Question generation method and device, computer device and storage medium

Abstract

This application relates to a question generation method, apparatus, computer device, and storage medium. The method includes: parsing a target paragraph and its corresponding target answer using a question generation model to obtain a target prediction question; parsing the target prediction question, the target paragraph, and the position of the target answer within the target paragraph using a reading comprehension model to determine a quality label for the target prediction question; adding the quality label to the target prediction question; and outputting the quality-labeled target prediction question. This method, by introducing a reading comprehension model, automatically determines the quality label for the target prediction question based on the target prediction question, the target paragraph, and the position of the answer within the target paragraph, eliminating the need for manual quality control of the generated target prediction questions, further improving the efficiency of database construction and reducing database construction costs.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a problem generation method, apparatus, computer device, and storage medium. Background Technology

[0002] With the development of artificial intelligence technology, products that provide intelligent question answering based on question generation technology are becoming increasingly popular. For example, such products can predict answers based on the target paragraph and question input by the user.

[0003] In order to provide users with information Q&A services quickly and accurately, intelligent Q&A products usually need to build a question database, which records text paragraphs, answers, and predicted user questions.

[0004] While existing technologies can predict user questions based on text paragraphs and answers using neural network models when building databases, the prediction results of neural network models are unstable, and manual quality control of the predicted questions is still required. This results in low efficiency and high cost in the database construction process. Summary of the Invention

[0005] Therefore, it is necessary to provide a problem generation method, apparatus, computer equipment, and storage medium that can improve the accuracy of problem generation in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a problem generation method. The method includes:

[0007] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0008] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0009] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0010] In one embodiment, a question generation model is used to parse the target paragraph and the target answer corresponding to the target paragraph to obtain the target prediction question, including:

[0011] The target paragraph and the target answer corresponding to the target paragraph are encoded through the coding network of the question generation model, so as to obtain the paragraph semantic coding result and paragraph position coding result corresponding to the target paragraph, as well as the answer semantic coding result and answer position coding result corresponding to the target answer;

[0012] By using the feature fusion network of the question generation model, the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are fused to obtain the fused features;

[0013] The target prediction question is obtained by decoding the fused features through the decoding network of the question generation model.

[0014] In one embodiment, the feature fusion network of the question generation model is used to fuse the paragraph semantic encoding result, the paragraph position encoding result, the answer semantic encoding result, and the answer position encoding result to obtain fused features, including:

[0015] The paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are concatenated to obtain the concatenated encoding result;

[0016] By using the feature fusion network of the question generation model, the splicing encoding results are fused to obtain fused features.

[0017] In one embodiment, a reading comprehension model is used to analyze the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the quality label of the target prediction question, including:

[0018] The reading comprehension model analyzes the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the degree of overlap between the answer to the target prediction question and the target answer.

[0019] The quality label for the target prediction problem is determined based on the degree of overlap between target locations.

[0020] In one embodiment, the question generation model is trained as follows:

[0021] Input the sample paragraph and the corresponding sample answer into the question generation model to obtain the sample prediction question;

[0022] Based on the sample prediction question and the sample standard question corresponding to the sample answer, determine the question similarity between the sample prediction question and the sample standard question;

[0023] Based on the sample prediction question, the sample paragraph, and the position of the sample answer within the sample paragraph, determine the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer.

[0024] The target loss value is determined based on the problem similarity and sample location overlap.

[0025] Update the model parameters of the problem generation model based on the target loss value.

[0026] In one embodiment, determining the question similarity between the sample prediction question and the sample standard question corresponding to the sample answer includes:

[0027] Input the sample prediction question and the corresponding standard question of the sample answer into the paraphrase recognition model to obtain the question similarity between the sample prediction question and the standard question.

[0028] In one embodiment, the degree of overlap between the sample position of the answer to the sample prediction question and the sample answer is determined based on the sample prediction question, the sample paragraph, and the position of the sample answer within the sample paragraph. This includes:

[0029] The sample prediction question, sample paragraph, and the position of the sample answer in the sample paragraph are input into the reading comprehension model to obtain the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer.

[0030] In one embodiment, the target loss value is determined based on problem similarity and sample location overlap, including:

[0031] The first loss value is determined based on the similarity of the questions and the overlap of the sample locations;

[0032] The second loss value is determined based on the first loss value, the restatement loss value and the restatement loss weight of the restatement recognition model;

[0033] The third loss value is determined based on the first loss value, the comprehension loss value of the reading comprehension model, and the comprehension loss weight.

[0034] The target loss value is determined based on the second and third loss values.

[0035] Secondly, this application also provides a problem generation apparatus. The apparatus includes:

[0036] The question generation module is used to parse the target paragraph and the target answer corresponding to the target paragraph through the question generation model to obtain the target prediction question;

[0037] The label determination module is used to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph through the reading comprehension model, and determine the quality label of the target prediction question.

