CN114817478A - Text-based question answering method, device, computer equipment and storage medium - Google Patents

Abstract

The embodiment of the application belongs to the field of artificial intelligence and relates to a question and answer method and device based on a text, computer equipment and a storage medium, wherein the method comprises the following steps: acquiring a question text and an answer text corresponding to the question text; segmenting the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks; combining the question text with each semantic block respectively to obtain a plurality of question and answer value evaluation texts; respectively inputting the question-answer value evaluation texts into the question-answer value evaluation model to obtain question-answer value scores of semantic blocks in the question-answer value evaluation texts; screening each semantic block according to the question-answer value score to obtain a semantic block queue comprising at least one semantic block; and inputting the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text. The method and the device improve the question and answer accuracy based on the text.

Description

Text-based question answering method, device, computer equipment and storage medium

technical field

The present application relates to the field of artificial intelligence, and in particular, to a text-based question answering method, device, computer equipment and storage medium.

Background technique

With the development of computer technology, there are more and more applications of computers in natural language processing. Text-based intelligent question answering is an important topic in natural language processing. Its application scenario can be to obtain the question text first, and then determine the answer information corresponding to the question text in the longer answer text. Text-based intelligent question answering usually inputs long text into a model built on a neural network, and the model outputs answer information. The pre-trained language model has a self-attention mechanism, which can learn the semantic relationship contained in the language, and has a good effect in text-based intelligent question answering.

The training of the pre-trained language model consumes a lot of computing resources, and the pre-trained language model also has a limit on the length of the input text. In practical applications, the text to be processed is usually long. For this reason, the text is usually divided into multiple sub-texts that are smaller than the maximum input length of the model, and then input to the model for question and answer processing. However, this will lose a lot of important semantic information, resulting in low question answering accuracy.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present application is to provide a text-based question answering method, device, computer equipment and storage medium, so as to solve the problem of low question answering accuracy.

In order to solve the above technical problems, the embodiments of the present application provide a text-based question answering method, which adopts the following technical solutions:

Get the question text and its corresponding answer text;

Segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks;

Combining the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts;

Input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text;

Screening each semantic block according to the question and answer value score to obtain a semantic block queue including at least one semantic block;

Inputting the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text.

In order to solve the above technical problems, the embodiments of the present application also provide a text-based question and answer device, which adopts the following technical solutions:

The text acquisition module is used to acquire the question text and its corresponding answer text;

a text segmentation module, configured to segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks;

a combination module for combining the question text and each semantic block respectively to obtain a plurality of question-and-answer value evaluation texts;

A text input module, configured to input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text;

A screening module, configured to screen each semantic block according to the question-and-answer value score, to obtain a semantic block queue containing at least one semantic block;

An answer extraction module, configured to input the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text.

In order to solve the above-mentioned technical problems, the embodiment of the present application also provides a computer device, which adopts the following technical solutions:

Get the question text and its corresponding answer text;

Segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks;

Combining the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts;

Input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text;

Screening each semantic block according to the question and answer value score to obtain a semantic block queue including at least one semantic block;

Inputting the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text.

In order to solve the above technical problems, the embodiments of the present application also provide a computer-readable storage medium, which adopts the following technical solutions:

Get the question text and its corresponding answer text;

Segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks;

Combining the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts;

Input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text;

Screening each semantic block according to the question and answer value score to obtain a semantic block queue including at least one semantic block;

Inputting the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text.

Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects: obtaining question text and corresponding answer text, segmenting the answer text according to a preset semantic segmentation algorithm, and ensuring that the obtained semantic block has sufficiently rich semantic information; The question text and each semantic block are combined to obtain multiple question and answer value evaluation texts; the question and answer value evaluation text is input into the question and answer value evaluation model to obtain the question and answer value score. The question and answer value score measures the semantic block in the question and answer value evaluation text. Contribution and value, so as to select the semantic blocks with high question-answer value to form the semantic block queue; then input the question text and semantic block queue into the answer information extraction model, so that the model can accurately output the answer information according to the semantic blocks with high question-answer value In the present application, the semantic information of the semantic block is guaranteed during the text segmentation, and the semantic block with higher question-answer value is selected to extract the answer information, which improves the accuracy of the question-and-answer.

Description of drawings

In order to illustrate the solutions in the present application more clearly, the following will briefly introduce the accompanying drawings used in the description of the embodiments of the present application. For those of ordinary skill, other drawings can also be obtained from these drawings without any creative effort.

FIG. 1 is an exemplary system architecture diagram to which the present application can be applied;

Figure 2 is a flowchart of one embodiment of a text-based question answering method according to the present application;

3 is a schematic structural diagram of an embodiment of a text-based question answering device according to the present application;

FIG. 4 is a schematic structural diagram of an embodiment of a computer device according to the present application.

Detailed ways

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field of this application; the terms used herein in the specification of the application are for the purpose of describing specific embodiments only It is not intended to limit the application; the terms "comprising" and "having" and any variations thereof in the description and claims of this application and the above description of the drawings are intended to cover non-exclusive inclusion. The terms "first", "second" and the like in the description and claims of the present application or the above drawings are used to distinguish different objects, rather than to describe a specific order.

