: Construct and preprocess the question-and-answer data;

[0115] Collect question and answer data, including: questions and answers and their answerers. Questions include: question category, question topic, and question text. Answers include: answer text, posting time of each answer, and genuine evaluation tags.

[0116] Any question from the question-and-answer data After sorting all the corresponding answers by publication time, the question is obtained. The corresponding set of answer texts ;in, Indicate the problem The The sorted answer text Indicate the problem The total number of answers to the question The topic of the question is denoted as ; the problem Corresponding The set of respondents is denoted as ,in, express The respondent;

[0117] In practice, the collected community Q&A data is stored in XML files, with each record being a question-answer pair, collectively forming a question-answer pair set. Each question records the question ID, question category, question topic, and question body; each answer records the answer body, posting time, genuine evaluation tags, and the answerer ID; finally, the answers to each question are sorted chronologically.

[0118] Step 2, the BERT model can effectively capture the dependency between distant words in the input text, and a text encoding module is constructed to splice the question theme and the question text to obtain the description text of the question ; After that, the BERT model is used to obtain the word vector matrix and the sentence vector of the description text of the question , the word vector matrix and the sentence vector of the rearranged answer text of the question , and the word vector matrix of the document of the answerer ; Step 2.1, after tokenizing the question using a tokenizer, the text sequence of the question is converted and input into the BERT model for processing, so as to obtain the word vector matrix of the description text of the question , wherein represents the vector corresponding to the i-th tokenized word, and represents the length of the question ;

[0119] The vector corresponding to the token [CLS] aggregates the information of the entire input sequence, and in this embodiment, the vector corresponding to the token [CLS] in the question is taken as the sentence vector of the description text of the question . Step 2.2, after tokenizing the answer text using a tokenizer, the text sequence of the question is converted and input into the BERT model for processing, so as to obtain the word vector matrix of the question , wherein represents the vector corresponding to the i-th tokenized word, and represents the length of the question ;

[0120] The vector corresponding to the token [CLS] aggregates the information of the entire input sequence, and in this embodiment, the vector corresponding to the token [CLS] in the question is taken as the sentence vector of the description text of the question . Step 2.2, after tokenizing the answer text using a tokenizer, the text sequence of the question

[0121] is converted and input into the BERT model for processing, so as to obtain the word vector matrix of the question , wherein represents the vector corresponding to the i-th tokenized word, and represents the length of the question ; The vector corresponding to the token [CLS] aggregates the information of the entire input sequence, and in this embodiment, the vector corresponding to the token [CLS] in the question is taken as the sentence vector of the description text of the question . Step 3, the question

[0122] is input into the BERT model, and the question ​​The vector corresponding to the token [CLS] is used as the answer text. sentence vector .

[0123] Step 2.3: The answers posted by the respondent to other questions within the same question category demonstrate the breadth of the respondent's knowledge related to the question category, based on the question. Question categories, obtaining respondents from question-and-answer data. Answers to other questions within the same question category are concatenated to form the [number]. One answerer Documents ,in, express The first in One word, express Length;

[0124] Using a word segmenter After word segmentation, it is converted into The text sequence is input into the BERT model for processing, resulting in... word vector matrix .

[0125] Step 3: Construct a module representing the respondent's professional competence, including a TextCNN network and an attention mechanism module, and use it for... and Processing and obtaining responses On the issue The expression of professional competence ;

[0126] Step 3.1: When processing text tasks, TextCNN can efficiently extract local features, avoid text sequence dependencies, and perform multi-scale feature representation. In this embodiment, the TextCNN network is used for processing. word vector matrix ,get vector representation For the respondent In the question The answer text is below. and the respondent Corresponding document By processing and then merging the responses separately, the respondents can be obtained dynamically. An intermediate representation of the level of expertise for different questions within the same question category; in specific implementation, and After splicing, we get the first... concatenated vectors After, input to the full connection layer for processing, get the professional ability of the answerer of the question , where [ : ] represents the splicing operation, and D represents the dimension of the column vector.

