CN112685543B - A method and device for answering questions based on text - Google Patents

Abstract

The embodiment of the invention provides a method and a device for answering questions based on texts, wherein the method comprises the following steps: inputting the question semantic information and the text semantic information into a preset question answer model, and determining a first parameter corresponding to each word in the text semantic information and a second parameter corresponding to each word in the question semantic information according to the question semantic information, the text semantic information and a classifier in the preset question answer model; determining context characterization parameters of the perceivable problem according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset problem answer model; and outputting a start index and a stop index of the answer questions in the text semantic information according to the question semantic information and the context characterization parameters. The apparatus performs the above method. The method and the device provided by the embodiment of the invention can accurately output the starting index and the ending index of the answer questions in the text semantic information.

Description

A method and device for answering questions based on text

Technical field

The present invention relates to the field of artificial intelligence technology, and in particular, to a method and device for answering questions based on text.

Background technique

With the development of artificial intelligence technology, people need to let computers replace people in part of their activities. For example, let computers read text like humans and answer questions based on their understanding of the text. This kind of reading comprehension is like asking a computer to do reading comprehension questions similar to college entrance examination English. Currently, neural network-based methods have become a mainstream trend because they can capture the semantic and grammatical relationships between questions and texts.

The existing technology uses LSTM to answer questions based on text. Specifically, it can include: using the Seq2Seq model solution, using the LSTM encoder-decoder to learn sequence-to-sequence model reading comprehension, including two LSTMs: one encoder and one decoder. The encoder takes a sequence as input and processes one symbol at each time point. Its purpose is to convert the symbol sequence into a fixed-size feature vector that only encodes important information in the sequence while losing unnecessary information. . Reading comprehension can be achieved through the Seq2Seq model solution. The input is question (question semantic information) and documents (text semantic information). The model scans and outputs the starting index and ending index in the documents. However, the above method cannot solve this problem well. question, the accuracy of answering the question is very low.

Contents of the invention

In order to solve the problems existing in the existing technology, embodiments of the present invention provide a method and device for answering questions based on text.

Embodiments of the present invention provide a method for answering questions based on text, including:

Input the question semantic information and the text semantic information into the preset question answering model, and determine the first words corresponding to each word in the text semantic information based on the question semantic information, the text semantic information and the classifier in the preset question answering model. A parameter, and a second parameter respectively corresponding to each word in the question semantic information; wherein the first parameter includes semantic information in the question semantic information, respectively related to each word in the text semantic information, the said third parameter The second parameter contains the semantic information in the text semantic information and related to each word in the question semantic information;

Determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model;

According to the question semantic information and the context characterization parameter, a start index and an end index of answering the question are output in the text semantic information.

Wherein, based on the question semantic information, the text semantic information and the classifier in the preset question answering model, the first parameters respectively corresponding to each word in the text semantic information and the first parameters corresponding to each word in the question semantic information are determined. The corresponding second parameters include:

According to the question semantic information and the text semantic information, a similarity matrix is determined; the rows of the similarity matrix are the similarities between all words in the question semantic information corresponding to each word in the text semantic information, and the similarity is The columns of the matrix are the similarities between all words in the text semantic information corresponding to each word in the question semantic information;

The first parameter is determined based on the similarity matrix, the question semantic information and the classifier, and the second parameter is determined based on the similarity matrix, the text semantic information and the classifier .

Wherein, determining the first parameter according to the similarity matrix, the question semantic information and the classifier includes: determining the first parameter according to the following formula:

U'＝Σ _j (softmax(S _t: )*U _:j )

Wherein, U' is the first parameter, softmax is the classifier, S _t: is all the data in the t-th row in the similarity matrix, and U _:j is all the data in the j-th column in the question semantic information.

Wherein, determining the second parameter based on the similarity matrix, the text semantic information and the classifier includes: calculating the second parameter according to the following formula:

H'＝Σ _t (softmax(max(S _:j ))*H _t: )

Wherein, H' is the second parameter, softmax is the classifier, max is the maximum function, S _:j is all the data in the jth column in the similarity matrix, Ht _: is the text semantic information. All data in row t.

