Test question explanation generation method and related device, electronic device, and storage medium

Abstract

The present application discloses a method for generating explanations of test questions and related devices, electronic devices, and storage media, wherein the method for generating explanations of test questions includes: obtaining the text of the test question to be processed, and the text of the test question to be processed includes a stem text and a plurality of answer texts, and then predicting based on the stem text and the plurality of answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, explanation data of the text of the test question to be processed is generated. The above scheme can generate more detailed and more targeted explanation data.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a test question explanation generation method and related devices, electronic equipment, and storage media. Background Art

[0002] In the learning process, test practice is often one of the essential and important links. Through test practice, you can test whether you have fully understood the knowledge and help deepen your understanding of the knowledge.

[0003] In real-life scenarios, it is inevitable to encounter difficult questions during test practice. According to general learning theories and experience, timely explanations after encountering difficult questions are usually more beneficial to improving the level of knowledge acquisition. However, compared to the school environment, the available explanation content in the self-study scenario is also relatively limited, and usually only the standard answers to the questions can be used as a reference. However, the standard answers do not answer the questions well in many cases. In view of this, how to generate more detailed and targeted test explanations has become an urgent problem to be solved. Summary of the invention

[0004] The main technical problem solved by this application is to provide a test question explanation generation method and related devices, electronic equipment, and storage media, which can generate more detailed and targeted test question explanations.

[0005] In order to solve the above-mentioned technical problems, the first aspect of the present application provides a method for generating test question explanations, including: obtaining the text of the test question to be processed; wherein the text of the test question to be processed includes a stem text and a plurality of answer texts; making predictions based on the stem text and the plurality of answer texts to obtain the target examination points involved in the text of the test question to be processed and the importance of each target examination point; and generating explanation data for the text of the test question to be processed based on the importance of the target examination points and the explanation data.

[0006] In order to solve the above-mentioned technical problems, the second aspect of the present application provides a test question explanation generation device, including: an acquisition module, a prediction module and a generation module, the acquisition module is used to obtain the text of the test question to be processed; wherein the text of the test question to be processed includes a stem text and a plurality of answer texts; the prediction module is used to make predictions based on the stem text and the plurality of answer texts to obtain the target examination points involved in the text of the test question to be processed and the importance of each target examination point; the generation module is used to generate explanation data of the text of the test question to be processed based on the importance of the target examination points and the explanation data.

[0007] In order to solve the above technical problems, the third aspect of the present application provides an electronic device, including a memory and a processor coupled to each other, the memory storing program instructions, and the processor being used to execute the program instructions to implement the test question explanation generation method of the above first aspect.

[0008] In order to solve the above technical problems, the fourth aspect of the present application provides a computer-readable storage medium storing program instructions that can be executed by a processor, and the program instructions are used to implement the test question explanation generation method of the above first aspect.

[0009] The above scheme obtains the text of the test question to be processed, and the text of the test question to be processed includes the text of the question stem and several answer texts, and then predicts based on the text of the question stem and several answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, the explanation data of the text of the test question to be processed is generated. Therefore, the explanation data of the text of the test question to be processed is generated based on the importance of the target inspection point and the explanation data, so that the explanation data finally generated can take into account as many and important target inspection points as possible. Therefore, more detailed and more targeted explanation data can be generated. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of an embodiment of a method for generating test question explanations of the present application;

[0011] Figure 2 is a flow chart of an embodiment of importance prediction;

[0012] Figure 3 It is a schematic diagram of a framework of an embodiment of an importance prediction model;

[0013] Figure 4 This is a flow chart of an embodiment of inspection point prediction;

[0014] Figure 5 is a schematic diagram of an embodiment of a second sample test text;

[0015] Figure 6 It is a schematic diagram of a framework of an embodiment of an inspection point prediction model;

[0016] Figure 7 It is a schematic diagram of the framework of an embodiment of the test question explanation generating device of the present application;

[0017] Figure 8 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application;

[0018] Fig. 9 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0019] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0020] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0021] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0022] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the method for generating test question explanations of this application. Specifically, it may include the following steps:

[0023] Step S11: Obtain the text of the test question to be processed.

[0024] In the disclosed embodiment, the test question text to be processed includes a question stem text and a plurality of answer texts. It should be noted that the plurality of answer texts included in the test question text to be processed can constitute the standard answer of the test question text to be processed. In addition, the specific number of the plurality of answer texts can be 1, 2, 3, 4, etc., which is not limited here.

[0025] In an implementation scenario, the text of the test question to be processed can usually include two parts: the question stem text and the standard answer. On this basis, the standard answer can be split into a step sequence consisting of multiple steps. For example, a step is a sentence of the standard answer, or the step granularity can be defined according to the characteristics of different subjects, which is not limited here. After obtaining the step sequence, each step in the step sequence can be used as the answer text.

[0026] In an implementation scenario, the test text to be processed may involve several target examination points, and the specific number of the target examination points may be 1, 2, 3, 4, etc., which is not limited here. Taking the subject of the test text to be processed as mathematics as an example, the target examination point involved may be trigonometric functions, or it may also involve two target examination points of solid geometry and trigonometric functions. Other situations can be deduced by analogy, and no examples are given here one by one.

[0027] Step S12: Predictions are made based on the question stem text and several answer texts to obtain the target examination points involved in the question text to be processed and the importance of each target examination point.