Reference herein to an "embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor a separate or alternative embodiment that is mutually exclusive of other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

In order to make those skilled in the art better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings.

As shown in FIG. 1 , the system architecture 100 may include terminal devices 101 , 102 , and 103 , a network 104 and a server 105 . The network 104 is a medium used to provide a communication link between the terminal devices 101 , 102 , 103 and the server 105 . The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.

The user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages and the like. Various communication client applications may be installed on the terminal devices 101 , 102 and 103 , such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, and the like.

The terminal devices 101, 102, and 103 may be various electronic devices that have a display screen and support web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III, moving picture experts). Compression Standard Audio Layer 3), MP4 (Moving PictureExperts Group Audio Layer IV, Moving Picture Experts Compression Standard Audio Layer 4) Players, Laptops and Desktops, etc.

The server 105 may be a server that provides various services, such as a background server that provides support for the pages displayed on the terminal devices 101 , 102 , and 103 .

It should be noted that the text-based question answering method provided by the embodiments of the present application is generally executed by the server, and accordingly, the text-based question answering device is generally set in the server.

It should be understood that the numbers of terminal devices, networks and servers in FIG. 1 are merely illustrative. There can be any number of terminal devices, networks and servers according to implementation needs.

Continuing to refer to FIG. 2, a flowchart of one embodiment of a text-based question answering method according to the present application is shown. The text-based question answering method includes the following steps:

Step S201, acquiring question text and its corresponding answer text.

In this embodiment, the electronic device (for example, the server shown in FIG. 1 ) on which the text-based question answering method runs may communicate with the terminal through a wired connection or a wireless connection. It should be pointed out that the above wireless connection methods may include, but are not limited to, 3G/4G/5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connections now known or developed in the future. connection method.

Specifically, the question text and the answer text corresponding to the question text are obtained. Among them, the question text is used to ask the question; the answer text contains the answer corresponding to the question text. This application seeks to locate specific and direct answer fragments in the answer text.

Step S202, segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks.

Specifically, the answer text is segmented to obtain multiple semantic blocks. The present application can identify text elements with actual semantics in the answer text (such as words of a specific category in the text) according to a preset semantic segmentation algorithm, and segment the question and answer text based on the identified text elements to ensure that each semantic Blocks have sufficiently rich semantic information.

Step S203, combining the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts.

Specifically, the question text and each semantic block are combined one by one to obtain a plurality of question-answer value evaluation texts. In one embodiment, for the i-th semantic block obtained by segmentation, the question-answer value evaluation text is in the form of [[CLS]Query[SEP]Si], where CLS and SEP represent separators, Query represents question text, and Si represents The i-th semantic block.

Step S204, input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text.

Among them, the question and answer value evaluation model is used to evaluate the value or contribution of semantic blocks in question and answer.

Specifically, each obtained semantic block is input into the question-and-answer value evaluation model, respectively, and the question-and-answer value score of each semantic block is obtained. In general, the more useful a semantic block is in question answering, the higher the question answering value score of the semantic block.

In one embodiment, the question-answer value evaluation model is constructed based on a pre-trained language model, where the pre-trained language model may be a model such as Bert, Roberta, or the like. The question answering value evaluation model needs to be pre-trained.

In step S205, each semantic block is screened according to the question-and-answer value score, and a semantic block queue including at least one semantic block is obtained.

Specifically, the question-answer value score is a number, and the higher the value, the greater the role of the semantic block in question-answering. Each semantic block can be screened according to the question-and-answer value score to obtain a semantic block queue, wherein the semantic block queue includes at least one semantic block.

In one embodiment, a preset value score threshold is obtained, the question and answer value score is compared with the value score threshold, and a semantic block whose question and answer value score is greater than the value score threshold is selected to construct a semantic block queue.

In one embodiment, the semantic blocks are sorted in descending order according to the size of the question and answer value score, and then at least one semantic block is selected to obtain a semantic block queue.

Step S206, input the question text and the semantic block queue into the answer information extraction model, and obtain the answer information corresponding to the question text in the answer text.

The answer information extraction model may be a model for extracting answer information from the semantic block queue. The answer information here refers to a text fragment directly related to the question text. Although the answer text contains the answer corresponding to the question text, the answer text may be longer and contain more information. Therefore, text fragments directly related to the question text are extracted from the semantic block queue by the answer information extraction model. For example, the question text is "Where is the capital of China?" and the answer text is a longer introductory text containing the sentence: "The capital of China is Beijing". The answer information extraction model of this application aims to extract "Beijing" from the answer text as answer information.

Specifically, the question text and the semantic block queue are input into the answer information extraction model, and the answer information extraction model can determine the text segment most closely related to the question text in each semantic block in the semantic block queue, and output it as the answer information .

In one embodiment, the answer information extraction model is constructed based on a pre-trained language model, where the pre-trained language model may be a Bert, Roberta or other model.