Wherein, determining the context representation parameters of perceptible questions according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model includes, based on the following splicing The context characterization parameters are determined by:

G=β(H;U';H*U';H*H')

Wherein, G is the context representation parameter, β is the fully connected network, H is the text semantic information, U' is the first parameter, and H' is the second parameter.

Wherein, after the step of inputting question semantic information and text semantic information into the preset question answering model, the method of answering questions based on text also includes:

Encoding the text semantic information and the question semantic information respectively;

Use the encoded text semantic information and the encoded question semantic information to replace the classifier in the question semantic information, the text semantic information and the preset question answering model, and determine the difference between each word in the text semantic information. The corresponding first parameter, the question semantic information and the text semantic information in the second parameter step respectively corresponding to each word in the question semantic information, and perform subsequent steps.

Wherein, encoding the text semantic information and the question semantic information separately includes:

The text semantic information is processed as follows:

Perform vectorization processing on the text semantic information to obtain text semantic vectors;

Input the text semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the text semantic vector to obtain the first processing result;

Input the first processing result to the feedforward neural network in the preset question answering model, and use L2norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result;

The semantic information of the problem is processed as follows:

Vectorize the question semantic information to obtain a question semantic vector;

Input the question semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the question semantic vector to obtain the first processing result;

The first processing result is input to the feedforward neural network in the preset question answering model, and L2norm is used to process the output result of the feedforward neural network and the first processing result to obtain the second processing result.

An embodiment of the present invention provides a device for answering questions based on text, including:

The first determination unit is used to input question semantic information and text semantic information into the preset question answering model, and determine the text semantic information based on the question semantic information, text semantic information and the classifier in the preset question answering model. The first parameter corresponding to each word in the question semantic information, and the second parameter corresponding to each word in the question semantic information; wherein, the first parameter includes the question semantic information, respectively related to each word in the text semantic information. The semantic information, the second parameter includes semantic information in the text semantic information that is respectively related to each word in the question semantic information;

A second determination unit configured to determine context representation parameters of perceptible questions based on the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model;

An answering unit, configured to output a start index and an end index of answering the question in the text semantic information according to the question semantic information and the context characterization parameter.

An embodiment of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein,

When the processor executes the computer program, the following method steps are implemented:

Input the question semantic information and the text semantic information into the preset question answering model, and determine the first words corresponding to each word in the text semantic information based on the question semantic information, the text semantic information and the classifier in the preset question answering model. A parameter, and a second parameter respectively corresponding to each word in the question semantic information; wherein the first parameter includes semantic information in the question semantic information, respectively related to each word in the text semantic information, the said third parameter The second parameter contains the semantic information in the text semantic information and related to each word in the question semantic information;

Determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model;

According to the question semantic information and the context characterization parameter, a start index and an end index of answering the question are output in the text semantic information.

Embodiments of the present invention provide a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the following method steps are implemented:

Input the question semantic information and the text semantic information into the preset question answering model, and determine the first words corresponding to each word in the text semantic information based on the question semantic information, the text semantic information and the classifier in the preset question answering model. A parameter, and a second parameter respectively corresponding to each word in the question semantic information; wherein the first parameter includes semantic information in the question semantic information, respectively related to each word in the text semantic information, the said third parameter The second parameter contains the semantic information in the text semantic information and related to each word in the question semantic information;

Determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model;

According to the question semantic information and the context characterization parameter, a start index and an end index of answering the question are output in the text semantic information.

The method and device for answering questions based on text provided by the embodiments of the present invention determine the contextual representation parameters of the perceptible question through the first parameter and the second parameter that can reflect the bidirectional semantic relationship between the question semantic information and the text semantic information, and then It can accurately output the starting index and ending index of answering the question in the text semantic information.