[0028] Specifically, predictions can be made based on the question text and several answer texts to obtain the target inspection points involved in the answer text. It should be noted that in real scenarios, each answer text may involve inspection points. In this case, the target inspection points involved in each answer text can be obtained through prediction; of course, only part of the answer text may involve inspection points. In this case, the target inspection points involved in this part of the answer text can be obtained through prediction. At the same time, predictions can be made based on the question text and several answer texts to obtain the importance corresponding to the answer text, and the importance corresponding to the answer text indicates the importance of the target inspection point involved in the answer text when the answer text involves the target inspection point. On this basis, the answer text involving the target inspection point can be used as the target text, and the importance corresponding to the target text can be used as the importance of the target inspection point involved in the target text. For example, the test text to be processed contains three answer texts, which can be named answer text 1, answer text 2 and answer text 3 respectively. After prediction, it can be obtained that answer text 1 involves the target inspection point "solid geometry", answer text 2 does not involve the target inspection point, and answer text 3 involves the target inspection point "trigonometric function"; at the same time, it can be predicted that the importance corresponding to answer text 1 is 0.8, the importance corresponding to answer text 2 is 0.3, and the importance corresponding to answer text 3 is 0.9. On this basis, answer text 1 and answer text 3 involving the target inspection point can be used as target texts respectively, and the importance of 0.8 corresponding to target text 1 (i.e. answer text 1) is used as the importance of the target inspection point "solid geometry" involved in target text 1 (i.e. answer text 1), and the importance of 0.9 corresponding to target text 2 (i.e. answer text 3) is used as the importance of the target inspection point "trigonometric function" involved in target text 2 (i.e. answer text 3). Other situations can be deduced by analogy, and no examples are given here one by one. It should be noted that the above inspection point prediction and importance prediction can be performed in a certain order, such as the inspection point prediction can be performed first, and then the importance prediction, or the importance prediction can be performed first, and then the inspection point prediction. Of course, the above inspection point prediction and importance prediction can also be performed at the same time, which is not limited here.

[0029] In one implementation scenario, in order to obtain the importance, the correlation calculation can be performed based on the first feature representation of the question text and the second feature representation of each answer text to obtain the first correlation of each answer text with respect to the importance prediction, and then the second feature representation of each answer text and the first correlation are weighted to obtain the first weighted representation of each answer text. On this basis, predictions can be made based on the first weighted representation of each answer text to obtain the importance of the target inspection point involved in each answer text. The specific process of feature interaction and importance prediction can be referred to the following public embodiments, which will not be described here. The above method, through the first feature representation of the question text and the second feature representation of each answer text, the first correlation of each answer text with respect to the importance prediction is obtained, and then the second feature representation is weighted by the first correlation, so that the importance prediction is made based on the weighted feature representation, which can improve the accuracy of the importance prediction.

[0030] In one implementation scenario, in order to obtain the target inspection point, the correlation calculation can be performed based on the first feature representation of the question text and the second feature representation of each answer text to obtain the second correlation of each answer text with respect to the inspection point prediction, and then the second feature representation and the second correlation of each answer text are weighted to obtain the second weighted representation of each answer text. On this basis, predictions can be made based on the second weighted representation of each answer text to obtain the target inspection point involved in the answer text. The specific process of feature interaction and inspection point prediction can be referred to the following public embodiments, which will not be repeated here. The above method, through the first feature representation of the question text and the second feature representation of each answer text, the second correlation of each answer text with respect to the inspection point prediction is obtained, and then the second feature representation is weighted by the second correlation, so that the inspection point prediction is made based on the weighted feature representation, which can improve the accuracy of the inspection point prediction.

[0031] Step S13: Based on the importance of the target examination point and the explanation data, generate explanation data for the test question text to be processed.

[0032] In an implementation scenario, after obtaining the importance of each target examination point, the target examination points can be sorted in descending order of importance. On this basis, the target examination points ranked in the front preset order can be selected, and based on the explanation data of the selected target examination points, the explanation data of the test question text to be processed can be generated.

[0033] In a specific implementation scenario, explanation data for each inspection point can be pre-constructed, and the explanation data for the inspection point can include but are not limited to: explanation video, explanation audio, explanation diagram, explanation text, etc., which are not limited here. Still taking the inspection point "trigonometric function" as an example, the explanation data for the inspection point "trigonometric function" can include video data for explaining "trigonometric function"; or, the explanation data for the inspection point "trigonometric function" can also include audio data for explaining "trigonometric function"; or, the explanation data for the inspection point "trigonometric function" can also include image data for explaining "trigonometric function", or, the explanation data for the inspection point "trigonometric function" can also include text data for explaining "trigonometric function". Other situations can be deduced by analogy, and examples are not given one by one here.

[0034] In a specific implementation scenario, the preset rank can be set according to actual conditions. For example, the preset rank can be set to a numerical value, such as 2, 3, 4, etc., which is not limited here; or the preset rank amount can be set to a percentage, such as 20%, 30%, 40%, etc., which is not limited here.

[0035] In a specific implementation scenario, after selecting the target inspection points according to the importance of each target inspection point, the explanation data of the selected target inspection points can be comprehensively selected to obtain the explanation data of the test question text to be processed. For example, the set of explanation data of the selected target inspection points can be directly used as the explanation data of the test question text to be processed.

[0036] In one implementation scenario, in order to facilitate the user to view the explanation data, the explanation data of the target inspection point involved in the selected answer text can be displayed in response to the user's selection instruction for the answer text. For example, if the explanation data of the selected target inspection point is video data, the video data can be played; or if the explanation data of the selected target inspection point is audio data, the audio data can be played; or if the explanation data of the selected target inspection point is image data, the image data can be displayed; or if the explanation data of the selected target inspection point is text data, the text data can be displayed.

[0037] In an implementation scenario, the importance of the target inspection point can be predicted based on the importance prediction model, and the importance prediction model is trained based on the first sample test text. The training process of the model can refer to the following public embodiments, which will not be described here. On this basis, the number of selections of each answer text in the test text to be processed can be counted, and based on the number of selections, the importance of the target inspection point involved in each answer text in the test text to be processed can be updated, so that the test text to be processed marked with the updated importance can be used as the new first sample test text, and then the importance prediction model can be retrained based on the new first sample test text. The above method continuously optimizes the importance prediction model and improves the accuracy of importance prediction by collecting the number of selections of each answer text in the real scene.

[0038] In a specific implementation scenario, the greater the number of times the answer text is selected, the higher the importance of the target examination point involved in the answer text can be considered. Conversely, the smaller the number of times the answer text is selected, the lower the importance of the target examination point involved in the answer text can be considered.

[0039] In a specific implementation scenario, after accumulating a certain number of new first sample test question texts, the importance prediction model can be retrained based on the new first sample test question texts, so that during the use of the importance prediction model, it can be continuously adjusted and updated so that its importance prediction results continue to approach the user's actual situation.

[0040] In a specific implementation scenario, after retraining the importance prediction model, the retrained importance prediction model can be used to perform importance prediction and re-execute the above-mentioned step of "generating explanation data for the test text to be processed based on the importance of the target examination point and the explanation data", so that the finally generated explanation data can be as consistent as possible in terms of coverage of examination points and problems encountered by users during actual use.