In this embodiment, the question text and its corresponding answer text are obtained, and the answer text is segmented according to a preset semantic segmentation algorithm to ensure that the obtained semantic block has sufficient semantic information; A question and answer value evaluation text; input the question and answer value evaluation text into the question and answer value evaluation model to obtain the question and answer value score. Then, the question text and the semantic block queue are input into the answer information extraction model, so that the model can accurately output the answer information according to the semantic blocks with high question and answer value; this application guarantees the semantic block during text segmentation. Semantic information, selecting semantic blocks with high question-answer value to extract answer information, which improves the accuracy of question-answering.

Further, the above step S202 may include: identifying the target word in the answer text; segmenting the answer text according to the target word and a preset text length condition to obtain a plurality of semantic blocks; wherein, the number of target words in the semantic block is equal to the preset number The threshold value and the text length of the semantic block are within the preset length range; or, the text length of the semantic block is equal to the value of the right endpoint of the preset length range; the quantity threshold and the preset length range are parameters in the preset semantic segmentation algorithm.

Specifically, natural language processing is performed on the answer text to identify target words in the answer text. Among them, the target word can be a specific category of words in the answer text. In one embodiment, the target words may be named entities as well as verbs in the answer text. The identified target words will be used as a reference element during text segmentation.

Semantic segmentation algorithms limit text segmentation from two dimensions: target word and text length. Based on the semantic segmentation algorithm, the answer text is segmented according to the target word and the preset text length condition. The text length condition means that the text length of the intercepted semantic block should be within the preset length interval [B _min , B _max ], B _min and B _max are preset values, for example, B _min =32 and B _max =64.

When dividing, it is necessary to intercept the text of length B _min first. If the text of length B _min already contains target words with a preset number of threshold _Ns , the text of length B _min is directly regarded as a semantic block; and the segmentation continues from the next position of the answer text.

If the number of target words in the text of length B _min does not reach the preset number threshold N _s , continue to search for the target words in the text of length B _min to B _max until the found target words reach the preset number threshold N _s , at this time, the answer text is truncated with the position of the last found target word as the end position, and a semantic block is obtained; and the segmentation is continued from the next position of the answer text.

If the number of target words in the text of length B _min does not reach the preset number threshold N _s , and the length of B _max is found, the number of target words still cannot reach the preset number threshold N _s , at this time at the length of B _max Truncate the answer text as the end position to get a semantic block; and continue the segmentation from the next position of the answer text.

In one embodiment, the semantic block includes a first part and a second part, wherein the text length of the first part is _Bmin . When dividing, it is necessary to intercept the text of length B _min to obtain the first part. Then, search for the target word in the text with the length of B _min and B _max until the found target word reaches the preset number threshold N _s , at this time, the answer text is truncated with the position of the last found target word as the end position to obtain For the second part, get a semantic block based on the first part and the second part; and continue the segmentation from the next position of the answer text. When searching from the length of B _min to the length of B _max , the number of target words still cannot reach the preset number threshold N _s . At this time, the answer text is truncated at the length of B _max as the end position to obtain the second part. According to the first part and the first part The second part gets a semantic block; and continues the segmentation from the next position in the answer text.

In this embodiment, the answer text is segmented according to the target word and the text length condition to obtain a plurality of semantic blocks, and the semantic block also has a sufficient number of target words under the condition that the text length condition is met, ensuring that the semantic block has enough rich semantic information .

Further, the above step S204 may include: for each question and answer value evaluation text, input the question and answer value evaluation text into each sub-model in the question and answer value evaluation model, respectively, to obtain multiple question and answer value sub-scores; Operation is performed to obtain the question and answer value score of the semantic block in the question and answer value evaluation text.

Specifically, the question answering value evaluation model can include multiple parallel sub-models, and the sub-models can be composed of a pre-trained language model, a fully connected layer, and an activation function. The pre-trained language model may include Bert, Roberta, LMO, GPT and other models. The activation function can be relu, sigmoid and other functions. In one embodiment, the question-answer value evaluation model may be composed of a pre-trained language model component, a fully connected layer, and an activation function, wherein the pre-trained language model component may be composed of multiple pre-trained language models. The question and answer value score output by the question and answer value evaluation model can be between 0 and 1.

For each semantic block, the question and answer value evaluation text is input into each sub-model in the question and answer value evaluation model, and each sub-model outputs the question and answer value sub-score of the semantic block in the question and answer value evaluation text, and a linear operation can be performed on each question and answer value sub-score. , such as a direct summation operation, to obtain the question-and-answer value score of the semantic block.

In one embodiment, the average of each question and answer value sub-score is calculated to obtain the question and answer value score of the semantic block.

In one embodiment, the weight of each question and answer value sub-score is determined by a preset weighting algorithm, for example, the weight of each question and answer value sub-score is determined by the AHP, and then the weighted sum of each question and answer value sub-score is calculated according to the weight, to obtain Question Answer Value Score for Semantic Blocks. By assigning different weights to the question-and-answer value sub-scores, the differences between different sub-models and the question-and-answer value sub-scores output by the sub-models can be introduced to ensure the accuracy of the question-and-answer value score calculation.

In this embodiment, the question-and-answer value evaluation model may include multiple different sub-models, each sub-model outputs question-and-answer value sub-scores for semantic blocks, and performs linear operations on the question-and-answer value sub-scores output by each sub-model to obtain a comprehensive number of sub-scores. The Q&A value score of the model evaluation results ensures the accuracy of the Q&A value score.