Description of the drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of an embodiment of a method for answering questions based on text according to the present invention;

Figure 2 is a flow chart of another embodiment of the method for answering questions based on text according to the present invention;

Figure 3 is a schematic structural diagram of an embodiment of a device for answering questions based on text according to the present invention;

FIG. 4 is a schematic diagram of the physical structure of an electronic device according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

Figure 1 is a flow chart of a method for answering questions based on text according to the present invention. As shown in Figure 1, a method for answering questions based on text provided by an embodiment of the present invention includes the following steps:

S101: Input question semantic information and text semantic information into the preset question answering model, and determine the corresponding words in the text semantic information based on the question semantic information, text semantic information and the classifier in the preset question answering model. The first parameter, and the second parameter respectively corresponding to each word in the question semantic information; wherein, the first parameter includes the semantic information in the question semantic information, respectively related to each word in the text semantic information, the The second parameter includes semantic information in the text semantic information that is respectively related to each word in the question semantic information.

Specifically, the question semantic information and the text semantic information are input to the preset question answering model, and based on the question semantic information, the text semantic information and the classifier in the preset question answering model, the respective words in the text semantic information are determined. The corresponding first parameter, and the second parameter respectively corresponding to each word in the question semantic information; wherein, the first parameter includes the semantic information in the question semantic information, respectively related to each word in the text semantic information, The second parameter includes semantic information in the text semantic information that is respectively related to each word in the question semantic information. The method steps may be performed by a computer device. The preset question answering model can be a neural network and is not specifically limited. The main source of data for question semantic information and text semantic information can come from Baidu Zhidao data, because there are a large number of topic posts on Baidu Zhidao, and many people's extended meaning interpretation, irony and other information under the topic are real people. Data information communicated with people. The question on Baidu Knows is the query in the model, and several answers to the corresponding question are documents.

Modular design can be used to implement the embodiments of the present invention. Figure 2 is a flow chart of another embodiment of the method for answering questions based on text of the present invention. As shown in Figure 2, it mainly includes three modules: (1) Encoding module (corresponding to Query Encoder and Documents Encoder): query and documents are input to the encoding module after Embedding (vectorization). Among them, the encoding module can use three transformer (corresponding to u1 ~ u3 or h1 ~ h3) encoders, using the self-attention mechanism. It speeds up training and allows the model to remember long-distance semantic information; (2) Bidirectional attention mechanism module (corresponding to Attention Flow Match Layer): query semantic information and documents semantic information can be obtained from the encoding module, through the similarity matrix Calculate the relationship between words, design a two-way attention mechanism, and input it into the feedforward neural network to obtain their semantic relationship vectors; (3) Decoding module: With the help of the existing attention mechanism formula, Pointer Networks is used to obtain documents the starting index and ending index.

The detailed description of the above three modules is as follows:

For (1) encoding module:

First, the input of the encoding module is Embedding's query and documents. You can use Tencent's word vector file to look up the table and map each word into a 200-dimensional vector.

The encoding module is mainly composed of three transformers. The word vector at each position of each transformer will pass through a self-attention layer, and then the previous word vectors and outputs will be added together and processed using L2norm, so as to avoid Gradient dispersion. The final output is then passed through the feedforward neural network, and then the previous outputs are added together and processed using L2norm. That is, the text semantic information and the question semantic information are encoded separately. Referring to Figure 2, the encoded text semantic information and the encoded question semantic information can be used to replace the above-mentioned problem semantic information, text semantic information and The classifier in the preset question answering model determines the first parameter corresponding to each word in the text semantic information and the second parameter corresponding to each word in the question semantic information. The question semantic information and Text semantic information, and perform subsequent steps, such as subsequent steps S102 and S103. Further, taking text semantic information as an example, the following processing is performed:

Perform vectorization processing on the text semantic information to obtain text semantic vectors (corresponding to words in Figure 2); input the text semantic vector to the self-attention layer in the preset question answering model, and use L2norm (that is, Euclidean distance) processes the output result of the self-attention layer and the text semantic vector to obtain the first processing result; input the first processing result to the feedforward neural in the preset question answering model network, and use L2norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result. For question semantic information, please refer to the description of text semantic information and will not go into details.