[0041] The above scheme obtains the text of the test question to be processed, and the text of the test question to be processed includes the text of the question stem and several answer texts, and then predicts based on the text of the question stem and several answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, the explanation data of the text of the test question to be processed is generated. Therefore, the explanation data of the text of the test question to be processed is generated based on the importance of the target inspection point and the explanation data, so that the explanation data finally generated can take into account as many and important target inspection points as possible. Therefore, more detailed and more targeted explanation data can be generated.

[0042] See also Figure 2 , Figure 2This is a flow chart of an embodiment of importance prediction. Specifically, it may include the following steps:

[0043] Step S21: Calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text to obtain the first relevance of each answer text with respect to importance prediction.

[0044] In an implementation scenario, for the question text or the answer text, the embedding representation of each character in the text can be obtained first, and then the embedding representations of each character in the text are merged (such as addition, concatenation, etc.) to obtain the feature representation of the text. It should be noted that when the text is the question text, the feature representation of the text is the first feature representation, and when the text is the answer text, the feature representation of the text is recorded as the second feature representation. In addition, the specific process of obtaining the embedded representation can refer to the technical details of vectorization tools such as word2vec, which will not be repeated here.

[0045] In an implementation scenario, the first feature representation of the question text and the second feature representation of each answer text can be interactively represented based on the attention mechanism to obtain the first relevance of each answer text with respect to importance prediction. It should be noted that the higher the first relevance of the answer text, the stronger the relevance of the answer text for importance prediction. Conversely, the lower the first relevance of the answer text, the weaker the relevance of the answer text for importance prediction. In addition, for the specific process of feature interaction through the attention mechanism, please refer to the relevant technical details of the attention mechanism, which will not be repeated here.

[0046] Step S22: performing weighting based on the second feature representation and the first relevance of each answer text to obtain a first weighted representation of each answer text.

[0047] Specifically, the second feature representation of the answer text may be weighted using the first relevance of the answer text to obtain a first weighted representation of the answer text.

[0048] Step S23: Predict the importance of the target inspection point involved in the answer text based on the first weighted representation of each answer text.

[0049] Specifically, in order to improve the efficiency of importance prediction, an importance prediction model can be pre-trained, and as described in the aforementioned disclosed embodiment, the importance prediction model can be trained based on the first sample test text, the first sample test text can include a first sample question text and a plurality of first sample answer texts, the first sample answer text is annotated with sample importance, and the sample importance indicates the criticality of the sample inspection point involved in the first sample answer text. It should be noted that the higher the criticality of the sample inspection point involved in the first sample answer text, the greater the sample importance annotated by the first sample answer text, and conversely, the lower the criticality of the sample inspection point involved in the first sample answer text, the smaller the sample importance annotated by the first sample answer text. In addition, in the case where the first sample answer text does not involve a sample inspection point, the sample importance annotated by the first sample answer text can be a preset importance, and the preset importance is lower than the sample importance annotated by any first sample answer text involving a sample inspection point.

[0050] In an implementation scenario, the sample importance can be obtained based on the analysis of the answer texts of several subjects to the first sample test text. It should be noted that the specific number of several subjects can be set according to the actual situation, such as 10, 50, 100, etc., which is not limited here. Specifically, the first sample answer texts involving the sample inspection points can be used as the current sample answer texts, and the sample inspection points involved in the current answer texts can be used as the current inspection points, and the miss rates of several subjects at the current inspection points can be obtained based on the semantic similarity between the current sample answer texts and the answer texts of each subject. On this basis, the sample importance of each first sample answer text is obtained based on the miss rates corresponding to the sample inspection points involved in the first sample answer text in the first sample test text. In the above method, by counting the miss rates of several subjects at the current inspection points, and by the miss rates of each sample inspection point, the sample importance of each sample inspection point is obtained, so that the importance of the sample inspection point can be measured from the perspective of the miss rates of the subjects to the sample inspection points, and the accuracy of the sample importance can be improved.

[0051] In a specific implementation scenario, take the example that the first sample test text contains three first sample answer texts, which are named first sample answer text 1, first sample answer text 2 and first sample answer text 3 respectively, and the sample inspection point involved in the first sample answer text 1 is "solid geometry", the first sample answer text 2 does not involve a sample inspection point, and the sample inspection point involved in the first sample answer text 3 is "trigonometric function". On this basis, the first sample answer text 1 and the first sample answer text 3 can be used as the current sample answer texts, and the sample inspection point "solid geometry" and the sample inspection point "trigonometric function" can be used as the current inspection point. Other situations can be deduced by analogy, and no examples are given here one by one.

[0052] In a specific implementation scenario, a pre-trained language model such as BERT (Bidirectional Encoder Representation from Transformers) can be used to perform semantic extraction on the current sample answer text and the answer text, respectively, to obtain the semantic representation of the current sample answer text and the semantic representation of the answer text. It should be noted that in order to improve the accuracy of semantic extraction, the pre-trained language model can be pre-trained in advance with domain data (such as test data, etc.). For the specific process of pre-training, please refer to the technical details of pre-trained language models such as BERT, which will not be repeated here. In addition, the answer text can specifically contain several sub-texts, and the several sub-texts represent different answer steps respectively, so the semantic representations of the several sub-texts can be extracted respectively. On this basis, if the semantic similarity between the semantic representation of the answer text (or a sub-text) and the semantic representation of the current sample answer text is higher than a preset threshold (e.g., 0.8, 0.9, etc.), it can be considered that the subject to whom the answer text (or a sub-text) belongs has hit the current test point. Otherwise, it can be considered that the subject to whom the answer text (or a sub-text) belongs has not hit the current test point. Exemplarily, for the first sample test question i involving sample test point j, the answer texts of n subjects can be collected. If it is determined through the above analysis that m subjects have not hit the sample test point j, it can be determined that the miss rate of the sample test point j involved in the first sample test question text i is The same can be said for other situations, which will not be listed here one by one.