Further, before the above step S204, it may also include: acquiring a plurality of question-and-answer value training texts; each question-and-answer value training text is obtained by combining the training question text and each semantic block obtained by dividing the training answer text; Input the initial question and answer value evaluation model, and obtain the question and answer value prediction score of the semantic block in each question and answer value training text; obtain the semantic block label of the semantic block in each question and answer value training text; the semantic block label identifies whether the semantic block is related to the answer information; based on the obtained Calculate the value evaluation loss based on the Q&A value prediction score and semantic block label; adjust the model parameters of the initial Q&A value evaluation model according to the value evaluation loss, until the model converges, and the Q&A value evaluation model is obtained.

Specifically, this application requires pre-training to obtain a question-and-answer value evaluation model. During training, the training question text and the training answer text are obtained, and the training answer text is segmented according to the semantic segmentation algorithm to obtain multiple semantic blocks. The training question text and each semantic block are combined to obtain multiple question-answer value training texts.

The question and answer value training text is input into the initial question and answer value evaluation model, and the question and answer value prediction scores of the semantic blocks in each question and answer value training text are obtained. The Q&A value prediction score is the Q&A value score during the model training phase. Each semantic block has a semantic block label, and the semantic block label is used to display whether the semantic block is associated with the answer information. For example, when the semantic block is associated with the answer information, the semantic block label is 1; when the semantic block is not associated with the answer information, the semantic block label is 1. The block label is 0.

The value evaluation loss loss(Q) can be calculated according to the question and answer value prediction score and the semantic block label, loss(Q)=CrossEntropy(dis(Q),label(Q)). Among them, CrossEntropy represents the cross entropy loss function, dis(Q) represents the question-answer value prediction score of each semantic block, and label(Q) represents the semantic block label of each semantic block.

In one embodiment, after obtaining the question and answer value prediction score, filter each semantic block according to the question and answer value prediction score to obtain a semantic block queue, and then calculate the value according to the question and answer value prediction score and the semantic block label of each semantic block in the semantic block queue. Evaluation loss loss(Q), loss(Q) = CrossEntropy(dis(Q), label(Q)). Among them, CrossEntropy represents the cross entropy loss function, dis(Q) represents the question-answer value prediction score of each semantic block in the semantic block queue, and label(Q) represents the semantic block label of each semantic block in the semantic block queue.

In this embodiment, a plurality of question-and-answer value training texts are obtained and an initial question-and-answer value evaluation model is input to obtain the question-and-answer value prediction scores of the semantic blocks in each question-and-answer value training text; the semantic block labels of each semantic block are obtained; The semantic block label calculates the value evaluation loss, and then adjusts the initial question and answer value evaluation model, so that the question and answer value evaluation model obtained after training can accurately output the question and answer value score of the semantic block in the question and answer value text.

Further, the above step S205 may include: arranging each semantic block in descending order according to the question-and-answer value score to obtain a candidate queue; according to the arrangement order of the semantic blocks in the candidate queue, sequentially adding each semantic block to the initial semantic block queue, and counting The current queue length of the initial semantic block queue; when the current queue length is equal to the preset length threshold, the current initial semantic block queue is determined as the semantic block queue; when the current queue length is greater than the preset length threshold, according to the length threshold value The last added semantic block is truncated to get the semantic block queue.

Specifically, according to the descending order of the question and answer value scores of the semantic blocks, each semantic block is sorted in descending order to obtain a candidate queue. The initial state of the initial semantic block queue can be empty, and then the semantic blocks arranged in descending order in the candidate queue are added to the initial semantic block queue in sequence, and the current queue length of the initial semantic block queue is counted after each addition.

If the current queue length is less than the preset length threshold, continue to add semantic blocks; if the current queue length is exactly equal to the length threshold, for example, the length threshold is 512, 8 semantic blocks are added, and the length of each semantic block is 64 , at this time, stop adding semantic blocks, and determine the current initial semantic block queue as the semantic block queue. If after adding a semantic block, the current queue length is greater than the length threshold, truncate the last added semantic block according to the length threshold to ensure that the length of the remaining semantic block queue is equal to the length threshold to obtain the semantic block queue. For example, after adding 6 semantic blocks, the queue length is 384, the length of the 7th semantic block is 40, the length of the 8th semantic block is 42, and the length of the 9th semantic block is 50. After adding the 9th semantic block, the current If the queue length will exceed the length threshold, the 9th semantic block will be truncated. Among them, the length threshold can be the maximum text length allowed by the answer information extraction model.

In this embodiment, the semantic blocks are sorted in descending order according to the question-and-answer value score, and then the descending-ordered semantic blocks are added to the initial semantic block queue one by one, so as to select the most valuable semantic block for question-and-answer; when the current queue length exceeds When the length threshold is set, the last added semantic block is truncated in time to ensure that the length of the semantic block queue does not exceed the maximum text length allowed by the answer information extraction model.