The reason why LSTM is not used to obtain the semantic information of sentences is that the average number of words in documents is more than 300, so it is difficult to train LSTM and gradient dispersion is easy to occur. The use of three-layer transformer can effectively avoid this problem, and can also speed up training.

For (2) bidirectional attention mechanism module:

The input of this module is the query semantic information U and documents semantic information H obtained through the encoding module. The output is a problem-aware context representation parameter G.

First, the similarity matrix of H and U can be calculated (corresponding to determining the similarity matrix based on the question semantic information and text semantic information):

S _tj =α(H _t ,U _j )

S _tj represents the similarity matrix between the t-th column vector h in H and the j-th column vector u in U. α represents the mapping function that can be trained, where α = W [h; u; h*u], W represents the weight matrix, as shown in Figure 2, the similarity matrix is used to share the matrix documents-to-query and query- to-documents attention in two directions, where the t-th row represents: the similarity between the t-th word in the text and each word in the question, and the j-th column represents: the j-th word in the question and The similarity of each word in the text.

(1).Documents-to-query Attention calculates: for each document word, which query words are most relevant to it. The calculation formula is as follows (corresponding to determining the first parameter based on the similarity matrix, the question semantic information and the classifier):

U'＝Σ _j (softmax(S _t: )*U _:j )

Wherein, U' is the first parameter, softmax is the classifier, S _t: is all the data in the t-th row in the similarity matrix, and U _:j is all the data in the j-th column in the question semantic information.

Each row of the similarity matrix S _tj is directly used as an attention value through the softmax layer, because each row in S _tj represents the similarity between the t-th word in the text and each word in the question, documents-to- Query represents the impact of documents on query, and then weighted and summed with each column of U to obtain U'.

(2).query-to-documents Attention calculates: for each query words, which documents words are most relevant to it. The calculation formula is as follows (corresponding to determining the second parameter based on the similarity matrix, the text semantic information and the classifier):

H'＝Σ _t (softmax(max(S _:j ))*H _t: )

Wherein, H' is the second parameter, softmax is the classifier, max is the maximum function, S _:j is all the data in the jth column in the similarity matrix, Ht _: is the text semantic information. All data in row t.

Since these texts are very important to answer the question, we directly take the largest column in the similarity matrix, perform softmax normalization on it, calculate the weighted sum of the documents vectors, and obtain H'.

(3). The U' and H' obtained in the first two steps are spliced and input into the fully connected network to obtain the problem-aware context representation G. The splicing method can be G=β(H; U'; H*U'; H *H'), where β is a fully connected network.

For (3) decoding module:

The input of this module is the query semantic information U obtained through the encoding module and the problem-aware context representation G obtained through the bidirectional attention mechanism module. The output is the starting index and ending index on the documents.

Based on the characteristics of machine reading comprehension, the problem is solved in a pointer-like manner. Each pointer corresponds to an element of the input sequence, so that the input sequence can be directly manipulated without the need to specifically set the output vocabulary.

The query semantic information U and the context representation G are combined through the tanh activation function, and finally the maximum starting and ending index points of the documents are output through softmax.

Through machine reading and understanding based on the transformer network structure, the shortcomings of slow training time and long update time of the previous use of recurrent neural networks are overcome. This machine reading comprehension also combines information between questions and documents to improve output accuracy.

S102: Determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model.

Specifically, the context representation parameters of the perceptible question are determined according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model. Further, the context characterization parameters can be determined according to the following splicing method:

G=β(H;U';H*U';H*H')

Wherein, G is the context representation parameter, β is the fully connected network, H is the text semantic information, U' is the first parameter, and H' is the second parameter. Please refer to the above description and will not repeat them again.

S103: According to the question semantic information and the context representation parameter, output the start index and the end index of answering the question in the text semantic information.

Specifically, according to the question semantic information and the context representation parameter, the starting index and the ending index of answering the question are output in the text semantic information. Please refer to the above description and will not repeat them again.