[0127] Step 3.2, the question category involves a wide range of content, and the question description text often only reflects one aspect or a few aspects of the question category. The attention mechanism can obtain the professional ability of the answerer focus on the question . In this embodiment, the attention mechanism module is used to process , so as to obtain the professional ability representation of the answerer to the question ; ;

[0128] Step 3.2.1, copy to the professional ability matrix with dimension D x , so as to obtain the fusion representation of and using formula (1):

[0129] (1)

[0130] In formula (1), represents the weight matrix of , represents the weight matrix of , represents the weight matrix of the interaction between and , represents the element-wise product of two matrices, is an activation function.

[0131] Step 3.2.2, after normalizing using formula (2), the attention score vector of the answerer is obtained:

[0132] (2)

[0133] In formula (2), represents the parameter vector of , represents the transpose, is an activation function;

[0134] Step 3.2.3, after weighting on , the professional ability of the answerer​​​ On the issue professional competence .

[0135] Step 4, as follows Figure 2 As shown, in this embodiment, an LSTM module based on an attention mechanism is constructed. An attention mechanism is added to the basic LSTM cell unit, and the input is no longer the first... The output of the LSTM cell unit at time 1 The answer text The representation of answer information interaction Instead, the front The aggregated information, representing the answer information output by the LSTM cell at each time step, is used to analyze the results. and Processing is performed to obtain the first... The answer text The representation of answer information interaction ;

[0136] Step 4.1: Define the total number of time steps as... , No. The input at the nth time step is the nth time step. The answer text sentence vector , the respondent On the issue professional competence , No. The output of the nth time step is the nth time step. The answer text Answer information interaction representation ;

[0137] Step 4.2, based on the previous The answer text ,in, Indicates the first Each answer text is used to calculate the respondent's score using formula (3). Acquisition and Problems The relevant information A degree , The higher the number, the better. The answer text is for the respondent. The answer provided The greater the impact on quality:

[0138] (3)

[0139] In equation (3), It is the first The answer text The respondent On the issue This is an indication of professional competence.

[0140] Step 4.3, Calculate using equation (4) The fusion vector representing the interactive representation of answer information in the answer body. :

[0141] (4)

[0142] In equation (4), It is the first The answer text The representation of interactive answer information;

[0143] Step 4.4: Using equations (5)-(10) to obtain The representation of answer information interaction :

[0144] (5)

[0145] (6)

[0146] (7)

[0147] (8)

[0148] (9)

[0149] (10)

[0150] In equations (5)-(10), , , , , They represent the first The input gate, forget gate, output gate, candidate cell state, and cell state at each time step. Indicates the first Cell state at each time step Represents element-wise product. This represents the sigmoid activation function. This represents the tanh activation function. , , They represent the input gates respectively. The answer text weight matrix, the answer information fusion weight matrix, and the respondent's professional ability weight matrix. Indicates input gate The bias vector, , , These represent the gates of forgetting. The answer text weight matrix, the answer information fusion weight matrix, and the respondent's professional ability weight matrix. Represents the Gate of Oblivion The bias vector, , , These represent the output gates. The answer text weight matrix, the answer information fusion weight matrix, and the respondent's professional ability weight matrix. Indicates output gate The bias vector, , , These represent the candidate cell states. The answer text weight matrix, the answer information fusion weight matrix, and the respondent's professional ability weight matrix. Indicates the state of candidate cells The bias vector.

[0151] Step 5, Question Description text The overall semantics of the answer text The representation is very important. In this embodiment, an answer representation module based on the sentence vectors of the question description text is constructed and used to calculate... With the answer text No. Word vectors of each segment Match score In practice, The calculation method is as follows transpose vector and The dot product, then using Function to answer text of The matching scores of each word segment are normalized to obtain... No. Word vectors of each segment weight value Thus, the answer text of The word vector matrix of each segment. Perform weighted aggregation to obtain Representation based on sentence vectors of the problem description text This allows for more accurate capture of the answer text. The semantics of.