[0038] The problem output module is used to add quality labels to the target prediction problem and output the target prediction problem with the added quality labels.

[0039] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0040] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0041] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0042] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0043] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0044] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0045] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0046] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0047] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0048] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0049] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0050] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0051] The aforementioned question generation method, apparatus, computer equipment, and storage medium, through a question generation model, parse the target paragraph and its corresponding target answer to obtain the target prediction question. Then, through a reading comprehension model, they analyze the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the quality label for the target prediction question. This quality label is then added to the target prediction question, and the tagged target prediction question is output. This solution, by introducing a reading comprehension model, automatically determines the quality label for the target prediction question based on the target prediction question, the target paragraph, and the position of the answer within the target paragraph. This eliminates the need for manual quality control of the generated target prediction questions, further improving the efficiency of database construction and reducing database construction costs. Attached Figure Description

[0052] Figure 1 This is an application environment diagram of a problem generation method provided in this embodiment;

[0053] Figure 2 This is a flowchart illustrating the first problem generation method provided in this embodiment;

[0054] Figure 3 This is a schematic diagram of a process for obtaining a target prediction problem provided in this embodiment;

[0055] Figure 4 This is a schematic diagram of the process of training a problem generation model provided in this embodiment;

[0056] Figure 5 This is a schematic diagram of a process for determining a target loss value provided in this embodiment;

[0057] Figure 6 This is a flowchart illustrating the second problem generation method provided in this embodiment;

[0058] Figure 7 This is a structural block diagram of the first problem generation device provided in this embodiment;

[0059] Figure 8 This is a structural block diagram of the second problem generation device provided in this embodiment;

[0060] Figure 9 This is a structural block diagram of the third problem generation device provided in this embodiment;

[0061] Figure 10 This is a structural block diagram of the fourth problem generation device provided in this embodiment;

[0062] Figure 11 This is a structural block diagram of the fifth problem generation device provided in this embodiment;

[0063] Figure 12 This is an internal structural diagram of a computer device provided in this embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] The problem generation method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, in one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 1 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores paragraph and answer data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a problem generation method.

[0066] This embodiment discloses a question generation method, apparatus, computer device, and storage medium. Through a question generation model, the target paragraph and the target answer corresponding to the target paragraph are analyzed to obtain the target prediction question. Then, through a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question. The quality label is then added to the target prediction question, and the target prediction question with the added quality label is output.

[0067] In one embodiment, such as Figure 2 As shown, a question generation method is provided, applicable to scenarios where questions corresponding to target answers are generated based on target paragraphs and their corresponding target answers. This method is then applied to… Figure 1 Taking the server in the example, the following steps are included:

[0068] S201 uses a question generation model to analyze the target paragraph and its corresponding target answer to obtain the target prediction question.

[0069] The question generation model can be a model used to predict the question corresponding to the target answer based on the target paragraph and the target answer corresponding to the target paragraph. Optionally, the question generation model in this embodiment can be built based on the Transform model.

[0070] The target paragraph can be a paragraph used to generate the predicted question during the question generation model phase. Optionally, this paragraph can be an article paragraph, or a paragraph of regulations, rules, or clauses.

[0071] The target answer can be one or more sentences from the target paragraph.

[0072] The target prediction question can be generated by a question generation model and corresponds to the target answer and the target paragraph.

[0073] Optionally, the server inputs the target paragraph and the target answer corresponding to the target paragraph into a pre-trained question generation model, which then parses the input target paragraph and target answer to obtain the target prediction question.

[0074] S202 uses a reading comprehension model to analyze the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the quality label of the target prediction question.

[0075] The reading comprehension model can be a model used to judge the degree of overlap between the generated answer and the target answer to the target prediction question. This reading comprehension model can perform prediction tasks using an open reading comprehension dataset (SQuAD, i.e., the QA dataset), or it can perform prediction tasks based on a pre-trained model (such as a pre-trained language representation model, Bidirectional Encoder Representation from Transformers, BERT), using an open-source Chinese machine reading comprehension dataset (Cloze-Style Machine Reading Comprehension, SC-MRC, CMRC).

[0076] The quality label can be a label that characterizes the quality of the target prediction question. Optionally, in this embodiment, the quality of the target prediction question can be measured by the accuracy of the target prediction question. In this embodiment, the quality of the target prediction question can be measured by the degree of overlap between the position of the target answer in the target paragraph and the position of the answer corresponding to the target prediction question in the target paragraph. The higher the position overlap, the better the quality of the target prediction question.

[0077] Optionally, this embodiment can use a reading comprehension model to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph, determine the target position overlap between the answer corresponding to the target prediction question and the target answer, and determine the quality label of the target prediction question based on the target position overlap.