The text-based question answering method provided by the embodiment of the present invention determines the context representation parameters of the perceptible question through the first parameter and the second parameter that can reflect the bidirectional semantic relationship between the question semantic information and the text semantic information, and thereby can Accurately output the starting index and ending index of answering the question in the text semantic information.

On the basis of the above embodiments, based on the question semantic information, the text semantic information and the classifier in the preset question answering model, the first parameters respectively corresponding to each word in the text semantic information and the question are determined. The second parameters corresponding to each word in the semantic information include:

Specifically, a similarity matrix is determined based on the question semantic information and the text semantic information; the rows of the similarity matrix are the similarities between all words in the question semantic information corresponding to each word in the text semantic information, so The columns of the similarity matrix are the similarities between all words in the text semantic information corresponding to each word in the question semantic information; please refer to the above description and will not be described again.

Specifically, the first parameter is determined based on the similarity matrix, the question semantic information and the classifier, and the first parameter is determined based on the similarity matrix, the text semantic information and the classifier. The second parameter. Please refer to the above description and will not repeat them again.

The method for answering questions based on text provided by the embodiment of the present invention determines the first parameter and the second parameter by introducing a similarity matrix, ensuring the rationality of the determination of the first parameter and the second parameter, and further being able to accurately output answers in the text semantic information. The starting index and ending index of the question.

On the basis of the above embodiment, determining the first parameter based on the similarity matrix, the question semantic information and the classifier includes: specifically, determining the first parameter according to the following formula:

U'＝Σ _j (softmax(S _t: )*U _:j )

Wherein, U' is the first parameter, softmax is the classifier, S _t: is all the data in the t-th row in the similarity matrix, and U _:j is all the data in the j-th column in the question semantic information. Please refer to the above description and will not repeat them again.

The method for answering questions based on text provided by the embodiment of the present invention determines the first parameter through a specific formula, further ensuring the rationality of the determination of the first parameter, and helping to accurately output the starting index and ending index of answering the question in the text semantic information. .

Based on the above embodiment, determining the second parameter based on the similarity matrix, the text semantic information and the classifier includes: specifically, calculating the second parameter according to the following formula:

H'＝Σ _t (softmax(max(S _:j ))*H _t: )

Wherein, H' is the second parameter, softmax is the classifier, max is the maximum function, S _:j is all the data in the jth column in the similarity matrix, Ht _: is the text semantic information. All data in row t. Please refer to the above description and will not repeat them again.

The method for answering questions based on text provided by the embodiment of the present invention determines the second parameter through a specific formula, further ensuring the rationality of the determination of the second parameter, and helping to accurately output the starting index and ending index of answering the question in the text semantic information. .

On the basis of the above embodiments, the context representation parameters of perceptible questions are determined based on the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model. , including, specifically, determining the context characterization parameters according to the following splicing method:

G=β(H;U';H*U';H*H')

Wherein, G is the context representation parameter, β is the fully connected network, H is the text semantic information, U' is the first parameter, and H' is the second parameter. Please refer to the above description and will not repeat them again. Please refer to the above description and will not repeat them again.

The method for answering questions based on text provided by embodiments of the present invention determines context representation parameters through specific splicing methods, and can further accurately output the starting index and ending index of answering questions in the text semantic information.

Based on the above embodiments, after the step of inputting question semantic information and text semantic information into the preset question answering model, the method for answering questions based on text also includes:

Specifically, the text semantic information and the question semantic information are encoded separately; reference can be made to the above description, which will not be described again.

Specifically, the coded text semantic information and the coded question semantic information are used to replace the classifier in the question semantic information, the text semantic information and the preset question answering model, and the classifier in the text semantic information is determined. The first parameter corresponding to each word, and the question semantic information and text semantic information in the second parameter step corresponding to each word in the question semantic information, and subsequent steps are performed. Please refer to the above description and will not repeat them again.

The text-based question answering method provided by the embodiment of the present invention can avoid gradient dispersion, thereby improving the robustness of the model.