[0152] Step 6: Utilizing knowledge graphs can establish closer semantic connections between entity words in questions and entity words in answers, enhancing the model's reasoning ability and interpretability. In this embodiment, an answer representation module based on a question-answering working graph is constructed and used to construct... Corresponding question-and-answer working subgraph Question and Answer Working Diagram Thus obtain Based on question-answering working graph The representation of ;

[0153] Step 6.1: Utilize ConceptNet knowledge graph from and Extract the corresponding entity word sets respectively and and its public additional entity word set ;

[0154] The problem Description text The algorithm performs word segmentation, part-of-speech tagging, and stemming, and then identifies entity words matching the ConceptNet knowledge graph from the processed description text. entity word set ;

[0155] Will The algorithm performs word segmentation, part-of-speech tagging, and stemming, and then identifies entity words matching the ConceptNet knowledge graph from the processed answer text. entity word set ;

[0156] Will The entity words in the ConceptNet knowledge graph are connected to adjacent entity word nodes, and The intersection of adjacent entity node connections between entity words in the ConceptNet knowledge graph is used as the basis for understanding the concept. The corresponding additional entity word set .

[0157] Step 6.2: Obtain from ConceptNet knowledge graph , and The nodes corresponding to each entity word in the text are constructed. Corresponding question-and-answer working subgraph In this embodiment, and The same entity words are assigned to question node type and answer node type respectively, thus serving as Different entity nodes in the knowledge graph;

[0158] Step 6.3, retrieving the question and answer work subgraph from the knowledge graph Some entity nodes irrelevant to the question are included to bridge the question Description text The difference between the overall semantic information of the description text and the information of the entity nodes in the knowledge graph is that, in this embodiment, the After linear transformation, a low-dimensional vector representation of is generated and added to the question and answer work subgraph as a text node, and is connected to each entity word in and respectively, thereby obtaining the corresponding question and answer work graph , as shown in Figure 3 , which contains all the entity nodes in .

[0159] Step 6.4, question and answer work subgraph The entity words in the question and answer work subgraph that are more closely related to the description text better reflect the important information of the description text In this embodiment, the concatenation text of any entity word n in the question and answer work subgraph and is obtained by concatenating the entity word n and , the is segmented into a text sequence , and input into the BERT model for processing to obtain the vector representation of each segmented word in , thereby obtaining the predicted score of each segmented word in using a linear layer;

[0160] The integer identifier of the text sequence is used as the actual label, and the cross-entropy loss function is used to calculate the difference between the predicted score of each segmented word in and the actual label, thereby obtaining the loss value of the entity word n, and taking the negative value of the loss value as the score of the entity word n , The higher the score of the entity word n, the higher the degree of association between the entity word n and the description text .

[0161] Step 6.5, using a graph neural network to iteratively update the embedding representation of all entity nodes in , and using an attention mechanism to process and​ all entity word nodes in the answer text based on the question-answer work graph representation ;

[0162] Step 6.5.1, the graph neural network can utilize the question-answer work graph all nodes, thereby effectively capturing the information interaction between the answer text and the question description text . For any entity word node n and its adjacent entity word node s on the question-answer work graph , the node type of the adjacent entity word node s, the node type of the entity word node n and the edge between the adjacent entity word node s and the entity word node n are one-hot encoded respectively, to obtain the one-hot encoded representation of the node type of the adjacent entity word node s , the one-hot encoded representation of the node type of the entity word node n , and the one-hot encoded representation of the edge between the entity word node n and the adjacent entity word node s , thereby obtaining the relation type vector representation of the entity word node n and the adjacent entity word node s using formula (11) :

[0163] (11)

[0164] In formula (11), , , , represents the number of node types in the question-answer work graph , , , represents the number of types of edges in the question-answer work graph , is a multilayer perceptron to obtain the relation type.

[0165] The entity word node n and the adjacent entity word node s on the question-answer work graph are represented by pre-trained vectors respectively, to obtain the embedding representation of the adjacent entity word node s and the embedding representation of the entity word node n ;

[0166] The one-hot encoded representation of the node type of the adjacent entity word node s is linearly transformed to obtain the vector representation of the node type of the adjacent entity word node s ;

[0167] The one-hot encoded representation of the node type of the entity word node n A linear transformation is performed to obtain a vector representation of the node type of the entity word node n ;

[0168] Step 6.5.2, define the current iteration number as , and initialize = 1, let denote the maximum number of iterations;

[0169] Step 6.5.3, obtain the message passed from the adjacent entity word node s to the entity word node n in the th iteration by using formula (12) :

[0170] (12)

[0171] In formula (12), denotes the embedding representation of the adjacent entity word node s in the th iteration, , is a message linear layer.