[0078] Specifically, the server inputs the target prediction question, the target paragraph, and the position of the target answer within the target paragraph into a pre-set reading comprehension model. The reading comprehension model parses the input target prediction question, target paragraph, and position of the target answer within the target paragraph to obtain the positional overlap between the position of the answer to the target prediction question and the position of the target answer within the target paragraph, i.e., the target position overlap. Based on this target position overlap, it is compared with a pre-set target position overlap threshold. If the target position overlap obtained by the reading comprehension model is greater than the target position overlap threshold, then the target prediction question corresponding to the target position overlap obtained by the reading comprehension model meets the requirements, and the quality label of the target prediction question is determined to be a high-quality label; otherwise, the quality label of the target prediction question is determined to be a low-quality label.

[0079] S203 adds quality labels to the target prediction problem and outputs the target prediction problem with added quality labels.

[0080] Optionally, the server adds quality labels to the target prediction problem based on the determined quality labels, and outputs the target prediction problem with added quality labels for further human screening, or the server directly selects the target prediction problem with high-quality labels to perform subsequent operations.

[0081] This embodiment can construct a question database based on target prediction questions with added quality labels. When providing information question-and-answer services, the question database can prioritize displaying target prediction questions with high-quality labels for users to select and further obtain the answers to the target prediction questions.

[0082] The aforementioned question generation method uses a question generation model to parse the target paragraph and its corresponding target answer to obtain the target predicted question. Then, using a reading comprehension model, it analyzes the target predicted question, the target paragraph, and the position of the target answer within the target paragraph to determine the quality label for the target predicted question. This quality label is then added to the target predicted question, and the tagged target predicted question is output. This solution, by introducing a reading comprehension model, automatically determines the quality label for the target predicted question based on the target predicted question, the target paragraph, and the position of the answer within the target paragraph. This eliminates the need for manual quality checks on the generated target predicted questions, further improving the efficiency of database construction and reducing database construction costs.

[0083] In one embodiment, the problem generation model in the above embodiment can be a Transform model, and the problem generation model can include: an encoding network, a feature fusion network, and a decoding network, such as... Figure 3 As shown in the figure, this embodiment presents an optional method for obtaining the target prediction problem based on an encoding network, feature fusion network, and decoding network of a problem generation model, including the following steps:

[0084] S301 uses the coding network of the question generation model to encode the target paragraph and the target answer corresponding to the target paragraph, and obtains the paragraph semantic coding result and paragraph position coding result corresponding to the target paragraph, as well as the answer semantic coding result and answer position coding result corresponding to the target answer.

[0085] The encoding network can be a network used to encode the target paragraph and the target answer.

[0086] The paragraph semantic encoding result can be a multi-dimensional encoding result obtained by semantically parsing each word in the paragraph. The answer semantic encoding result can be a multi-dimensional encoding result obtained by semantically parsing each word in the answer.

[0087] The paragraph position encoding result can be the relative position encoding result of each character in the paragraph, obtained through sinusoidal position encoding processing, based on the paragraph semantic encoding result. Similarly, the answer position encoding result can be the relative position encoding result of each character in the answer, obtained through sinusoidal position encoding processing, based on the answer semantic encoding result.

[0088] Optionally, the server inputs the target paragraph and the corresponding target answer into the encoding network of a pre-set question generation model. The encoding network performs semantic parsing based on the input target paragraph and target answer to obtain the paragraph semantic encoding result of the target paragraph and the answer semantic encoding result of the target answer. Then, based on the paragraph semantic encoding result of the target paragraph, the pre-set calculation logic (such as sine position encoding) is used to obtain the paragraph position encoding result of the target paragraph. Based on the answer semantic encoding of the target answer, the pre-set calculation logic (such as sine position encoding) is used to obtain the answer position encoding result of the target answer.

[0089] S302 uses a feature fusion network based on the question generation model to fuse the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result to obtain fused features.

[0090] The feature fusion network can be a network that fuses features from paragraph semantic encoding results, paragraph position encoding results, answer semantic encoding results, and answer position encoding results. Optionally, the feature fusion network can be a multi-head attention mechanism network.

[0091] Optionally, the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are concatenated to obtain a concatenated encoding result. The feature fusion network of the question generation model is then used to perform feature fusion on the concatenated encoding result to obtain fused features. Specifically, the server calculates the paragraph summation encoding result based on pre-set computational logic (such as addition) using the paragraph semantic encoding result and the paragraph position encoding result. Similarly, it calculates the answer summation encoding result based on pre-set computational logic (such as addition) using the answer semantic encoding result and the answer summation encoding result. The paragraph summation encoding result is then added to the answer summation encoding result to obtain the concatenated encoding result. This concatenated encoding result is then input into the feature fusion network, where feature fusion is performed based on a multi-head attention mechanism to obtain a fused representation of the concatenated encoding result. This fused representation of the concatenated encoding result is then used as the fused feature.

[0092] S303 decodes the fused features through the decoding network of the question generation model to obtain the target prediction question.

[0093] The decoding network can contain multiple decoding layers connected in sequence.