[0053] In a specific implementation scenario, after obtaining the miss rates of several subjects at the current inspection point, the miss rates corresponding to the sample inspection points involved in the current sample test text can be normalized, such as using a softmax function for normalization, to obtain the sample importance corresponding to each sample inspection point. Still taking the first sample test text containing 3 first sample answer texts as an example, after the above statistical operation, it can be obtained that the miss rate of the sample inspection point "solid geometry" is 0.7, and the miss rate of the sample inspection point "trigonometric function" is 0.8. Then the softmax function can be used to normalize the above two miss rates, and the sample importance of the sample inspection point "solid geometry" can be obtained as 0.48, and the sample importance of the sample inspection point "trigonometric function" is 0.52. Other situations can be deduced by analogy, and no examples are given here one by one.

[0054] In an implementation scenario, after obtaining the first sample test question text marked with sample importance, the first sample question stem text and several first sample answer texts can be predicted based on the importance prediction model to obtain the predicted importance of each first sample answer text, and then based on the difference between the sample importance and the predicted importance of the first sample answer text involving the sample inspection point, the network parameters of the importance prediction model are adjusted. In other words, in the difference measurement process, the importance prediction error of the first sample answer text that does not involve the sample inspection point can be shielded. It should be noted that the specific measurement method of the difference can refer to the technical details of the loss function such as the mean square error loss, and the specific adjustment process of the parameters can refer to the technical details of the optimization method such as gradient descent, which will not be repeated here. In the above method, the first sample question text and several first sample answer texts are predicted through the importance prediction model to obtain the predicted importance of each first sample answer text, and based on the difference between the sample importance and the predicted importance of the first sample answer text involving the sample inspection point, the network parameters of the importance prediction model are adjusted. That is, during the training process, the importance prediction error of the first sample answer text that does not involve the sample inspection point is shielded, thereby reducing the impact of noise on model training as much as possible, which is conducive to improving the prediction accuracy of the importance prediction model.

[0055] In an implementation scenario, after the importance prediction model is trained, the trained importance prediction model can be used to process the question text and several answer texts to obtain the importance. Figure 3 , Figure 3 Schematic diagram of the framework of an embodiment of the importance prediction model. Figure 3As shown, the importance prediction model may include a first attention network, a first semantic extraction network and a first prediction network. The first attention network is specifically used to calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text, obtain the first relevance of each answer text regarding the importance prediction, and weight each answer text based on the second feature representation and the first relevance to obtain the first weighted representation of each answer text. In addition, the first semantic extraction network is specifically used to perform semantic extraction on the first weighted representation of each answer text to obtain the first semantic representation of each answer text. The first prediction network is specifically used to predict the importance based on the first semantic representation of each answer text. It should be noted that the first semantic extraction network may include but is not limited to: Bi-LSTM (Bi-directional Long Short-Term Memory, bidirectional long short-term memory network), etc., which is not limited here. In addition, the first prediction network may include but is not limited to: DNN (Deep Neural Networks, deep neural network), etc., which is not limited here. The above method, by setting the importance prediction network to include a first attention network, a first semantic extraction network and a first prediction network, and the first attention network is used to perform feature interaction, the first semantic extraction network is used for semantic extraction, and the first prediction network is used for importance prediction, can help greatly improve the efficiency of importance prediction.

[0056] The above scheme performs feature interaction through the first feature representation of the question text and the second feature representation of each answer text to obtain the first relevance of each answer text regarding importance prediction, and then weights the second feature representation by the first relevance, so as to perform importance prediction based on the weighted feature representation, which can improve the accuracy of importance prediction.

[0057] See also Figure 4 , Figure 4 This is a flow chart of an embodiment of observation point prediction. Specifically, it may include the following steps:

[0058] Step S41: Calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text to obtain the second relevance of each answer text with respect to the prediction of the examination point.

[0059] In one implementation scenario, the method of obtaining the first feature representation of the question text and the second feature representation of the answer text can be found in the relevant description in the aforementioned public embodiment, which will not be repeated here.

[0060] In an implementation scenario, similar to obtaining the first relevance in the aforementioned public embodiment, the first feature representation of the question text and the second feature representation of each answer text can also be subjected to feature interaction based on the attention mechanism to obtain the second relevance of each answer text regarding importance prediction. It should be noted that the higher the second relevance of the answer text, the stronger the relevance of the answer text for the prediction of the examination point. Conversely, the lower the second relevance of the answer text, the weaker the relevance of the answer text for the prediction of the examination point. In addition, for the specific process of feature interaction through the attention mechanism, please refer to the relevant technical details of the attention mechanism, which will not be repeated here.

[0061] Step S42: performing weighting based on the second feature representation and the second relevance of each answer text to obtain a second weighted representation of each answer text.

[0062] Specifically, the second feature representation of the answer text may be weighted using the second relevance of the answer text to obtain a second weighted representation of the answer text.

[0063] Step S43: Predict the target inspection point involved in the answer text based on the second weighted representation of each answer text.

[0064] In one implementation scenario, a binary classification prediction can be performed based on the second weighted representation of each answer text to obtain a first prediction result for each answer text, and a multi-classification prediction can be performed based on the second weighted representation of each answer text to obtain a second prediction result for each answer text, and the first prediction result includes the possibility that the answer text involves an inspection point, and the second prediction result includes the candidate inspection points involved in the answer text. On this basis, in response to the first prediction result of the answer text satisfying a preset condition, the target inspection point involved in the answer text can be obtained based on the candidate inspection points involved in the answer text. The above method determines the target inspection point involved in the answer text by combining the binary classification prediction and the multi-classification prediction of the inspection point, thereby avoiding as much as possible the prediction error that may be caused by determining the target inspection point only through multi-classification prediction, and thus can be conducive to improving the robustness of the inspection point prediction.

[0065] In a specific implementation scenario, the first prediction result may specifically include a probability value of the answer text involving the inspection point. It should be noted that the larger the probability value, the higher the possibility that the answer text involves the inspection point, and conversely, the smaller the probability value, the lower the possibility that the answer text involves the inspection point.

[0066] In a specific implementation scenario, as described above, the first prediction result may include a probability value that the answer text involves an inspection point, and the preset condition may be set to include: the probability value is greater than a preset threshold. That is, when the probability value of the answer text being predicted to involve an inspection point is greater than the preset threshold, it can be considered that the answer text involves an inspection point, and when the second prediction result includes multiple candidate inspection points, the candidate inspection point with the highest confidence can be used as the target inspection point involved in the answer text.