On the basis of the above embodiments, encoding the text semantic information and the question semantic information respectively includes:

The text semantic information is processed as follows:

Specifically, the text semantic information is vectorized to obtain the text semantic vector; reference can be made to the above description, which will not be described again.

Specifically, the text semantic vector is input to the self-attention layer in the preset question answering model, and L2norm is used to process the output result of the self-attention layer and the text semantic vector to obtain the first processing result. ; Please refer to the above description and will not go into details.

Specifically, the first processing result is input to the feedforward neural network in the preset question answering model, and L2norm is used to process the output result of the feedforward neural network and the first processing result to obtain the second Processing results; please refer to the above description and will not be repeated.

The semantic information of the problem is processed as follows:

Specifically, the question semantic information is vectorized to obtain the question semantic vector; reference can be made to the above description, which will not be described again.

Specifically, the question semantic vector is input to the self-attention layer in the preset question answering model, and L2norm is used to process the output result of the self-attention layer and the question semantic vector to obtain the first processing result. ; Please refer to the above description and will not go into details.

Specifically, the first processing result is input to the feedforward neural network in the preset question answering model, and L2norm is used to process the output result of the feedforward neural network and the first processing result to obtain the second process result. Please refer to the above description and will not repeat them again.

The text-based question answering method provided by the embodiment of the present invention can further avoid gradient dispersion and better improve the robustness of the model.

Figure 3 is a schematic structural diagram of an embodiment of a device for answering questions based on text of the present invention. As shown in Figure 3, an embodiment of the present invention provides a device for answering questions based on text, including a first determination unit 301, a second determination unit 302 and Answer unit 303, where:

The first determining unit 301 is used to input question semantic information and text semantic information into the preset question answering model, and determine the text semantic information based on the question semantic information, text semantic information and the classifier in the preset question answering model. The first parameter corresponding to each word in the question semantic information, and the second parameter corresponding to each word in the question semantic information; wherein, the first parameter includes the question semantic information, respectively related to each word in the text semantic information. The semantic information and the second parameter include semantic information in the text semantic information that is respectively related to each word in the question semantic information; the second determination unit 302 is used to determine according to the first parameter, the second parameter, The text semantic information and the fully connected network in the preset question answering model determine the contextual representation parameters of the perceptible question; the answering unit 303 is configured to determine the contextual representation parameters in the text according to the question semantic information and the contextual representation parameters. The starting index and ending index of the answer to the question are output in the semantic information.

Specifically, the first determination unit 301 is used to input question semantic information and text semantic information into the preset question answering model, and determine the text according to the question semantic information, text semantic information and the classifier in the preset question answering model. A first parameter corresponding to each word in the semantic information, and a second parameter corresponding to each word in the question semantic information; wherein, the first parameter includes each parameter in the question semantic information and in the text semantic information. The second parameter includes semantic information related to each word in the text semantic information and semantic information related to each word in the question semantic information; the second determination unit 302 is configured to determine according to the first parameter, the second parameter and the semantic information related to each word in the question semantic information. The fully connected network in the two parameters, the text semantic information and the preset question answering model, determines the contextual representation parameters of the perceptible question; the answering unit 303 is used to determine the context representation parameters of the perceptible question according to the question semantic information and the context representation parameters. The starting index and ending index of answering the question are output in the text semantic information.

The device for answering questions based on text provided in the embodiment of the present invention determines the context representation parameters of the perceptible question through the first parameter and the second parameter that can reflect the bidirectional semantic relationship between the question semantic information and the text semantic information, and thereby can Accurately output the starting index and ending index of answering the question in the text semantic information.

The device for answering questions based on text provided by the embodiment of the present invention can be specifically used to execute the processing procedures of each of the above method embodiments. Its functions will not be described in detail here, and reference can be made to the detailed description of the above method embodiments.