[0094] Optionally, the server inputs the fused features into the decoding network of the problem generation model. The decoding network is based on a pre-set teacher-forcing learning method. The input of the first decoding layer of the decoding network is the fused features, and the input of each subsequent decoding layer is the weighted sum of the output solutions of the previous decoding layers. The weighted sum of the outputs of the last decoding layer is used as the target prediction problem.

[0095] The aforementioned method for predicting the target question encodes the target paragraph and its corresponding target answer using the encoding network of the question generation model. This yields the paragraph semantic encoding and paragraph position encoding results for the target paragraph, as well as the answer semantic encoding and answer position encoding results for the target answer. The feature fusion network of the question generation model then fuses these results to obtain fused features. Finally, the decoding network of the question generation model decodes these fused features to obtain the target prediction question. This method considers the paragraph semantic encoding, paragraph position encoding, answer semantic encoding, and answer position encoding results when encoding the target paragraph and target answer, increasing the diversity of the encoding results' dimensions. The multi-dimensional encoding feature calculation makes the obtained fused features more reasonable, and the target prediction question is then obtained based on these fused features, thus improving accuracy.

[0096] Figure 4 This is a flowchart illustrating the training process of a problem generation model in one embodiment. Currently, in existing technologies, optimization of the model is based solely on its own loss function, resulting in low accuracy and a lack of rationality. Therefore, to improve the rationality of model optimization, this embodiment provides an optional method for training a problem generation model, including the following steps:

[0097] S401 inputs the sample paragraph and the corresponding sample answer into the question generation model to obtain the sample prediction question.

[0098] The sample paragraph can be a paragraph used to generate questions during the training phase of the question generation model.

[0099] The sample answer can be a single sentence or multiple sentences from the sample paragraph.

[0100] Optionally, the server inputs the sample paragraph and the corresponding sample answer into a pre-set question generation model, which parses the sample paragraph and the sample answer to obtain the sample prediction question.

[0101] Specifically, the encoding network of the question generation model encodes the sample paragraph and the corresponding sample answer, obtaining the paragraph semantic encoding result and paragraph position encoding result, as well as the answer semantic encoding result and answer position encoding result. The feature fusion network of the question generation model fuses the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result and answer position encoding result to obtain fused features. The decoding network of the question generation model decodes the fused features to obtain the sample prediction question.

[0102] S402 determines the question similarity between the sample prediction question and the sample standard question corresponding to the sample answer based on the sample prediction question and the sample standard question.

[0103] Among them, question similarity can be the semantic similarity between the sample prediction question and the sample standard question.

[0104] Optionally, there are various ways to determine the question similarity between the sample prediction question and the sample standard question corresponding to the sample answer, and this application does not limit this.

[0105] One possible implementation is that the server calculates the question similarity between the sample prediction question and the sample standard question corresponding to the sample answer through pre-set calculation logic.

[0106] Another alternative implementation is to input the sample prediction question and the sample standard question corresponding to the sample answer into the paraphrase recognition model to obtain the question similarity between the sample prediction question and the sample standard question.

[0107] The paraphrase recognition model can predict whether the sample prediction question is similar to the sample standard question based on the sample prediction question and the sample standard question, and output the predicted binary classification probability (i.e., question similarity). This paraphrase recognition model can be fine-tuned based on a pre-trained model (such as a pre-trained language representation model, Bidirectional Encoder Representation from Transformers, BERT) using the open-source Chinese similarity question task dataset (ALarge-scale Chinese Question Matching Corpus, LCQMC).

[0108] Specifically, the server inputs the sample prediction question and the sample standard question corresponding to the sample answer into a pre-set paraphrase recognition model. The paraphrase recognition model calculates the question similarity between the sample prediction question and the sample standard question, obtains the binary classification probability between the sample prediction question and the sample standard question, and uses the binary classification probability as the question similarity between the sample prediction question and the sample standard question.

[0109] S403 determines the degree of overlap between the sample position of the answer to the sample prediction question and the sample answer based on the sample prediction question, the sample paragraph, and the position of the sample answer in the sample paragraph.

[0110] Optionally, there are various ways to determine the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer, based on the sample prediction question, the sample paragraph, and the position of the sample answer in the sample paragraph. This application does not limit this method.

[0111] One possible implementation is that the server inputs the sample prediction question and sample paragraph into a pre-set answer generation model to obtain the sample prediction answer corresponding to the sample prediction question, and parses the position of the sample prediction answer in the sample paragraph. Then, based on the pre-set calculation logic, the server calculates the sample position overlap between the answer corresponding to the sample prediction question and the sample answer.

[0112] Another alternative implementation involves inputting the sample prediction question, sample paragraph, and the position of the sample answer within the sample paragraph into a reading comprehension model to obtain the sample position overlap between the answer to the sample prediction question and the sample answer. Specifically, the server inputs the sample prediction question, sample paragraph, and the position of the sample answer within the sample paragraph into a pre-set reading comprehension model. This model parses the positions of the sample prediction question, sample paragraph, and sample answer within the sample paragraph to obtain the sample position overlap between the answer to the sample prediction question and the sample answer.