[0172] Step 6.5.4, in this embodiment, in order to maintain the consistency of the data format, the score of the adjacent entity word node s and the score of the entity word node n are respectively input into the multilayer perceptron of the node score mapping for processing, and the vector representation of and the vector representation of are obtained accordingly;

[0173] Step 6.5.5, calculate the query vector of the adjacent entity word node s in the th iteration by using formula (13) , wherein

[0174] (13)

[0175] In formula (13), is a query vector linear layer.

[0176] Step 6.5.6, calculate the key vector of the entity word node n in the th iteration by using formula (14) , wherein

[0177] (14)

[0178] In formula (14), It is a linear layer of key vectors.

[0179] Step 6.5.7: Calculate using equation (15) Attention weights :

[0180] (15)

[0181] In equation (15), It is the neighborhood of the adjacent entity word node s; Indicates transpose. It is the first The key vector of entity word node n′ in the next iteration;

[0182] Step 6.5.8: Use equation (16) to obtain the first... Embedding representation of entity word node n in the next iteration Thus, the first Embedding representation of entity word node n after the next iteration :

[0183] (16)

[0184] In equation (16), Represents the neighborhood of entity word node n. It is a multilayer perceptron for message fusion; the first In the next iteration, the neighborhood of entity word node n The messages passed from different entity word nodes to entity word node n have different levels of importance. Indicates the first In the next iteration, the message passed from the adjacent entity node s' to the entity node n Attention weights The larger the value, For entity word node n, the first Each iteration is more important.

[0185] Step 6.5.9: Calculate the first step using equation (17). The weight value of entity node n after the next iteration ;

[0186] (17)

[0187] Step 6.5.10: Calculate the question-and-answer working subgraph After assigning weight values ​​to each entity node, the question-answering working subgraph is processed. We perform a weighted summation on all entity word nodes to obtain... Based on question-answering working graph The representation of .

[0188] Step 7: By combining different representations of the answer text, richer semantic information of the answer text can be obtained. In this embodiment, an answer representation fusion module is constructed to process the answer representation based on the question-answering working graph. The representation of Representation based on sentence vectors of problem description text Representation of interaction with answer information By splicing and merging, we obtain Multi-view representation ;

[0189] Step 7.1: Obtain using equation (18) Based on the question The final expression :

[0190] (18)

[0191] In equation (18), yes and concatenated vector The weight matrix, yes The bias vector, It is an activation function.

[0192] Step 7.2, using equation (19) to obtain Multi-view representation :

[0193] (19)

[0194] In equation (19), yes and concatenated vector The weight matrix, yes The bias vector.

[0195] Step 8: In this embodiment, the answer selection task is treated as a binary classification problem. A prediction module is constructed and used to process the multi-view-based representation. , the respondent On the issue The expression of professional competence and the problem Description text sentence vector Information fusion is used to predict problems. and the the matching score of the answer text , , between 0 and 1, indicating the first answer text probability of being a high-quality answer;

[0196] Step 8.1, obtaining a fusion representation based on the multi-view representation and the representation of the professional ability of the answerer to the question : :

[0197] (20)

[0198] In formula (20), is the weight matrix of the concatenated vector and , is the bias vector of .

[0199] Step 8.2, obtaining the matching score of the answer text to the question , so as to obtain the predicted evaluation label of :

[0200] (21)

[0201] In formula (21), is the weight matrix of the concatenated vector and , is the bias vector of , is the parameter vector of and after information fusion through the fully connected layer, is the bias term of and after information fusion through the fully connected layer, and sigmoid is the activation function.

[0202] ​​​Step 9, the intelligent question answering network composed of the text encoding module, the answerer professional ability representation module, the attention mechanism based LSTM module, the answer representation module based on the question description text sentence vector, the answer representation module based on the question answering work graph, the answer representation fusion module and the prediction module is trained by using the gradient descent method, and the cross entropy loss value based on the true evaluation label and the predicted evaluation label is calculated to update the network parameters, and the training is stopped when the cross entropy loss value converges or the training times reach a predetermined threshold, so that the optimal intelligent question answering model is obtained, and is used for predicting the best answer for each input question.

[0203] In this embodiment, an electronic device includes a memory for storing a program supporting the processor to execute the above method, and a processor configured to execute the program stored in the memory.

[0204] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program performs the steps of the above method when executed by a processor.

