Test question processing method and device, computer readable storage medium and electronic device

Abstract

The present disclosure discloses a method and device for processing test questions, a computer-readable storage medium, and an electronic device, and relates to the field of natural language processing technology. The test question processing method includes: determining the original test question text of the target science test question; using an anchor prediction model, based on the original test question text of the target science test question, determining the anchor data corresponding to the target science test question, wherein the anchor data is data associated with the examination content of the target science test question, and the anchor prediction model is trained based on the analytical test question text and anchor label data of each of M science test question samples, where M is a positive integer. The test question processing method provided by the present disclosure can use the anchor prediction model to accurately predict the anchor data corresponding to the target science test question, thereby improving the user's learning effect and experience favorability. In other words, compared with the prior art, the present disclosure can be applied to science test question scenarios, and the prediction accuracy is higher.

Description

Technical Field

[0001] The present disclosure relates to the technical field of natural language processing, and in particular to a test question processing method and device, a computer-readable storage medium, and an electronic device. Background Art

[0002] As the core element of the learning machine graph, anchor labels can directly affect the degree of knowledge mastered by users. Especially for science questions, the importance of anchor labels is self-evident. However, since the original test texts of science questions contain many formulas and professional terms that are difficult for machines to understand, the existing anchor label prediction methods based on the original test texts have poor accuracy, which greatly affects the user experience. Summary of the invention

[0003] In view of this, the present disclosure provides a test question processing method and device, a computer-readable storage medium and an electronic device to solve the problem that the accuracy of existing anchor point label prediction methods is relatively poor compared to science test questions.

[0004] In a first aspect, an embodiment of the present disclosure provides a test question processing method, which includes: determining the original test question text of a target science test question; using an anchor prediction model to determine the anchor data corresponding to the target science test question based on the original test question text of the target science test question, wherein the anchor data is data associated with the examination content of the target science test question, and the anchor prediction model is trained based on the analyzed test question texts and anchor label data of M science test question samples, where M is a positive integer.

[0005] In combination with the first aspect, in certain implementations of the first aspect, before using the anchor prediction model to determine the anchor data corresponding to the target science test question based on the original test text of the target science test question, the method also includes: generating analyzed test texts for each of the M science test question samples based on the original test texts of each of the M science test question samples, wherein the original test texts include question stem texts; and training an initial network model based on the analyzed test texts and anchor label data of each of the M science test question samples to obtain an anchor prediction model.

[0006] In combination with the first aspect, in some implementations of the first aspect, the initial network model corresponds to a full-word masking loss function. Based on the parsed test texts and anchor label data of each of the M science test question samples, the initial network model is trained to obtain an anchor prediction model, including: based on the parsed test texts and full-word masking loss function of each of the M science test question samples, the initial network model is pre-trained to obtain a pre-trained model, wherein in the pre-trained model, the full-word masking loss function satisfies a preset convergence condition; based on the parsed test texts and anchor label data of each of the M science test question samples, the pre-trained model is trained to obtain an anchor prediction model.

[0007] In combination with the first aspect, in certain implementations of the first aspect, based on the original test texts of M science test question samples, analyzed test texts of M science test question samples are generated, including: for each science test question sample in the M science test question samples, regularizing the original test question text of the science test question sample to obtain a regularized test question text; and performing formula analysis on the regularized test question text to obtain an analyzed test question text of the science test question sample.

[0008] In combination with the first aspect, in certain implementations of the first aspect, an anchor point prediction model is used to determine anchor point data corresponding to the target science question based on the original question text of the target science question, including: determining N related science questions corresponding to the target science question based on the original question text of the target science question, where N is a positive integer; determining the analytical question text of each of the N related science questions; using the anchor point prediction model, based on the analytical question text of each of the N related science questions, predicting the anchor point data corresponding to each of the N related science questions; determining the anchor point data corresponding to the target science question based on the anchor point data corresponding to each of the N related science questions.

[0009] In combination with the first aspect, in certain implementations of the first aspect, the anchor prediction model includes an anchor prediction model based on the question stem and an anchor prediction model based on the entire question. Using the anchor prediction model, based on the respective analytical question texts of N relevant science questions, the anchor data corresponding to each of the N relevant science questions is predicted, including: for each of the N relevant science questions, based on the question source information of the relevant science question, determining the anchor prediction model corresponding to the relevant science question, wherein the anchor prediction model corresponding to the relevant science question is an anchor prediction model based on the question stem or an anchor prediction model based on the entire question; inputting the analytical question text of the relevant science question into the anchor prediction model corresponding to the relevant science question, and obtaining the anchor data corresponding to the relevant science question.