[0113] S404 determines the target loss value based on the similarity of the questions and the overlap of the sample locations.

[0114] The target loss value can be a parameter value used to optimize the problem generation model.

[0115] Optionally, the server calculates the loss value between the question similarity and the sample location overlap based on pre-set calculation logic (such as summation or mean calculation), and uses this loss value as the target loss value.

[0116] S405 updates the model parameters of the problem generation model based on the target loss value.

[0117] Optionally, the server replaces the model parameters of the problem generation model based on the calculated target loss value, that is, updates the model parameters of the problem generation model.

[0118] The aforementioned method for training the question generation model inputs sample paragraphs and their corresponding sample answers into the model to generate predicted questions. Based on these predicted questions and their corresponding standard questions, the similarity between the predicted and standard questions is determined. Then, based on the predicted questions, sample paragraphs, and the position of the sample answer within the paragraph, the overlap between the predicted and standard answers is determined. Finally, a target loss value is determined based on the question similarity and overlap, and the model parameters are updated accordingly. This method performs quality checks on the questions generated by the model, checking the similarity and overlap, making the calculated target loss value more reasonable. The model is then updated based on this target loss value, further improving the semantic accuracy and logical controllability of the generated questions.

[0119] Figure 5 This is a flowchart illustrating the process of determining the target loss value in one embodiment. Currently, in existing technologies, the accuracy of problem generation models is relatively low. Therefore, to improve the accuracy of the final generated problem, this embodiment provides an optional method for determining the target loss value, including the following steps:

[0120] S501 determines the first loss value based on the similarity of the questions and the overlap of the sample locations.

[0121] Optionally, the server calculates the loss value between the question similarity and the sample location overlap based on pre-set calculation logic, and uses this loss value as the first loss value.

[0122] S502 determines the second loss value based on the first loss value, the restatement loss value of the restatement recognition model, and the restatement loss weight.

[0123] Optionally, the server calculates a second loss value based on the first loss value, the paraphrase loss value and the paraphrase loss weight of the paraphrase recognition model, according to a pre-set calculation logic. For example, the formula for calculating the second loss value is shown in (1-1).

[0124]

[0125] in, It is the second loss value, α pqg It is a restatement of the loss weight. It is a restatement of the loss value, L mL This is the first loss value.

[0126] S503 determines the third loss value based on the first loss value, the comprehension loss value of the reading comprehension model, and the comprehension loss weight.

[0127] Optionally, the server calculates a third loss value based on the first loss value, the comprehension loss value of the reading comprehension model, and the comprehension loss weight, according to pre-set calculation logic. For example, the formula for calculating the third loss value is shown in (1-2).

[0128]

[0129] in, It is the third loss value, α mrc It's about understanding loss weights. It's about understanding the loss value, L mL This is the first loss value.

[0130] It should be noted that both the paraphrase loss and comprehension loss are derived from a reinforcement reward design based on an efficient gradient strategy (self-critical sequence training, SCST) and reward mechanism. The most important idea of ​​the gradient strategy is to use the encoding result of greedy search as a baseline. However, because the paraphrase recognition model and the reading comprehension model are trained using a fixed n:m ratio of training data, for example, if the paraphrase recognition model is trained with n-ratio training data and the machine reading comprehension model with m-ratio training data, the paraphrase loss and comprehension loss values, although calculated using the same reinforcement learning loss function, will yield different results. For example, the reinforcement learning loss function L for paraphrase loss and comprehension loss is... RL The calculation formula is shown in (1-3).

[0131]

[0132] Among them, L RL Here, r is the reinforcement learning loss function, and q is the reward function. s It is the encoding result of the generated questions obtained by sampling based on the output accuracy of all generated questions, q g P is the encoding result of the generator problem obtained by greedily sampling all generator problems. θ It is the gradient strategy for learning the parameters θ of the problem generation model.

[0133] S504 determines the target loss value based on the second and third loss values.

[0134] Optionally, the server calculates the sum of the second and third loss values ​​using pre-set calculation logic (such as addition), and uses this sum as the target loss value.

[0135] The aforementioned method for determining the target loss value determines a first loss value based on question similarity and sample location overlap. A second loss value is determined based on the first loss value and the paraphrase loss value and weight of the paraphrase recognition model. A third loss value is determined based on the first loss value and the comprehension loss value and weight of the reading comprehension model. Finally, the target loss value is determined based on the second and third loss values. This scheme, in determining the target loss, considers not only question similarity and sample location overlap, but also the paraphrase loss value and weight of the paraphrase recognition model, as well as the comprehension loss value and weight of the reading comprehension model. This multi-dimensional calculation makes the calculated second and third loss values ​​more specific, comprehensive, and reasonable. The target loss value is then calculated based on the determined second and third loss values, further improving the accuracy of the target loss value and making subsequent updates to the question generation model based on the target loss value more reasonable.