Figure 4 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. As shown in Figure 4, the electronic device includes: a processor (processor) 401, a memory (memory) 402 and a bus 403;

Among them, the processor 401 and the memory 402 complete communication with each other through the bus 403;

The processor 401 is used to call program instructions in the memory 402 to execute the methods provided by the above method embodiments, for example, including: inputting question semantic information and text semantic information into a preset question answering model, and answering the question according to the question. The classifier in the semantic information, the text semantic information and the preset question answering model determines the first parameters corresponding to each word in the text semantic information and the second parameters respectively corresponding to each word in the question semantic information. ; Wherein, the first parameter includes semantic information in the question semantic information that is respectively related to each word in the text semantic information, and the second parameter includes each word in the text semantic information and is respectively related to the question semantic information. Relevant semantic information; determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model; according to the question Semantic information and the context representation parameters, the starting index and the ending index of answering the question are output in the text semantic information.

This embodiment discloses a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer The methods provided by the above method embodiments can be executed, for example, including: inputting question semantic information and text semantic information into a preset question answering model, and classifying them according to the question semantic information, text semantic information and the preset question answering model. The device determines the first parameters corresponding to each word in the text semantic information and the second parameters corresponding to each word in the question semantic information; wherein the first parameter includes the question semantic information and the text. The semantic information and the second parameter respectively related to each word in the semantic information include the semantic information respectively related to each word in the question semantic information in the text semantic information; according to the first parameter and the second parameter , the text semantic information and the fully connected network in the preset question answering model determine the context representation parameters of the perceptible question; according to the question semantic information and the context representation parameters, output in the text semantic information Answer the question's starting index and ending index.

This embodiment provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions. The computer instructions cause the computer to execute the methods provided by the above method embodiments, for example, including : Input the question semantic information and the text semantic information into the preset question answering model, and determine the corresponding words respectively corresponding to each word in the text semantic information based on the question semantic information, the text semantic information and the classifier in the preset question answering model. The first parameter, and the second parameter respectively corresponding to each word in the question semantic information; wherein, the first parameter includes the semantic information in the question semantic information, respectively related to each word in the text semantic information, the said The second parameter includes semantic information in the text semantic information that is respectively related to each word in the question semantic information; according to the first parameter, the second parameter, the text semantic information and the preset question answering model The fully connected network in the system determines the context representation parameters of the perceptible question; according to the question semantic information and the context representation parameters, the starting index and the ending index of answering the question are output in the text semantic information.

Those of ordinary skill in the art can understand that all or part of the steps to implement the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, It includes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other various media that can store program codes.