Claims

1. An intelligent question answering method based on multi-view answer representation, characterized in that, is performed as follows: Step 1, constructing question and answer data and preprocessing; Collecting question and answer data, including: questions and answers and their answerers, wherein the questions include: question category, question topic, question text, the answers include: answer text, the publishing time of each answer and the true evaluation label; Any question from the question-and-answer data After sorting all the corresponding answers by publication time, the question is obtained. The corresponding set of answer texts ;in, Indicate the problem The The sorted answer text, Indicate the problem The total number of answers to the question The topic of the question is denoted as ; the problem Corresponding The set of respondents is denoted as ,in, express The respondent; Step 2, construct a text encoding module for splicing the question theme and the question After the question body is spliced, the description text of the question is obtained After that, the BERT model is used to obtain the word vector matrix and the sentence vector of the description text of the question and the word vector matrix and the sentence vector of the first arranged answer body of the question and the word vector matrix of the document of the answerer ;​​​​​​​ Step 3, constructing an answerer professional ability representation module, comprising: a TextCNN network and an attention mechanism module, and being used for processing and to obtain an answerer representation of professional ability for a question ; ; Step 4, construct an attention mechanism based LSTM module and use it to process and to derive the representation of the answer information interaction of the first answer text . ; Step 5, construct the answer representation module based on the question description text sentence vector, and used to calculate The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word The matching score of the word vector of the first segmented word Step 6, build answer representation module based on Q&A work graph and used to build Corresponding Q&A work sub-graph And Q&A work graph , so as to obtain Representation based on Q&A work graph ;​ Step 7, construct answer representation fusion module, used for splicing fusion of representations based on question-answer work graph representation representation based on question description text sentence vector and answer information interaction get multi-view based representation ; Step 8: Construct a prediction module and use it for multi-view-based representations. , the respondent On the issue The expression of professional competence and the problem Description text sentence vector Information fusion is used to predict problems. and the The answer text Matching score ; Step 9, training the intelligent question and answer network composed of the text encoding module, the answerer professional ability representation module, the attention mechanism based LSTM module, the question description text sentence vector based answer representation module, the question and answer work graph based answer representation module, the answer representation fusion module and the prediction module by using the gradient descent method, and calculating the cross entropy loss value based on the true evaluation label and the predicted evaluation label to update the network parameters, stopping training when the cross entropy loss value converges or the number of training reaches a predetermined threshold, thereby obtaining an optimal intelligent question and answer model, and using it to predict the best answer for each input question. 2.The intelligent question answering method based on multi-view answer representation according to claim 1, characterized in that, Step 2 is performed as follows: Step 2.1: Use a word segmenter to... After word segmentation, it is converted into a question. The text sequence is input into the BERT model for processing, thereby obtaining the problem. Description text word vector matrix ,in, express After word segmentation, the first The vector corresponding to each word segmentation express Length; The token [CLS] is used to mark the beginning of a new example. The vector corresponding to the token [CLS] is used as the question description text sentence vector ; Step 2.2: Use a word segmenter to analyze the answer text. After word segmentation, it is converted into The text sequence is input into the BERT model for processing, thereby obtaining... word vector matrix ,in, express After word segmentation, the first The vector corresponding to each word segmentation express Length; The method comprises the following steps: The vector corresponding to the token [CLS] in the answer text is taken as the answer text The sentence vector of the answer text ; Step 2.3, based on the question Question categories, obtaining respondents from question-and-answer data. Answers to other questions within the same question category are concatenated to form the [number]. One answerer Documents ,in, express The first in One word, express Length; Using a word segmenter After word segmentation, it is converted into The text sequence is input into the BERT model for processing, resulting in... word vector matrix . 3.The intelligent question answering method based on multi-view answer representation according to claim 2, characterized in that, Step 3 is performed as follows: Step 3.1: Process using TextCNN network word vector matrix ,get vector representation ,Will and After splicing, we get the first... concatenated vectors Then, the input is fed into a fully connected layer for processing to obtain the respondent's answer. Intermediate representation of professional competence Where [:] represents the concatenation operation, and D represents the dimension of the column vector; Step 3.2: Utilize the attention mechanism module to... Processing and obtaining responses On the issue professional competence ; Step 3.2.1, Copy to dimension D× Professional competency matrix Thus, by using equation (1), we can obtain and fusion representation : (1) In formula (1), denotes a weight matrix of denotes a weight matrix of denotes and a weight matrix of the interaction between denotes an element-wise product of two matrices, is an activation function; Step 3.2.2, using equation (2) to... After performing normalization, the respondent was obtained. Attention score vector : (2) In formula (2), denotes a parameter vector of denotes the transpose, is an activation function; Step 3.2.3, utilizing right After weighting, the respondents were obtained. On the issue professional competence . 4.The intelligent question answering method based on multi-view answer representation according to claim 3, characterized in that, Step 4 is performed as follows: Step 4.1: Define the total number of time steps as... , No. The input at the nth time step is the nth time step. The answer text sentence vector , the respondent On the issue professional competence , No. The output of the nth time step is the nth time step. The answer text Answer information interaction representation ; Step 4.2, according to the previous answer text wherein, denotes the answer text, the answerer is asked to obtain information related to the question with a degree of : (3) In formula (3), is the answer text of the answerer to the question Step 4.3, calculating the answer information interaction representation of the answer text using formula (4) fusion vector of answer information interaction representation of the answer text : (4) In formula (4), is the first answer text of the answer information interaction; Step 4.