[0136] In one embodiment, this embodiment provides an optional method for generating questions, using the application of this method to a server as an example for illustration. For example... Figure 6 As shown, the method includes the following steps:

[0137] S601 inputs the sample paragraph and the corresponding sample answer into the question generation model to obtain the sample prediction question.

[0138] S602 inputs the sample prediction question and the sample standard question corresponding to the sample answer into the paraphrase recognition model to obtain the question similarity between the sample prediction question and the sample standard question.

[0139] S603 inputs the sample prediction question, the sample paragraph, and the position of the sample answer in the sample paragraph into the reading comprehension model to obtain the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer.

[0140] S604 determines the first loss value based on the similarity of the questions and the overlap of the sample locations.

[0141] S605 determines the second loss value based on the first loss value, the restatement loss value of the restatement recognition model, and the restatement loss weight.

[0142] S606 determines the third loss value based on the first loss value, the comprehension loss value of the reading comprehension model, and the comprehension loss weight.

[0143] S607 determines the target loss value based on the second and third loss values.

[0144] S608 updates the model parameters of the problem generation model based on the target loss value.

[0145] S609 uses the coding network of the question generation model to encode the target paragraph and the target answer corresponding to the target paragraph, and obtains the paragraph semantic coding result and paragraph position coding result corresponding to the target paragraph, as well as the answer semantic coding result and answer position coding result corresponding to the target answer.

[0146] S610 concatenates the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result to obtain the concatenated encoding result.

[0147] S611 uses a feature fusion network of the problem generation model to perform feature fusion on the splicing and encoding results to obtain fused features.

[0148] S612 decodes the fused features through the decoding network of the question generation model to obtain the target prediction question.

[0149] S613 uses a reading comprehension model to analyze the target prediction question, the target paragraph, and the position of the target answer within the target paragraph, determining the degree of overlap between the answer to the target prediction question and the target answer.

[0150] S614 determines the quality label for the target prediction problem based on the overlap of target locations.

[0151] S615 adds quality labels to the target prediction problem and outputs the target prediction problem with added quality labels.

[0152] It should be noted that in this embodiment, steps S601-S608 are the process of training the problem generation model, and steps S609-S615 are the process of generating problems using the problem generation model. Optionally, steps S601-S608 can be executed first to train the problem generation model, and then steps S609-S615 can be executed to generate problems using the problem generation model.

[0153] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0154] Based on the same inventive concept, this application also provides a problem generation apparatus for implementing the problem generation method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations in one or more of the problem generation apparatus embodiments provided below can be found in the limitations of the problem generation method described above, and will not be repeated here.

[0155] In one embodiment, such as Figure 7 As shown, a problem generation device 1 is provided, including: a problem generation module 10, a tag determination module 11, and a problem output module 12, wherein:

[0156] The question generation module 10 is used to parse the target paragraph and the target answer corresponding to the target paragraph through the question generation model to obtain the target prediction question;

[0157] The label determination module 11 is used to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph through the reading comprehension model, and determine the quality label of the target prediction question.

[0158] Problem output module 12 is used to add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0159] In one embodiment, such as Figure 8 As shown, Figure 7 The problem generation module 10 includes:

[0160] The encoding unit 100 is used to encode the target paragraph and the target answer corresponding to the target paragraph through the encoding network of the question generation model, so as to obtain the paragraph semantic encoding result and paragraph position encoding result corresponding to the target paragraph, as well as the answer semantic encoding result and answer position encoding result corresponding to the target answer;

[0161] The feature fusion unit 101 is used to perform feature fusion on paragraph semantic encoding results, paragraph position encoding results, answer semantic encoding results and answer position encoding results through the feature fusion network of the question generation model to obtain fused features;

[0162] Decoding unit 102 is used to decode the fused features through the decoding network of the question generation model to obtain the target prediction question.

[0163] In one embodiment, Figure 8 The feature fusion unit 101 in the middle includes:

[0164] The encoding splicing subunit is used to splice the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result to obtain the spliced ​​encoding result;

[0165] The feature generation subunit is used to perform feature fusion on the splicing encoding results through the feature fusion network of the problem generation model to obtain fused features.

[0166] In one embodiment, such as Figure 9 As shown, Figure 7 The label determination module 11 in the middle includes:

[0167] The overlap determination unit 110 is used to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph through a reading comprehension model, and to determine the overlap between the target position of the answer corresponding to the target prediction question and the target answer.

[0168] The quality label determination unit 111 is used to determine the quality label of the target prediction problem based on the target location overlap.