[0067] In an implementation scenario, in order to improve the efficiency of test point prediction, a test point prediction model can be pre-trained, and the target test point can be predicted based on the test point prediction model. The test point prediction model can be specifically trained based on the second sample test question text, the second sample test question text includes a second sample question stem text and a plurality of second sample answer texts, and the second sample answer text can be annotated with a sample tag, and the sample tag indicates whether the second sample answer text actually involves a sample test point and the sample test point actually involved in the second sample answer text. Please refer to Figure 5 , Figure 5 Schematic diagram of an embodiment of the second sample test text. Figure 5 As shown, the second sample test text contains N second sample answer texts, which are respectively recorded as second sample answer text 1, second sample answer text 2, ..., second sample answer text N, and the sample mark of second sample answer text 1 is "examination point 1", indicating that the second sample answer text 1 actually involves sample examination points and the sample examination points involved are "examination point 1". Similarly, the sample mark of second sample answer text N is "examination point 2", indicating that the second sample answer text 2 actually involves sample examination points and the sample examination points involved are "examination point 2". In addition, the sample mark of second sample answer text 2 is empty, indicating that the second sample answer text 2 does not actually involve sample examination points. Of course, except for Figure 5 The marking method shown may also be other marking methods. For example, the number "1" or "0" may be used to indicate whether the second sample answer text actually involves a sample examination point, which is not limited here. It should be noted that the second sample test question text may be the same sample test question text as the first sample test question text. Of course, the second sample test question text may also be a different sample test question text from the first sample test question text, which is not limited here. The above method, by predicting the target examination point through the examination point prediction model, can help improve the prediction efficiency of the target examination point, and by machine identification of the target examination point, it can also avoid the problem of low human consistency in the process of manual identification of the target examination point, thereby improving the consistency of examination point identification.

[0068] In an implementation scenario, after obtaining the second sample test text marked with a sample mark, the second sample question stem text and several second sample answer texts can be predicted based on the test point prediction model to obtain the prediction mark of each second sample answer text, and the prediction mark indicates whether the second sample answer text is predicted to involve the sample test point and the sample test point involved in the prediction of the second sample answer text. On this basis, the network parameters of the test point prediction model can be adjusted based on the difference between the sample mark and the prediction mark of the second sample answer text actually involving the sample test point. In other words, in the difference measurement process, the test point prediction error of the second sample answer text that does not involve the sample test point can be shielded. It should be noted that the specific measurement method of the difference can refer to the technical details of the loss function such as cross entropy loss, and the specific adjustment process of the parameters can refer to the technical details of the optimization method such as gradient descent, which will not be repeated here. In the above method, the second sample question text and several second sample answer texts are predicted through the examination point prediction model to obtain the predicted examination points of each second sample answer text, and based on the difference between the sample mark and the predicted mark of the second sample answer text involving the sample examination point, the network parameters of the examination point prediction model are adjusted. That is, during the training process, the examination point prediction errors of the second sample answer text not involving the sample examination point are shielded, thereby reducing the impact of noise on model training as much as possible, which is conducive to improving the prediction accuracy of the examination point prediction model.

[0069] In a specific implementation scenario, as described above, the sample tag indicates whether the second sample answer text actually involves the sample inspection point and the sample inspection point actually involved by the second sample answer text, and the prediction tag indicates whether the second sample answer text is predicted to involve the sample inspection point and the sample inspection point predicted to involve by the second sample answer text. On this basis, the difference between the first sample sub-tag and the first prediction sub-tag can be measured based on the binary cross entropy loss to obtain the first loss, and the first sample sub-tag is derived from the sample tag, indicating whether the second sample answer text actually involves the sample inspection point (e.g., 0 indicates that the sample inspection point is not involved, and 1 indicates that the sample inspection point is involved), and the first prediction sub-tag is derived from the prediction tag, indicating whether the second sample answer text is predicted to involve the sample inspection point; at the same time, the difference between the second sample sub-tag and the second prediction sub-tag can be measured based on the multi-classification cross entropy loss to obtain the second loss, and the second sample sub-tag is derived from the sample tag, indicating the sample inspection point actually involved by the second sample answer text, and the second prediction sub-tag is derived from the prediction tag, indicating the sample inspection point predicted to involve by the second sample answer text. Based on this, the first loss and the second loss can be fused (e.g., weighted, added, etc.) to obtain a fused loss, and the network parameters of the observation point prediction model can be adjusted based on the loss fusion.

[0070] In a specific implementation scenario, the inspection point prediction model may include a binary classification network and a multi-classification network, and the binary classification network is used to perform the aforementioned binary classification to obtain the aforementioned first prediction label, and the multi-classification network is used to perform the aforementioned multi-classification to obtain the aforementioned second prediction label. The specific network structure of the inspection point prediction model can be referred to the following related description, which will not be repeated here.

[0071] In an implementation scenario, please refer to Figure 6 , Figure 6 FIG. 1 is a schematic diagram of a framework of an embodiment of an inspection point prediction model. Figure 6 As shown, the test point prediction model may include a second attention network, a second semantic extraction network, and a second prediction network. The second attention network is specifically used to calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text, obtain the second relevance of each answer text about the test point prediction, and weight the second feature representation and the second relevance of each answer text to obtain the second weighted representation of each answer text. In addition, the second semantic extraction network is specifically used to perform semantic extraction on the second weighted representation of each answer text to obtain the second semantic representation of each answer text. The second prediction network is specifically used to predict based on the second semantic representation of each answer text to obtain the target test point involved in the answer text. Further, the second prediction network may specifically include a binary classification network and a multi-classification network. The binary classification network is used to perform binary classification based on the second semantic representation of each answer text to obtain the aforementioned first prediction result, and the multi-classification network is used to perform multi-classification based on the second semantic representation of each answer text to obtain the aforementioned second prediction result. It should be noted that the second semantic extraction network may include but is not limited to: Bi-LSTM (Bi-directional Long Short-Term Memory, bidirectional long short-term memory network), etc., which is not limited here. In addition, the binary classification network may include but is not limited to: CRF (Conditional Random Field), etc., and the multi-classification network may include but is not limited to: DNN (Deep Neural Networks), etc., which are not limited here. The above method, by setting the inspection point prediction network to include a second attention network, a second semantic extraction network, and a second prediction network, and the second attention network is used to perform feature interaction, the second semantic extraction network is used for semantic extraction, and the second prediction network is used for inspection point prediction, can help greatly improve the efficiency of inspection point prediction.