Further, the above step S206 may include: combining the question text and each semantic block in the semantic block queue respectively to obtain a plurality of question and answer texts; inputting each question and answer text into an answer information extraction model respectively to obtain answer position information; based on the answer position information, the answer information corresponding to the question text is determined in the answer text.

Specifically, the question text and each semantic block in the semantic block queue are respectively combined to obtain a plurality of question and answer texts. In one embodiment, for the ith semantic block in the semantic block queue, the question and answer text is in the form of [[CLS]Query[SEP]Si], where CLS and SEP represent delimiters, Query represents question text, and Si represents The i-th semantic block.

Input each question and answer text into the answer information extraction model, and the answer information extraction model performs information processing on each question and answer text, and outputs the answer position information. The answer position information can be a vector, which can represent the start position and end position of the answer information in the semantic block; The answer position information may further include the number of the semantic block, which is used to indicate in which semantic block the answer information exists.

According to the answer position information, text fragments directly related to the question text can be cut out from the answer text to obtain the answer information.

In this embodiment, the question text and each semantic block in the semantic block queue are respectively combined to obtain the question and answer text, and the question and answer text is input into the answer information extraction model; the semantic blocks in the semantic block queue have the greatest contribution value to the question and answer, and the answer information The extraction model can accurately output the answer position information, so as to obtain the answer information accurately.

Further, before the above step S206, it may also include: obtaining the question and answer training text; the question and answer training text is obtained by combining the training question text and each semantic block in the training semantic block queue respectively; inputting the question and answer training text into the initial answer information extraction model to obtain Answer position prediction information of semantic blocks in Q&A training text; Obtain answer position label information of semantic blocks in Q&A training text; Calculate position evaluation loss based on the obtained answer position prediction information and answer position label information; Adjust initial answer information according to position evaluation loss The model parameters of the model are extracted until the model converges, and the answer information extraction model is obtained.

Specifically, the answer information extraction model needs to be obtained through pre-training. During training, the question-and-answer training text is obtained, and the question-and-answer training text is obtained by combining the training question text and each semantic block in the training semantic block queue, wherein the training semantic block queue comes from the training answer text corresponding to the training question text. The training of the answer information extraction model can be carried out on the basis of the question and answer value evaluation model, that is, the question and answer value evaluation model is trained first, and the training semantic block queue is generated through the trained question and answer value evaluation model.

The question and answer training text is input into the initial answer information extraction model, and the answer information extraction model predicts the position of the answer inside the semantic block in the question and answer training text, and outputs the answer position prediction information. Then, the answer position label information is obtained, and the answer position label information records the answer position inside the semantic block in the question answering training text in the form of a label.

The answer position prediction information and the answer position label information may be vectors composed of 0 and 1, where 1 indicates that the position belongs to the answer information, and 0 indicates that the position does not belong to the answer information.

The position evaluation loss can be calculated according to the answer position prediction information and the answer position label information, and the loss function can be a cross entropy loss function. With the goal of minimizing the loss of position evaluation, the model parameters of the initial answer information extraction model are adjusted until the model converges, and the answer information extraction model is obtained.

When optimizing the initial question and answer value evaluation model and the initial answer information extraction model, methods such as SGD and Adam can be used to optimize the model.

It can be understood that the initial answer information extraction model is trained based on the training semantic block queue, and the training semantic block queue is generated based on the question and answer value score output by the question and answer value evaluation model. The semantic blocks in the training semantic queue have high value in question and answer, and the initial answer The information extraction model can be trained on higher quality training samples, so the training time of the initial answer information extraction model can be shortened and the training overhead can be reduced.

In this embodiment, the question and answer training text is input into the initial answer information extraction model to obtain the answer position prediction information, and then the position evaluation loss is calculated with the answer position label information, and the model is adjusted according to the position evaluation loss to obtain the answer information extraction model, which ensures the answer information extraction. The model can accurately extract the answer information from the question and answer text.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. , when the program is executed, it may include the processes of the foregoing method embodiments. The aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM).

It should be understood that although the various steps in the flowchart of the accompanying drawings are sequentially shown in the order indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order and may be performed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but may be executed at different times, and the execution sequence is also It does not have to be performed sequentially, but may be performed alternately or alternately with other steps or at least a portion of sub-steps or stages of other steps.

Further referring to FIG. 3 , as an implementation of the method shown in FIG. 2 above, the present application provides an embodiment of a text-based question answering device. The device embodiment corresponds to the method embodiment shown in FIG. 2 . The device Specifically, it can be applied to various electronic devices.

As shown in FIG. 3 , the text-based question answering device 300 in this embodiment includes: a text acquisition module 301, a text segmentation module 302, a combination module 303, a text input module 304, a screening module 305, and an answer extraction module 306, wherein:

The text acquisition module 301 is used to acquire question text and corresponding answer text.

The text segmentation module 302 is configured to segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks.

The combination module 303 is used to combine the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts.

The text input module 304 is configured to input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text.

The screening module 305 is configured to screen each semantic block according to the question-and-answer value score to obtain a semantic block queue including at least one semantic block.

The answer extraction module 306 is configured to input the question text and the semantic block queue into the answer information extraction model to obtain answer information corresponding to the question text in the answer text.