[0169] In one embodiment, such as Figure 10 As shown, Figure 7 The problem generation device 1 in the middle also includes:

[0170] The sample question determination module 13 is used to input the sample paragraph and the sample answer corresponding to the sample paragraph into the question generation model to obtain the sample prediction question;

[0171] The similarity determination module 14 is used to determine the question similarity between the sample prediction question and the sample standard question corresponding to the sample answer based on the sample prediction question and the sample standard question.

[0172] The overlap determination module 15 is used to determine the overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer based on the sample prediction question, the sample paragraph, and the position of the sample answer in the sample paragraph.

[0173] The loss value determination module 16 is used to determine the target loss value based on the problem similarity and the overlap of sample positions;

[0174] The model parameter update module 17 is used to update the model parameters of the problem-generated model based on the target loss value.

[0175] In one embodiment, Figure 10 The similarity determination module 14 is also used to input the sample prediction question and the sample standard question corresponding to the sample answer into the paraphrase recognition model to obtain the question similarity between the sample prediction question and the sample standard question.

[0176] In one embodiment, Figure 10The overlap determination module 15 is also used to input the sample prediction question, sample paragraph, and the position of the sample answer in the sample paragraph into the reading comprehension model to obtain the overlap of sample position between the answer corresponding to the sample prediction question and the sample answer.

[0177] In one embodiment, such as Figure 11 As shown, Figure 10 The loss value determination module 16 includes:

[0178] The first loss value determination unit 160 is used to determine the first loss value based on the problem similarity and the sample location overlap.

[0179] The second loss value determination unit 161 is used to determine the second loss value based on the first loss value and the restatement loss value and restatement loss weight of the restatement recognition model;

[0180] The third loss value determination unit 162 is used to determine the third loss value based on the first loss value and the comprehension loss value and comprehension loss weight of the reading comprehension model.

[0181] The target loss value determination unit 163 is used to determine the target loss value based on the second loss value and the third loss value.

[0182] Each module in the aforementioned problem generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0183] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 12As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a problem-generating method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0184] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0185] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0186] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0187] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0188] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0189] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0190] The target paragraph and the target answer corresponding to the target paragraph are encoded through the coding network of the question generation model, so as to obtain the paragraph semantic coding result and paragraph position coding result corresponding to the target paragraph, as well as the answer semantic coding result and answer position coding result corresponding to the target answer;

[0191] By using the feature fusion network of the question generation model, the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are fused to obtain the fused features;

[0192] The target prediction question is obtained by decoding the fused features through the decoding network of the question generation model.

[0193] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0194] The paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are concatenated to obtain the concatenated encoding result;

[0195] By using the feature fusion network of the question generation model, the splicing encoding results are fused to obtain fused features.

[0196] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0197] The reading comprehension model analyzes the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the degree of overlap between the answer to the target prediction question and the target answer.

[0198] The quality label for the target prediction problem is determined based on the degree of overlap between target locations.

[0199] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0200] Input the sample paragraph and the corresponding sample answer into the question generation model to obtain the sample prediction question;

[0201] Based on the sample prediction question and the sample standard question corresponding to the sample answer, determine the question similarity between the sample prediction question and the sample standard question;

[0202] Based on the sample prediction question, the sample paragraph, and the position of the sample answer within the sample paragraph, determine the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer.

[0203] The target loss value is determined based on the problem similarity and sample location overlap.

[0204] Update the model parameters of the problem generation model based on the target loss value.

[0205] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0206] Input the sample prediction question and the corresponding standard question of the sample answer into the paraphrase recognition model to obtain the question similarity between the sample prediction question and the standard question.

[0207] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0208] The sample prediction question, sample paragraph, and the position of the sample answer in the sample paragraph are input into the reading comprehension model to obtain the degree of overlap between the sample position of the answer corresponding to the sample prediction question and the sample answer.

[0209] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0210] The first loss value is determined based on the similarity of the questions and the overlap of the sample locations;

[0211] The second loss value is determined based on the first loss value, the restatement loss value and the restatement loss weight of the restatement recognition model;

[0212] The third loss value is determined based on the first loss value, the comprehension loss value of the reading comprehension model, and the comprehension loss weight.

[0213] The target loss value is determined based on the second and third loss values.

[0214] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0215] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0216] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0217] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0218] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0219] By using a question generation model, the target paragraph and its corresponding target answer are analyzed to obtain the target prediction question;

[0220] By using a reading comprehension model, the target prediction question, the target paragraph, and the position of the target answer in the target paragraph are analyzed to determine the quality label of the target prediction question;

[0221] Add quality labels to the target prediction problem and output the target prediction problem with added quality labels.