[0072] The above scheme performs feature interaction through the first feature representation of the question text and the second feature representation of each answer text to obtain the second relevance of each answer text with respect to the examination point prediction, and then weights the second feature representation by the second relevance, so as to perform examination point prediction based on the weighted feature representation, which can improve the accuracy of examination point prediction.

[0073] See also Figure 7 , Figure 7 It is a schematic diagram of the framework of an embodiment of the test question explanation generating device 70 of the present application. The test question explanation generating device 70 comprises: an acquisition module 71, a prediction module 72 and a generation module 73, wherein the acquisition module 71 is used to acquire the test question text to be processed; wherein the test question text to be processed comprises a question stem text and a plurality of answer texts; the prediction module 72 is used to make predictions based on the question stem text and the plurality of answer texts, and obtain the target examination points involved in the test question text to be processed and the importance of each target examination point; the generation module 73 is used to generate the explanation data of the test question text to be processed based on the importance of the target examination point and the explanation data.

[0074] The above scheme obtains the text of the test question to be processed, and the text of the test question to be processed includes the text of the question stem and several answer texts, and then predicts based on the text of the question stem and several answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, the explanation data of the text of the test question to be processed is generated. Therefore, the explanation data of the text of the test question to be processed is generated based on the importance of the target inspection point and the explanation data, so that the explanation data finally generated can take into account as many and important target inspection points as possible. Therefore, more detailed and more targeted explanation data can be generated.

[0075] In some disclosed embodiments, the prediction module 72 includes an importance prediction submodule, and the importance prediction submodule includes a first interaction unit for performing relevance calculation based on a first feature representation of the question text and a second feature representation of each answer text to obtain a first relevance of each answer text regarding importance prediction; the importance prediction submodule includes a first weighting unit for performing weighting based on the second feature representation and the first relevance of each answer text to obtain a first weighted representation of each answer text; the importance prediction submodule includes a first prediction unit for performing prediction based on the first weighted representation of each answer text to obtain the importance of the target examination point involved in the answer text.

[0076] Therefore, feature interaction is performed through the first feature representation of the question text and the second feature representation of each answer text to obtain the second relevance of each answer text with respect to the examination point prediction, and then the second feature representation is weighted by the second relevance, so that the examination point prediction is performed based on the weighted feature representation, which can improve the accuracy of the examination point prediction.

[0077] In some public embodiments, the importance is predicted based on an importance prediction model, which is trained based on a first sample test question text. The first sample test question text includes a first sample question stem text and several first sample answer texts, and the first sample answer text is annotated with a sample importance, which indicates the criticality of a sample examination point involved in the first sample answer text.

[0078] Therefore, the importance is predicted by the importance prediction model, and the importance prediction model is trained based on the first sample test question text. The first sample test question text includes a first sample question stem text and several first sample answer texts, and the first sample answer text is marked with sample importance. The sample importance indicates the criticality of the sample examination point involved in the first sample answer text, which can help improve the prediction accuracy of the importance prediction model.

[0079] In some disclosed embodiments, the sample importance is obtained based on analysis of the answers of several subjects to the first sample test text.

[0080] Therefore, the sample importance is obtained based on the analysis of the answer texts of several subjects to the first sample test text. Therefore, the sample importance can be accurately labeled through the answer texts of several subjects to the first sample test text, which is conducive to avoiding labeling errors caused by manual intuitive labeling and improving the accuracy of sample importance.

[0081] In some disclosed embodiments, the importance prediction submodule also includes a selection unit for taking the first sample answer texts involving the sample examination points as the current sample answer texts, and taking the sample examination points involved in the current answer texts as the current examination points; the importance prediction submodule also includes a statistical unit for obtaining the miss rates of several subjects at the current examination points based on the semantic similarities between the current sample answer texts and the answer texts of each subject; the importance prediction submodule also includes a measurement unit for obtaining the sample importance of each first sample answer text based on the miss rates corresponding to the sample examination points involved in the first sample answer texts in the first sample test question texts.

[0082] Therefore, by counting the miss rates of several subjects at the current inspection point and the miss rates of each sample inspection point, the sample importance of each sample inspection point can be obtained. This makes it possible to measure the importance of the sample inspection point from the perspective of the subjects' miss rates of the sample inspection points, thereby improving the accuracy of the sample importance.

[0083] In some disclosed embodiments, the importance prediction model includes: a first attention network, used to perform relevance calculation based on a first feature representation of the question text and a second feature representation of each answer text, to obtain a first relevance of each answer text with respect to importance prediction, and to weight each answer text based on the second feature representation and the first relevance, to obtain a first weighted representation of each answer text; a first semantic extraction network, used to perform semantic extraction on the first weighted representation of each answer text, to obtain a first semantic representation of each answer text; and a first prediction network, used to predict the importance based on the first semantic representation of each answer text.

[0084] Therefore, by setting the importance prediction network to include a first attention network, a first semantic extraction network and a first prediction network, and the first attention network is used to perform feature interaction, the first semantic extraction network is used for semantic extraction, and the first prediction network is used for importance prediction, it can help to greatly improve the efficiency of importance prediction.

[0085] In some disclosed embodiments, the importance prediction submodule includes a first model prediction unit, which is used to predict a first sample question text and several first sample answer texts based on an importance prediction model to obtain a predicted importance of each first sample answer text; the importance prediction submodule includes a first parameter adjustment unit, which is used to adjust the network parameters of the importance prediction model based on the difference between the sample importance and the predicted importance of the first sample answer text involving a sample examination point.

[0086] Therefore, the first sample question text and several first sample answer texts are predicted through the importance prediction model to obtain the predicted importance of each first sample answer text, and based on the difference between the sample importance and the predicted importance of the first sample answer text involving the sample examination point, the network parameters of the importance prediction model are adjusted. That is, during the training process, the importance prediction error of the first sample answer text that does not involve the sample examination point is shielded, thereby reducing the impact of noise on model training as much as possible, which is conducive to improving the prediction accuracy of the importance prediction model.