In this embodiment, the question text and its corresponding answer text are obtained, and the answer text is segmented according to a preset semantic segmentation algorithm to ensure that the obtained semantic block has sufficient semantic information; A question and answer value evaluation text; input the question and answer value evaluation text into the question and answer value evaluation model to obtain the question and answer value score. Then, the question text and the semantic block queue are input into the answer information extraction model, so that the model can accurately output the answer information according to the semantic blocks with high question and answer value; this application guarantees the semantic block during text segmentation. Semantic information, selecting semantic blocks with high question-answer value to extract answer information, which improves the accuracy of question-answering.

In some optional implementations of this embodiment, the text segmentation module 302 may include: a target word recognition submodule and a text segmentation submodule, wherein:

The target word recognition submodule is used to identify target words in the answer text.

The text segmentation sub-module is used to segment the answer text according to the target word and the preset text length condition to obtain multiple semantic blocks; wherein, the number of target words in the semantic block is equal to the preset number threshold and the text length of the semantic block is at the preset value or, the text length of the semantic block is equal to the value of the right endpoint of the preset length interval; the quantity threshold and the preset length interval are parameters in the preset semantic segmentation algorithm.

In this embodiment, the answer text is divided according to the target word and the text length condition to obtain multiple semantic blocks. The semantic block also has a sufficient number of target words under the condition that the text length condition is met, which ensures that the semantic block has enough rich semantic information .

In some optional implementations of this embodiment, the text input module 304 may include: a text input sub-module and a score operation sub-module, wherein:

The text input sub-module is used to input the question-and-answer value evaluation text into each sub-model in the question-and-answer value evaluation model for each question-and-answer value evaluation text, and obtain multiple question-and-answer value sub-scores.

The score operation sub-module is used to perform linear operation on each question and answer value sub-score to obtain the question and answer value score of the semantic block in the question and answer value evaluation text.

In this embodiment, the question-and-answer value evaluation model may include multiple different sub-models, each sub-model outputs question-and-answer value sub-scores for semantic blocks, and performs linear operations on the question-and-answer value sub-scores output by each sub-model to obtain a comprehensive number of sub-scores. The Q&A value score of the model evaluation results ensures the accuracy of the Q&A value score.

In some optional implementations of this embodiment, the text-based question answering apparatus 300 may further include: a question and answer training acquisition module, a question and answer training input module, a label acquisition module, a value loss calculation module, and a value model adjustment module, wherein:

The question and answer training acquisition module is used to obtain multiple question and answer value training texts; each question and answer value training text is obtained by combining the training question text and each semantic block obtained by dividing the training answer text;

The question and answer training input module is used to input each question and answer value training text into the initial question and answer value evaluation model, and obtain the question and answer value prediction score of the semantic block in each question and answer value training text;

The label acquisition module is used to obtain the semantic block labels of the semantic blocks in each question-and-answer value training text; the semantic block labels identify whether the semantic blocks are related to the answer information;

The value loss calculation module is used to calculate the value evaluation loss based on the obtained question and answer value prediction scores and semantic block labels;

The value model adjustment module is used to adjust the model parameters of the initial question and answer value assessment model according to the value assessment loss, until the model converges, and the question and answer value assessment model is obtained.

In this embodiment, a plurality of question-and-answer value training texts are obtained and an initial question-and-answer value evaluation model is input to obtain the question-and-answer value prediction scores of the semantic blocks in each question-and-answer value training text; the semantic block labels of each semantic block are obtained; The semantic block label calculates the value evaluation loss, and then adjusts the initial question and answer value evaluation model, so that the question and answer value evaluation model obtained after training can accurately output the question and answer value score of the semantic block in the question and answer value text.

In some optional implementations of this embodiment, the screening module 305 may include: a descending order sub-module, a semantic block adding sub-module, a current determination sub-module, and a semantic-block truncation sub-module, wherein:

The descending order sub-module is used to sort each semantic block in descending order according to the question and answer value score to obtain a candidate queue.

The semantic block adding sub-module is used to sequentially add each semantic block to the initial semantic block queue according to the arrangement order of the semantic blocks in the candidate queue, and count the current queue length of the initial semantic block queue.

The current determining submodule is used to determine the current initial semantic block queue as the semantic block queue when the current queue length is equal to the preset length threshold.

The semantic block truncation sub-module is used to truncate the last added semantic block according to the length threshold when the current queue length is greater than the preset length threshold to obtain the semantic block queue.

In this embodiment, the semantic blocks are sorted in descending order according to the question-and-answer value score, and then the descending-ordered semantic blocks are added to the initial semantic block queue one by one, so as to select the most valuable semantic block for question-and-answer; when the current queue length exceeds When the length threshold is set, the last added semantic block is truncated in time to ensure that the length of the semantic block queue does not exceed the maximum text length allowed by the answer information extraction model.

In some optional implementations of this embodiment, the answer extraction module 306 may include: a combination submodule, a question and answer input submodule, and an answer determination submodule, wherein:

The combination sub-module is used to combine the question text and each semantic block in the semantic block queue to obtain multiple question and answer texts.