[0222] It should be noted that the data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0223] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0224] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0225] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A question generation method characterized by comprising: The method includes: The target paragraph and its corresponding target answer are parsed using a question generation model to obtain a target prediction question. This includes: encoding the target paragraph and its corresponding target answer using the encoding network of the question generation model to obtain the paragraph semantic encoding result and paragraph position encoding result corresponding to the target paragraph, as well as the answer semantic encoding result and answer position encoding result corresponding to the target answer; fusing the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result using the feature fusion network of the question generation model to obtain fused features; and decoding the fused features using the decoding network of the question generation model to obtain the target prediction question. The paragraph semantic encoding result is a multi-dimensional encoding result obtained by semantic parsing each character in the target paragraph; the answer semantic encoding result is a multi-dimensional encoding result obtained by semantic parsing each character in the target answer; the paragraph position encoding result is a relative position encoding result of each character in the target paragraph obtained by sinusoidal position encoding processing based on the paragraph semantic encoding result; and the answer position encoding result is a relative position encoding result of each character in the target answer obtained by sinusoidal position encoding processing based on the answer semantic encoding result. The reading comprehension model is used to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph to determine the quality label of the target prediction question. Add the quality label to the target prediction problem and output the target prediction problem with the added quality label.

2. The method of claim 1, wherein, The feature fusion network of the question generation model fuses the paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result to obtain fused features, including: The paragraph semantic encoding result, paragraph position encoding result, answer semantic encoding result, and answer position encoding result are concatenated to obtain the concatenated encoding result; The concatenated encoding result is fused using a feature fusion network based on the problem generation model to obtain fused features.

3. The method of claim 1, wherein, The step of using a reading comprehension model to analyze the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the quality label of the target prediction question includes: The reading comprehension model analyzes the target prediction question, the target paragraph, and the position of the target answer within the target paragraph to determine the overlap between the target position of the answer to the target prediction question and the target answer. The quality label for the target prediction problem is determined based on the overlap of the target locations.

4. The method according to any one of claims 1 to 3, characterized in that, The problem generation model is trained in the following manner: The sample paragraph and the corresponding sample answer are input into the question generation model to obtain the sample prediction question; determine a question similarity between the sample prediction question and a sample standard question corresponding to the sample prediction question and the sample answer; determine a sample position coincidence degree between an answer corresponding to the sample prediction question and the sample answer according to the sample prediction question, the sample passage, and a position of the sample answer in the sample passage; determine a target loss value according to the question similarity and the sample position coincidence degree; update a model parameter of the question generation model according to the target loss value.

5. The method of claim 4, wherein, The determining the question similarity between the sample prediction question and the sample standard question corresponding to the sample prediction question and the sample answer includes: inputting the sample prediction question, the sample standard question corresponding to the sample prediction question and the sample answer into a paraphrase recognition model to obtain the question similarity between the sample prediction question and the sample standard question.

6. The method of claim 5, wherein, The determining the sample position coincidence degree between the answer corresponding to the sample prediction question and the sample answer according to the sample prediction question, the sample passage, and the position of the sample answer in the sample passage includes: inputting the sample prediction question, the sample passage, and the position of the sample answer in the sample passage into the reading comprehension model to obtain the sample position coincidence degree between the answer corresponding to the sample prediction question and the sample answer.

7. The method of claim 6, wherein, The determining the target loss value according to the question similarity and the sample position coincidence degree includes: determining a first loss value according to the question similarity and the sample position coincidence degree; determining a second loss value according to the first loss value, a paraphrase loss value and a paraphrase loss weight of the paraphrase recognition model; determining a third loss value according to the first loss value, a comprehension loss value and a comprehension loss weight of the reading comprehension model; determining the target loss value according to the second loss value and the third loss value.

8. A question generation apparatus characterized by comprising: The apparatus includes: a question generation module configured to analyze a target passage and a target answer corresponding to the target passage by using a question generation model to obtain a target prediction question; the question generation module includes: an encoding unit configured to encode the target passage and the target answer corresponding to the target passage by using an encoding network of the question generation model to obtain a passage semantic encoding result and a passage position encoding result corresponding to the target passage, and an answer semantic encoding result and an answer position encoding result corresponding to the target answer; a feature fusion unit configured to fuse features of the passage semantic encoding result, the passage position encoding result, the answer semantic encoding result, and the answer position encoding result by using a feature fusion network of the question generation model to obtain fused features; and a decoding unit configured to decode the fused features by using a decoding network of the question generation model to obtain the target prediction question. The paragraph semantic encoding result is a multi-dimensional encoding result obtained by performing semantic analysis on each word in the target paragraph. The answer semantic encoding result is a multi-dimensional encoding result obtained by performing semantic analysis on each word in the target answer. The paragraph position encoding result is a relative position encoding result of each word in the target paragraph obtained by performing sine position encoding processing based on the paragraph semantic encoding result. The answer position encoding result is a relative position encoding result of each word in the target answer obtained by performing sine position encoding processing based on the answer semantic encoding result. The label determination module is configured to analyze the target prediction question, the target paragraph, and the position of the target answer in the target paragraph by using a reading comprehension model, and determine a quality label of the target prediction question. The question output module is configured to add the quality label to the target prediction question, and output the target prediction question with the quality label. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

11. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images