4. Obtaining the answer information interaction representation of the answer information interaction using the formula (5) - formula (10) :​ (5) (6) (7) (8) (9) (10) (5) - (10) in the formula, , , , , respectively represent the input gate, the forget gate, the output gate, the candidate cell state and the cell state of the first time step, represent the cell state of the first time step, represent the element-wise product, represent the sigmoid activation function, represent the tanh activation function, , , respectively represent the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the input gate , represent the bias vector of the input gate , , , respectively represent the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the forget gate , represent the bias vector of the forget gate , , , respectively represent the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the output gate , represent the bias vector of the output gate , , , respectively represent the answer text weight matrix, the answer information fusion weight matrix and the answerer professional ability weight matrix of the candidate cell state , represent the bias vector of the candidate cell state . 5.The intelligent question answering method based on multi-view answer representation according to claim 4, characterized in that, The step 6 is performed as follows: Step 6.1, extract the corresponding entity word sets from and respectively and and their common extra entity word set ; The problem Description text The algorithm performs word segmentation, part-of-speech tagging, and stemming, and then identifies entity words matching the ConceptNet knowledge graph from the processed description text. entity word set ; performing word segmentation, part-of-speech tagging, and stemming operations, and identifying entity words matching the ConceptNet knowledge graph from the answer text after the operations to obtain an entity word set ;​​ Will The entity words in the ConceptNet knowledge graph are connected to adjacent entity word nodes, and The intersection of adjacent entity node connections between entity words in the ConceptNet knowledge graph is used as... The corresponding additional entity word set ; Step 6.

2. Obtain the nodes corresponding to each entity word in , and from the ConceptNet knowledge graph, and construct the corresponding question and answer work subgraph . ; Step 6.3, add to the low-dimensional vector representation of after linear transformation and as a text node, add to the question and answer work subgraph , and connect to each entity word in and respectively, so as to obtain the corresponding question and answer work graph ; Step 6.4, concatenate any entity word n in the Q&A work subgraph with the answer to the question, to obtain the concatenated text of the entity word n and the answer to the question , and perform word segmentation on , to obtain the text sequence of , and input it into the BERT model for processing, to obtain the vector representation of each word segmentation in , so as to obtain the predicted score of each word segmentation in by using a linear layer. With the integer identifier of the text sequence of as the actual label, the difference between the predicted score of each segmented word in and the actual label is calculated using the cross-entropy loss function to obtain the loss value of the entity word n, and the negative value of the loss value is taken as the score of the entity word n ; Step 6.