The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., including a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A method of answering questions based on text, which is characterized by including: Input the question semantic information and the text semantic information into the preset question answering model, and determine the first words corresponding to each word in the text semantic information based on the question semantic information, the text semantic information and the classifier in the preset question answering model. A parameter, and a second parameter respectively corresponding to each word in the question semantic information; wherein the first parameter includes semantic information in the question semantic information, respectively related to each word in the text semantic information, the said third parameter The second parameter contains the semantic information in the text semantic information and related to each word in the question semantic information; Determine the context representation parameters of the perceptible question according to the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model; According to the question semantic information and the context representation parameters, output the starting index and the ending index of answering the question in the text semantic information; After the step of inputting question semantic information and text semantic information into the preset question answering model, the method for answering questions based on text also includes: Encoding the text semantic information and the question semantic information respectively; Use the encoded text semantic information and the encoded question semantic information to replace the classifier in the question semantic information, the text semantic information and the preset question answering model, and determine the difference between each word in the text semantic information. The corresponding first parameter, and the question semantic information and text semantic information in the second parameter step corresponding to each word in the question semantic information, and perform subsequent steps; The encoding of the text semantic information and the question semantic information respectively includes: The text semantic information is processed as follows: Perform vectorization processing on the text semantic information to obtain text semantic vectors; Input the text semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the text semantic vector to obtain the first processing result; Input the first processing result to the feedforward neural network in the preset question answering model, and use L2 norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result ; The semantic information of the problem is processed as follows: Vectorize the question semantic information to obtain a question semantic vector; Input the question semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the question semantic vector to obtain the first processing result; Input the first processing result to the feedforward neural network in the preset question answering model, and use L2 norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result . 2. The method for answering questions based on text according to claim 1, characterized in that the classifier in the question semantic information, text semantic information and the preset question answering model is determined to be consistent with the text semantic information. The first parameter corresponding to each word, and the second parameter corresponding to each word in the question semantic information, include: According to the question semantic information and the text semantic information, a similarity matrix is determined; the rows of the similarity matrix are the similarities between all words in the question semantic information corresponding to each word in the text semantic information, and the similarity is The columns of the matrix are the similarities between all words in the text semantic information corresponding to each word in the question semantic information; The first parameter is determined based on the similarity matrix, the question semantic information and the classifier, and the second parameter is determined based on the similarity matrix, the text semantic information and the classifier . 3. The method for answering questions based on text according to claim 2, wherein determining the first parameter according to the similarity matrix, the question semantic information and the classifier includes: The following formula determines the first parameter: U'＝Σ _j (softmax(S _t: )*U _:j ) Wherein, U' is the first parameter, softmax is the classifier, S _t: is all the data in the t-th row in the similarity matrix, and U _:j is all the data in the j-th column in the question semantic information. 4. The method for answering questions based on text according to claim 2, wherein determining the second parameter according to the similarity matrix, the text semantic information and the classifier includes: The second parameter is calculated using the following formula: H'＝Σ _t (softmax(max(S _:j ))*H _t: ) Wherein, H' is the second parameter, softmax is the classifier, max is the maximum function, S _:j is all the data in the jth column in the similarity matrix, Ht _: is the text semantic information. All data in row t. 5. The method for answering questions based on text according to claim 1, characterized in that the method according to the first parameter, the second parameter, the text semantic information and the preset question answering model. The fully connected network determines the context representation parameters of the perceptible problem, including determining the context representation parameters according to the following splicing method: G=β(H;U';H*U';H*H') Wherein, G is the context representation parameter, β is the fully connected network, H is the text semantic information, U' is the first parameter, and H' is the second parameter. 6. A device for answering questions based on text, characterized by including: The first determination unit is used to input question semantic information and text semantic information into the preset question answering model, and determine the text semantic information based on the question semantic information, text semantic information and the classifier in the preset question answering model. The first parameter corresponding to each word in the question semantic information, and the second parameter corresponding to each word in the question semantic information; wherein, the first parameter includes the question semantic information, respectively related to each word in the text semantic information. The semantic information, the second parameter includes semantic information in the text semantic information that is respectively related to each word in the question semantic information; A second determination unit configured to determine context representation parameters of perceptible questions based on the first parameter, the second parameter, the text semantic information and the fully connected network in the preset question answering model; An answering unit, configured to output the starting index and ending index of answering the question in the text semantic information according to the question semantic information and the context characterization parameter; The device for answering questions based on text also includes a coding module for: Encoding the text semantic information and the question semantic information respectively; Use the encoded text semantic information and the encoded question semantic information to replace the classifier in the question semantic information, the text semantic information and the preset question answering model, and determine the difference between each word in the text semantic information. The corresponding first parameter, and the question semantic information and text semantic information in the second parameter step corresponding to each word in the question semantic information, and perform subsequent steps; The encoding of the text semantic information and the question semantic information respectively includes: The text semantic information is processed as follows: Perform vectorization processing on the text semantic information to obtain text semantic vectors; Input the text semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the text semantic vector to obtain the first processing result; Input the first processing result to the feedforward neural network in the preset question answering model, and use L2 norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result ; The semantic information of the problem is processed as follows: Vectorize the question semantic information to obtain a question semantic vector; Input the question semantic vector to the self-attention layer in the preset question answering model, and use L2norm to process the output result of the self-attention layer and the question semantic vector to obtain the first processing result; Input the first processing result to the feedforward neural network in the preset question answering model, and use L2 norm to process the output result of the feedforward neural network and the first processing result to obtain the second processing result . 7. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that when the processor executes the computer program, any one of claims 1 to 5 is implemented. One step of the method. 8. A non-transitory computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.