[0087] In some disclosed embodiments, the prediction module 72 includes an examination point prediction submodule, which includes a second interaction unit for performing relevance calculation based on a first feature representation of the question text and a second feature representation of each answer text to obtain a second relevance of each answer text with respect to the examination point prediction; the examination point prediction submodule includes a second weighting unit for performing weighting based on the second feature representation and the second relevance of each answer text to obtain a second weighted representation of each answer text; the examination point prediction submodule includes a second prediction unit for performing prediction based on the second weighted representation of each answer text to obtain a target examination point involved in the answer text.

[0088] Therefore, feature interaction is performed through the first feature representation of the question text and the second feature representation of each answer text to obtain the second relevance of each answer text with respect to the examination point prediction, and then the second feature representation is weighted by the second relevance, so that the examination point prediction is performed based on the weighted feature representation, which can improve the accuracy of the examination point prediction.

[0089] In some disclosed embodiments, the second prediction unit includes a binary classification subunit, which is used to perform binary classification prediction based on the second weighted representation of each answer text to obtain a first prediction result for each answer text, and the second prediction unit includes a multi-classification subunit, which is used to perform multi-classification prediction based on the second weighted representation of each answer text to obtain a second prediction result for each answer text; wherein the first prediction result includes the possibility that the answer text involves an examination point, and the second prediction result includes the candidate examination points involved in the answer text; the second prediction unit includes an examination point determination unit, which is used to obtain a target examination point involved in the answer text based on the candidate examination points involved in the answer text in response to the first prediction result of the answer text satisfying a preset condition.

[0090] Therefore, by combining the binary classification prediction and multi-classification prediction of the examination points, the target examination points involved in the answer text are determined, so as to avoid as much as possible the prediction errors that may be caused by determining the target examination points only through multi-classification prediction, which can help improve the robustness of the examination point prediction.

[0091] In some disclosed embodiments, a target examination point is predicted based on an examination point prediction model, which is trained based on a second sample test question text. The second sample test question text includes a second sample question stem text and several second sample answer texts, and the second sample answer text is annotated with a sample tag, which indicates whether the second sample answer text actually involves a sample examination point and the sample examination point that the second sample answer text actually involves.

[0092] Therefore, predicting the target inspection point through the inspection point prediction model can help improve the prediction efficiency of the target inspection point. And by machine identifying the target inspection point, it can also avoid the problem of low human consistency in the process of manual identification of the target inspection point, and improve the consistency of inspection point identification.

[0093] In some disclosed embodiments, the examination point prediction model includes: a second attention network, which is used to calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text, to obtain the second relevance of each answer text with respect to the examination point prediction, and to weight each answer text based on the second feature representation and the second relevance, to obtain a second weighted representation of each answer text; a second semantic extraction network, which is used to perform semantic extraction on the second weighted representation of each answer text, to obtain a second semantic representation of each answer text; and a second prediction network, which is used to perform prediction based on the second semantic representation of each answer text, to obtain the target examination point involved in the answer text.

[0094] Therefore, by setting the inspection point prediction network to include a second attention network, a second semantic extraction network and a second prediction network, and the second attention network is used to perform feature interaction, the second semantic extraction network is used for semantic extraction, and the second prediction network is used for inspection point prediction, it can help to greatly improve the efficiency of inspection point prediction.

[0095] In some disclosed embodiments, the examination point prediction submodule includes a second model prediction unit, which is used to predict the second sample question text and several second sample answer texts based on the examination point prediction model to obtain a prediction mark of each second sample answer text; wherein the prediction mark indicates whether the second sample answer text is predicted to involve a sample examination point and the sample examination point involved in the prediction of the second sample answer text; the examination point prediction submodule includes a second parameter adjustment unit, which is used to adjust the network parameters of the examination point prediction model based on the difference between the sample mark and the prediction mark of the second sample answer text actually involving the sample examination point.

[0096] Therefore, the second sample question text and several second sample answer texts are predicted through the examination point prediction model to obtain the predicted examination points of each second sample answer text, and based on the difference between the sample mark and the predicted mark of the second sample answer text involving the sample examination point, the network parameters of the examination point prediction model are adjusted. That is, during the training process, the examination point prediction errors of the second sample answer text not involving the sample examination point are shielded, thereby reducing the impact of noise on model training as much as possible, which is conducive to improving the prediction accuracy of the examination point prediction model.

[0097] In some disclosed embodiments, the generation module 73 includes an examination point sorting submodule, which is used to sort the target examination points in order of importance from high to low. The generation module 73 includes an examination point selection submodule, which is used to select the target examination points ranked in the front preset order. The generation module 73 includes an explanation data generation submodule, which is used to generate explanation data for the test text to be processed based on the explanation data of the selected target examination points.

[0098] See also Figure 8 , Figure 8 1 is a schematic diagram of a framework of an embodiment of an electronic device 80 of the present application. The electronic device 80 includes a memory 81 and a processor 82 coupled to each other, the memory 81 stores program instructions, and the processor 82 is used to execute the program instructions to implement the steps in any of the above-mentioned test question explanation generation method embodiments. Specifically, the electronic device 80 may include but is not limited to: a desktop computer, a laptop computer, a server, a mobile phone, a tablet computer, a learning machine, etc., which are not limited here.

[0099] Specifically, the processor 82 is used to control itself and the memory 81 to implement the steps in any of the above-mentioned test question explanation generation method embodiments. The processor 82 can also be called a CPU (Central Processing Unit). The processor 82 may be an integrated circuit chip with signal processing capabilities. The processor 82 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 82 can be implemented by an integrated circuit chip.

[0100] The above scheme obtains the text of the test question to be processed, and the text of the test question to be processed includes the text of the question stem and several answer texts, and then predicts based on the text of the question stem and several answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, the explanation data of the text of the test question to be processed is generated. Therefore, the explanation data of the text of the test question to be processed is generated based on the importance of the target inspection point and the explanation data, so that the explanation data finally generated can take into account as many and important target inspection points as possible. Therefore, more detailed and more targeted explanation data can be generated.