[0010] In combination with the first aspect, in certain implementations of the first aspect, the anchor data includes anchor label data and label confidence data. Based on the anchor data corresponding to each of the N related science questions, the anchor data corresponding to the target science question is determined, including: based on the anchor prediction models corresponding to each of the N related science questions, P related science questions corresponding to the anchor prediction model based on the whole question are determined, where P is less than or equal to N; if P is a positive integer, based on the label confidence data corresponding to each of the P related science questions and the preset confidence conditions, Q related science questions are selected from the P related science questions, where the label confidence data corresponding to each of the Q related science questions meets the preset confidence conditions, and Q is less than or equal to P; based on the anchor label data corresponding to each of the Q related science questions, the anchor data corresponding to the target science question is determined.

[0011] In combination with the first aspect, in certain implementations of the first aspect, based on the anchor label data corresponding to each of the Q related science questions, determining the anchor data corresponding to the target science question includes: determining the number of related science questions with the highest label confidence for the Q related science questions; if the number of related science questions with the highest label confidence is one, determining the anchor data corresponding to the target science question based on the anchor label data corresponding to the related science question with the highest label confidence; if the number of related science questions with the highest label confidence is multiple, then for the anchor label data corresponding to each of the multiple related science questions with the highest label confidence, based on the anchor label data with the highest number of occurrences, determining the anchor data corresponding to the target science question.

[0012] In combination with the first aspect, in certain implementations of the first aspect, after determining P related science questions corresponding to the anchor prediction model based on the entire question based on the anchor prediction models corresponding to each of the N related science questions, the method further includes: if P is zero, then for the anchor label data corresponding to each of the N related science questions, determining the anchor data corresponding to the target science question based on the anchor label data with the highest number of occurrences.

[0013] In combination with the first aspect, in certain implementations of the first aspect, based on the original test text of the target science test question, determining N related science test questions corresponding to the target science test question includes: obtaining a preset science test question bank; determining N related science test questions based on the original test text of the target science test question and the preset science test question bank.

[0014] In combination with the first aspect, in certain implementations of the first aspect, determining the original test text of the target science test question includes: obtaining a test question image corresponding to the target science test question; performing a character recognition operation on the test question image to obtain the original test text of the target science test question.

[0015] In combination with the first aspect, in certain implementations of the first aspect, the M science test question samples include at least one of mathematics test question samples, physics test question samples, and chemistry test question samples; and / or, the anchor point data includes anchor point label data, and the anchor point data also includes label confidence data.

[0016] In combination with the first aspect, in certain implementations of the first aspect, the original test question text also includes at least one of a knowledge point name, an answer text, and an answer analysis text.

[0017] In a second aspect, an embodiment of the present disclosure provides an anchor point prediction device, which includes: a first determination module, used to determine the original test question text of the target science test question; a second determination module, used to use an anchor point prediction model to determine the anchor point data corresponding to the target science test question based on the original test question text of the target science test question, wherein the anchor point data is data associated with the examination content of the target science test question, and the anchor point prediction model is trained based on the analyzed test question texts and anchor point label data of M science test question samples, where M is a positive integer.

[0018] In a third aspect, an embodiment of the present disclosure provides an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor, wherein the processor is used to execute the method mentioned in the first aspect above.

[0019] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is used to execute the method mentioned in the first aspect.

[0020] The test question processing method provided by the disclosed embodiment can use the anchor point prediction model to accurately predict the anchor point data corresponding to the target science test question, thereby improving the user's learning effect and experience. In other words, compared with the prior art, the disclosed embodiment can be applied to science test question scenarios and has higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Shown is a schematic diagram of an application scenario provided by an embodiment of the present disclosure.

[0022] Figure 2 The figure is a flow chart of a model training method provided in one embodiment of the present disclosure.

[0023] Figure 3 The figure is a schematic diagram of a process of obtaining an anchor prediction model based on training an initial network model according to an embodiment of the present disclosure.

[0024] Figure 4aThe figure is a schematic diagram of the training process of the pre-training model provided in one embodiment of the present disclosure.

[0025] Figure 4b The figure is a schematic diagram of the training process of the anchor point prediction model provided by an embodiment of the present disclosure.

[0026] Figure 5 The figure is a schematic diagram of a process for generating analytical test texts for M science test sample questions provided by an embodiment of the present disclosure.

[0027] Figure 6 The figure is a flowchart of a test question processing method provided in one embodiment of the present disclosure.

[0028] Figure 7 The figure is a schematic diagram of a process for determining anchor point data corresponding to a target science test question provided by an embodiment of the present disclosure.

[0029] Figure 8 The figure is a schematic diagram of a process for determining the anchor point data corresponding to each of N related science test questions provided by an embodiment of the present disclosure.

[0030] Fig. 9 The figure is a schematic diagram of a process for determining anchor point data corresponding to a target science test question provided by another embodiment of the present disclosure.

[0031] Fig.10 The figure is a schematic diagram of a process for determining anchor point data corresponding to a target science test question provided by another embodiment of the present disclosure.