The question and answer input sub-module is used to input each question and answer text into the answer information extraction model to obtain the answer position information.

The answer determination submodule is used to determine the answer information corresponding to the question text in the answer text based on the answer position information.

In this embodiment, the question text and each semantic block in the semantic block queue are respectively combined to obtain the question and answer text, and the question and answer text is input into the answer information extraction model; the semantic blocks in the semantic block queue have the greatest contribution value to the question and answer, and the answer information The extraction model can accurately output the answer position information, so as to obtain the answer information accurately.

In some optional implementations of this embodiment, the text-based question answering apparatus 300 may further include: a training text acquisition module, a training text input module, a position label acquisition module, a position loss calculation module, and an extraction model adjustment module, wherein:

The training text obtaining module is used to obtain the question and answer training text; the question and answer training text is obtained by combining the training question text and each semantic block in the training semantic block queue respectively.

The training text input module is used to input the question and answer training text into the initial answer information extraction model, and obtain the answer position prediction information of the semantic block in the question and answer training text.

The position label obtaining module is used to obtain the answer position label information of the semantic block in the question answering training text.

The position loss calculation module is used to calculate the position evaluation loss based on the obtained answer position prediction information and answer position label information.

The extraction model adjustment module is used to adjust the model parameters of the initial answer information extraction model according to the position evaluation loss, until the model converges, and the answer information extraction model is obtained.

In this embodiment, the question and answer training text is input into the initial answer information extraction model to obtain the answer position prediction information, and then the position evaluation loss is calculated with the answer position label information, and the model is adjusted according to the position evaluation loss to obtain the answer information extraction model, which ensures the answer information extraction. The model can accurately extract the answer information from the question and answer text.

To solve the above technical problems, the embodiments of the present application further provide computer equipment. Please refer to FIG. 4 for details. FIG. 4 is a block diagram of a basic structure of a computer device according to this embodiment.

The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that communicate with each other through a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure, but it should be understood that it is not required to implement all of the shown components, and more or less components may be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculation and/or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, special-purpose Integrated circuit (ApplicationSpecific Integrated Circuit, ASIC), programmable gate array (Field-Programmable GateArray, FPGA), digital processor (Digital Signal Processor, DSP), embedded equipment, etc.

The computer equipment may be a desktop computer, a notebook computer, a palmtop computer, a cloud server and other computing equipment. The computer device can perform human-computer interaction with the user through a keyboard, a mouse, a remote control, a touch pad or a voice control device.

The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static Random Access Memory (SRAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Programmable Read Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4 , such as a hard disk or a memory of the computer device 4 . In other embodiments, the memory 41 may also be an external storage device of the computer device 4 , such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (FlashCard) and so on. Of course, the memory 41 may also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of a text-based question-and-answer method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

The processor 42 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. This processor 42 is typically used to control the overall operation of the computer device 4 . In this embodiment, the processor 42 is configured to execute computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for executing the text-based question answering method.

The network interface 43 may include a wireless network interface or a wired network interface, and the network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

The computer device provided in this embodiment can execute the above text-based question answering method. The text-based question answering method here may be the text-based question answering method of each of the above embodiments.

In this embodiment, the question text and its corresponding answer text are obtained, and the answer text is segmented according to a preset semantic segmentation algorithm to ensure that the obtained semantic block has sufficient semantic information; A question and answer value evaluation text; input the question and answer value evaluation text into the question and answer value evaluation model to obtain the question and answer value score. Then, the question text and the semantic block queue are input into the answer information extraction model, so that the model can accurately output the answer information according to the semantic blocks with high question and answer value; this application guarantees the semantic block during text segmentation. Semantic information, selecting semantic blocks with high question-answer value to extract answer information, which improves the accuracy of question-answering.

The present application also provides another embodiment, that is, to provide a computer-readable storage medium, where the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to The at least one processor is caused to perform the steps of the text-based question answering method as described above.

In this embodiment, the question text and its corresponding answer text are obtained, and the answer text is segmented according to a preset semantic segmentation algorithm to ensure that the obtained semantic block has sufficient semantic information; A question and answer value evaluation text; input the question and answer value evaluation text into the question and answer value evaluation model to obtain the question and answer value score. Then, the question text and the semantic block queue are input into the answer information extraction model, so that the model can accurately output the answer information according to the semantic blocks with high question and answer value; this application guarantees the semantic block during text segmentation. Semantic information, selecting semantic blocks with high question-answer value to extract answer information, which improves the accuracy of question-answering.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence or in a part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, CD-ROM), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of this application.

Obviously, the above-described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The accompanying drawings show the preferred embodiments of the present application, but do not limit the scope of the patent of the present application. This application may be embodied in many different forms, rather these embodiments are provided so that a thorough and complete understanding of the disclosure of this application is provided. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or perform equivalent replacements for some of the technical features. . Any equivalent structure made by using the contents of the description and drawings of the present application, which is directly or indirectly used in other related technical fields, is also within the scope of protection of the patent of the present application.