[0101] See also Fig. 9 , Fig. 9 The schematic diagram is a schematic diagram of a computer-readable storage medium 90 of the present application. The computer-readable storage medium 90 stores program instructions 91 that can be executed by a processor, and the program instructions 91 are used to implement the steps in any of the above-mentioned test question explanation generation method embodiments.

[0102] The above scheme obtains the text of the test question to be processed, and the text of the test question to be processed includes the text of the question stem and several answer texts, and then predicts based on the text of the question stem and several answer texts to obtain the target inspection points involved in the text of the test question to be processed and the importance of each target inspection point. On this basis, based on the importance of the target inspection point and the explanation data, the explanation data of the text of the test question to be processed is generated. Therefore, the explanation data of the text of the test question to be processed is generated based on the importance of the target inspection point and the explanation data, so that the explanation data finally generated can take into account as many and important target inspection points as possible. Therefore, more detailed and more targeted explanation data can be generated.

[0103] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0104] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0105] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0107] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0108] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

Claims

1. A method for generating a test question explanation, characterized in that: include: Obtaining a text of the test question to be processed; wherein the text of the test question to be processed includes a text of the question stem and a plurality of answer texts, and the plurality of answer texts constitute a standard answer to the text of the test question to be processed; Based on the attention mechanism, the first feature representation of the question text and the second feature representation of each answer text are calculated for relevance, so as to obtain the first relevance of each answer text with respect to the importance prediction and the second relevance with respect to the inspection point prediction; Performing weighting based on the second feature representation and the first relevance of each of the answer texts to obtain a first weighted representation of each of the answer texts, and performing weighting based on the second feature representation and the second relevance of each of the answer texts to obtain a second weighted representation of each of the answer texts; Predicting based on the second weighted representation of each answer text to obtain the target inspection point involved in the answer text, and predicting based on the first weighted representation of each answer text to obtain the importance of the target inspection point involved in the answer text; Based on the importance of the target examination point and the explanation data, the explanation data of the test question text to be processed is generated.

2. The method according to claim 1, characterized in that The importance is predicted based on an importance prediction model, and the importance prediction model is trained based on a first sample test question text. The first sample test question text includes a first sample question stem text and several first sample answer texts, and the first sample answer text is marked with a sample importance, and the sample importance indicates the criticality of the sample examination point involved in the first sample answer text.

3. The method according to claim 2, characterized in that The step of obtaining the sample importance comprises: The first sample answer texts involving the sample inspection points are respectively used as current sample answer texts, and the sample inspection points involved in the current sample answer texts are used as current inspection points; Based on the semantic similarity between the current sample answer text and the answer text of each subject, the miss rate of several subjects at the current inspection point is obtained; Based on the miss rates corresponding to the sample examination points involved in the first sample answer texts in the first sample test question texts, the sample importance of each of the first sample answer texts is obtained.

4. The method according to claim 2, characterized in that: The importance prediction model includes: A first attention network is used to calculate the relevance based on the first feature representation of the question text and the second feature representation of each answer text to obtain a first relevance of each answer text with respect to importance prediction, and to perform weighting based on the second feature representation and the first relevance of each answer text to obtain a first weighted representation of each answer text; A first semantic extraction network is used to perform semantic extraction on the first weighted representation of each of the answer texts to obtain a first semantic representation of each of the answer texts; The first prediction network is used to predict the importance based on the first semantic representation of each answer text.

5. The method according to claim 2, characterized in that: The training steps of the importance prediction model include: Predicting the first sample question stem text and the plurality of first sample answer texts based on the importance prediction model to obtain a predicted importance of each of the first sample answer texts; Based on the difference between the sample importance and the predicted importance of the first sample answer text related to the sample inspection point, the network parameters of the importance prediction model are adjusted.

6. The method according to claim 1, characterized in that The step of performing prediction based on the second weighted representation of each answer text to obtain the target inspection point involved in the answer text includes: Perform binary classification prediction based on the second weighted representation of each answer text to obtain a first prediction result of each answer text, and perform multi-classification prediction based on the second weighted representation of each answer text to obtain a second prediction result of each answer text; wherein the first prediction result includes the possibility of the examination point involved in the answer text, and the second prediction result includes the candidate examination points involved in the answer text; In response to the first prediction result of the answer text satisfying a preset condition, a target inspection point involved in the answer text is obtained based on the candidate inspection points involved in the answer text.

7. The method according to claim 1, characterized in that The target examination point is predicted based on an examination point prediction model, and the examination point prediction model is trained based on a second sample test question text. The second sample test question text includes a second sample question stem text and several second sample answer texts, and the second sample answer text is marked with a sample tag, and the sample tag indicates whether the second sample answer text actually involves a sample examination point and the sample examination point actually involved by the second sample answer text.

8. The method according to claim 1, characterized in that The step of generating explanation data of the test question text to be processed based on the importance of the target examination point and the explanation data includes: Sorting the target inspection points in descending order of importance; A target examination point ranked in a front preset order is selected, and based on the explanation data of the selected target examination point, explanation data of the test question text to be processed is generated.

9. A test question explanation generating device, characterized in that: include: An acquisition module is used to acquire the text of the test question to be processed; wherein the text of the test question to be processed includes a question stem text and a plurality of answer texts, and the plurality of answer texts constitute a standard answer to the text of the test question to be processed; A prediction module, for calculating the correlation between the first feature representation of the question text and the second feature representation of each answer text based on an attention mechanism, and obtaining a first correlation of each answer text with respect to importance prediction and a second correlation with respect to examination point prediction; performing weighting based on the second feature representation and the first correlation of each answer text, and obtaining a first weighted representation of each answer text, and performing weighting based on the second feature representation and the second correlation of each answer text, and obtaining a second weighted representation of each answer text; performing prediction based on the second weighted representation of each answer text, and obtaining a target examination point involved in the answer text, and performing prediction based on the first weighted representation of each answer text, and obtaining the importance of the target examination point involved in the answer text; A generation module is used to generate explanation data of the test question text to be processed based on the importance of the target examination point and the explanation data.

10. An electronic device, characterized in that: It comprises a memory and a processor coupled to each other, wherein the memory stores program instructions, and the processor is used to execute the program instructions to implement the test question explanation generation method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: Program instructions that can be executed by a processor are stored, and the program instructions are used to implement the test question explanation generation method described in any one of claims 1 to 8.

Images