5. Iteratively update the embedding representations of all entity word nodes with a graph neural network and process the embedding representations of all entity word nodes with an attention mechanism and obtain an answer text and based on the representations of the question-answer work graph . . . 6.The intelligent question answering method based on multi-view answer representation according to claim 5, characterized in that, The step 6.5 is performed as follows: Step 6.5.1, for a Q&A work graph any entity word node n on the Q&A work graph and its adjacent entity word nodes s, respectively one-hot encoding the node type of the adjacent entity word node s, the node type of the entity word node n, and the edge between the adjacent entity word node s and the entity word node n, to obtain a one-hot encoding representation of the node type of the adjacent entity word node s a one-hot encoding representation of the node type of the entity word node n a one-hot encoding representation of the edge between the entity word node n and the adjacent entity word node s so as to obtain a relation type vector representation of the entity word node n and the adjacent entity word node s by using formula (11) : (11) In formula (11), is a multi-layer perceptron that obtains the relationship type; representing entity word nodes n and adjacent entity word nodes s on a question and answer work graph with pre-trained vectors, respectively, to obtain an embedding representation of the adjacent entity word nodes s and an embedding representation of the entity word nodes n ;​ one-hot encoding representation of the node type of the adjacent entity word node s performing a linear transformation to obtain a vector representation of the node type of the adjacent entity word node s ; one-hot encoding representation of the node type of the entity word node n performing a linear transformation to obtain a vector representation of the node type of the entity word node n ; Step 6.5.2, define the current iteration number as and initialize = 1, let denote the maximum number of iterations; Step 6.5.3, obtaining the first iteration of the message passed from entity word node s to entity word node n using formula (12) message passed from entity word node s to entity word node n in the next iteration : (12) In formula (12), denotes the embedding representation of the neighboring entity node s at the , message linear layer; Step 6.5.4, score of the entity word node s adjacent to the entity word node n and the score of the entity word node n are input into a multi-layer perceptron mapping of node scores, respectively, to obtain vector representations and , respectively;​​ Step 6.5.

5. Calculating the query vector for the neighboring entity node s in the kth iteration using formula (13) query vector for the neighboring entity node s in the kth iteration : (13) In formula (13), is a query vector linear layer; Step 6.5.

6. Calculating the key vector for entity node n in the next iteration using formula (14) Key vector for entity node n in the next iteration : (14) In formula (14), is a key vector linear layer; Step 6.5.7, calculating with formula (15) attention weights of : (15) In formula (15), is the neighborhood of the adjacent entity word node s; denotes the transpose, is the key vector of the entity word node n' at the iteration. Step 6.5.8, obtaining the embedding representation of the entity word node n at the i-th iteration using formula (16) the embedding representation of the entity word node n at the i-th iteration , thereby obtaining the embedding representation of the entity word node n at the i-th iteration the embedding representation of the entity word node n at the i-th iteration : (16) In formula (16), denotes the neighborhood of the entity word node n, is a message-passing multi-layer perceptron; denotes the th iteration of the attention weight of the message passed from the neighboring entity word node s' to the entity word node n. ​ Step 6.5.

9. Calculating the weight value of the entity word node n after the second iteration using formula (17) weight value of the entity word node n after the first iteration ; (17) Step 6.5.10, computing the Q&A work subgraph After weighting each entity word node by the weight value, the Q&A work subgraph is computed by weighted summing all entity word nodes in the Q&A work subgraph to obtain a representation based on the Q&A work graph .

7. The intelligent question answering method based on multi-view answer representation according to claim 6, characterized in that, The step 7 is performed as follows: Step 7.1, using formula (18) to obtain based on the question final representation : (18) In formula (18), is and the concatenated vector is a weight matrix for is is a bias vector for is an activation function; Step 7.2, obtaining of formula (19) from Multi-view based representation : (19) In formula (19), is and the concatenated vector is a weight matrix for is is a bias vector for 8.The intelligent question answering method based on multi-view answer representation according to claim 7, characterized in that, The step 8 is performed as follows: Step 8.1, obtaining with formula (20) Multi-view based representation And the answerer To the question Representation of the professional competence Fusion representation : (20) In formula (20), is and the concatenated vector is the weight matrix of is is the bias vector of Step 8.

2. Obtain the prediction rating label of the problem using the formula (21) as follows: ​​​​ (21) In formula (21), is and the concatenated vector is a weight matrix of is is a bias vector of is and is a parameter vector after information fusion through the fully connected layer, is and is a bias term after information fusion through the fully connected layer, and sigmoid is an activation function.

9. An electronic device comprising a memory and a processor, characterized in that The memory is used to store a program supporting the processor to execute the intelligent question and answer method of any one of claims 1-8, and the processor is configured to execute the program stored in the memory.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is run by the processor to execute the steps of the intelligent question and answer method of any one of claims 1-8.

Images