Claims (10)

1. a text-based question and answer method, is characterized in that, comprises the following steps: Get the question text and its corresponding answer text; Segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks; Combining the question text and each semantic block respectively to obtain a plurality of question and answer value evaluation texts; Input each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtain the question and answer value score of the semantic block in each question and answer value evaluation text; Screening each semantic block according to the question and answer value score to obtain a semantic block queue including at least one semantic block; Inputting the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text. 2. The text-based question answering method according to claim 1, wherein the step of dividing the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks comprises: identifying target words in the answer text; The answer text is divided according to the target word and the preset text length condition to obtain a plurality of semantic blocks; wherein, the number of target words in the semantic block is equal to the preset number threshold and the text length of the semantic block is within the preset length interval Or, the text length of the semantic block is equal to the value of the right endpoint of the preset length interval; the quantity threshold and the preset length interval are parameters in the preset semantic segmentation algorithm. 3. text-based question and answer method according to claim 1, is characterized in that, described each question and answer value evaluation text is input respectively into question and answer value evaluation model, obtains the question and answer value score of the semantic block in described each question and answer value evaluation text. Steps include: For each question and answer value evaluation text, the question and answer value evaluation text is respectively input into each sub-model in the question and answer value evaluation model to obtain a plurality of question and answer value sub-scores; Perform a linear operation on each question-and-answer value sub-score to obtain the question-and-answer value score of the semantic block in the question-and-answer value evaluation text. 4. The text-based question and answer method according to claim 1, wherein each question and answer value evaluation text is respectively input into the question and answer value evaluation model to obtain the question and answer value score of the semantic block in each question and answer value evaluation text Before the steps, also include: Obtaining multiple question-answer-value training texts; each question-answer-value training text is obtained by combining the training question text and each semantic block obtained by dividing the training answer text; Inputting each question-and-answer value training text into an initial question-and-answer value evaluation model, to obtain the question-and-answer value prediction score of the semantic block in each question-and-answer value training text; Obtain the semantic block label of the semantic block in each question-and-answer value training text; the semantic block label identifies whether the semantic block is associated with the answer information; Calculate the value evaluation loss based on the obtained Q&A value prediction scores and semantic block labels; The model parameters of the initial question and answer value assessment model are adjusted according to the value assessment loss until the model converges, and the question and answer value assessment model is obtained. 5. The text-based question and answer method according to claim 1, wherein the step of screening each semantic block according to the question and answer value score to obtain a semantic block queue comprising at least one semantic block comprises: Arrange the semantic blocks in descending order according to the question and answer value score to obtain a candidate queue; According to the arrangement order of the semantic blocks in the candidate queue, each semantic block is added to the initial semantic block queue in turn, and the current queue length of the initial semantic block queue is counted; When the current queue length is equal to the preset length threshold, determining the current initial semantic block queue as the semantic block queue; When the current queue length is greater than a preset length threshold, the last added semantic block is truncated according to the length threshold to obtain a semantic block queue. 6 . The text-based question answering method according to claim 1 , wherein the question text and the semantic block queue are input into an answer information extraction model to obtain the difference between the question text and the question text in the answer text. 7 . The steps corresponding to the answer information include: Combining the question text and each semantic block in the semantic block queue respectively to obtain a plurality of question and answer texts; Input each question and answer text into the answer information extraction model respectively to obtain the answer position information; Based on the answer position information, answer information corresponding to the question text is determined in the answer text. 7 . The text-based question answering method according to claim 6 , wherein, in the answer information extraction model by inputting the question text and the semantic block queue into the answer information, the answer text and the question text are obtained in the answer text. 8 . Before the steps corresponding to the answer information, it also includes: Obtaining the question-and-answer training text; the question-and-answer training text is obtained by combining the training question text and each semantic block in the training semantic block queue respectively; Inputting the question-and-answer training text into an initial answer information extraction model to obtain the answer position prediction information of the semantic block in the question-and-answer training text; Obtain the answer position label information of the semantic block in the question answering training text; Calculate the position evaluation loss based on the obtained answer position prediction information and answer position label information; The model parameters of the initial answer information extraction model are adjusted according to the position evaluation loss until the model converges, and an answer information extraction model is obtained. 8. A text-based question and answer device, comprising: The text acquisition module is used to acquire the question text and its corresponding answer text; a text segmentation module, configured to segment the answer text according to a preset semantic segmentation algorithm to obtain a plurality of semantic blocks; a combination module for combining the question text and each semantic block respectively to obtain a plurality of question-and-answer value evaluation texts; The text input module is used for inputting each question and answer value evaluation text into the question and answer value evaluation model respectively, and obtains the question and answer value score of the semantic block in each question and answer value evaluation text; A screening module, configured to screen each semantic block according to the question-and-answer value score, to obtain a semantic block queue containing at least one semantic block; An answer extraction module, configured to input the question text and the semantic block queue into an answer information extraction model to obtain answer information corresponding to the question text in the answer text. 9. A computer device comprising a memory and a processor, wherein computer-readable instructions are stored in the memory, the processor implementing the computer-readable instructions as claimed in any one of claims 1 to 7 when the processor executes the computer-readable instructions Steps of a text-based question answering method. 10. A computer-readable storage medium, wherein computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, any one of claims 1 to 7 is implemented. The steps of the text-based question answering method described in item.

Images