[0032] Fig.11 The figure is a schematic diagram of a process for determining N related science test questions provided by an embodiment of the present disclosure.

[0033] Fig.12 The figure is a schematic diagram of a flow chart of determining the original test text of a target science test question provided by an embodiment of the present disclosure.

[0034] Fig.13 Shown is a schematic diagram of the structure of a model training device provided in one embodiment of the present disclosure.

[0035] Fig.14 Shown is a schematic diagram of the structure of a test question processing device provided by an embodiment of the present disclosure.

[0036] Fig.15 Shown is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0038] The anchor tag of the test question is a further refinement of the knowledge point tag. Compared with the knowledge point tag, it contains more detailed information about the test content. For example, math test question A is: "Student Xiao Ming deposited 250 yuan into the bank, trying to deposit the whole amount for half a year. After half a year, he got a total of 252.7 yuan in principal and interest. What is the annual interest rate of the bank for half a year?" Correspondingly, under normal circumstances, the knowledge point tag corresponding to math test question A is: "One-variable linear equation", and the anchor tag is: "One-variable linear equation, interest rate problem, involving the calculation method of the bank's annual interest rate". It can be seen that the anchor tag can represent the test content information of the test question. How to accurately determine the anchor tag of the test question is a key research issue in current artificial intelligence education.

[0039] At present, one of the existing mainstream anchor label prediction methods is to use Chinese word segmentation tools to segment the original test text, and then use the Text Convolutional Neural Networks (TextCNN) model to predict anchor labels. However, the accuracy of the above-mentioned anchor label prediction method based on semantic understanding is not ideal. Especially for science test questions, because the original test text of science test questions contains many formulas and professional terms that are difficult to be understood by artificial intelligence machines, and these available contents are lost in large quantities during the word segmentation process, which leads to the TextCNN model being unable to predict anchor labels with higher accuracy. Therefore, when the existing anchor label prediction method is applied to the application scenario of science test questions, the prediction accuracy is even worse, which greatly affects the user's learning effect.

[0040] In order to solve the above problems, the embodiments of the present disclosure provide a model training method and device, and a test question processing method and device to achieve the purpose of improving the prediction accuracy of anchor point labels for science test questions.

[0041] Combine the following Figure 1 The application scenarios of the embodiments of the present disclosure are briefly introduced.

[0042] Figure 1 FIG. 1 is a schematic diagram of an application scenario provided by an embodiment of the present disclosure. Figure 1 As shown, the application scenario is a photo search scenario. Specifically, the photo search scenario includes a server 110 and a user terminal 120 that is communicatively connected to the server 110. An anchor prediction model may be deployed in the server 110, and the server 110 is used to execute the test question processing method mentioned in the embodiment of the present disclosure. The user terminal 120 may be a terminal device of the user. It can be understood that the user may be a student user, a parent user or a teacher user, and the embodiments of the present disclosure will not be described one by one.

[0043] Exemplarily, in actual application, the user uses the user terminal 120 to take a test image, and sends the test image to the server 110 with the help of the user terminal 120. After receiving the test image, the server 110 first obtains the original test text of the test based on the test image, and then predicts the anchor data corresponding to the test based on the original test text and the anchor prediction model, and sends the predicted anchor data to the user terminal 120. After receiving the anchor data, the user terminal 120 presents the anchor data to the user so that the user can understand the test content information of the test more clearly, so as to better learn knowledge. Among them, the anchor data is data associated with the test content of the test, for example, the anchor data includes anchor tag data (i.e., anchor tag) and tag confidence data (i.e., tag confidence).

[0044] Exemplarily, the training method of the anchor prediction model deployed in the server 110 includes: generating analyzed test texts for M science test question samples based on the original test question texts of each of the M science test question samples, wherein the original test question texts include the question stem texts and M is a positive integer; training an initial network model based on the analyzed test question texts and anchor label data of each of the M science test question samples to obtain an anchor prediction model, wherein the anchor prediction model is used to predict the anchor data corresponding to the science test question to be predicted based on the test texts of the science test question to be predicted.

[0045] Exemplarily, the user terminal 120 mentioned above is a mobile terminal such as a tablet computer or a mobile phone.

[0046] Combine the following Figures 2 to 5 A brief introduction to the model training method disclosed in the present invention is given.

[0047] Figure 2 FIG. 1 is a flow chart of a model training method provided by an embodiment of the present disclosure. Figure 2 As shown, the model training method provided by the embodiment of the present disclosure includes the following steps.

[0048] Step S210, generating respective analyzed test texts of the M science test question samples based on the respective original test texts of the M science test question samples.

[0049] Exemplarily, the M science test samples include at least one of mathematics test samples, physics test samples and chemistry test samples, and M is a positive integer. Since many formulas and professional terms in science subjects are universal, the disclosed embodiments do not uniformly limit the specific subjects to which the science test samples belong, as long as they are science subjects. In order to further optimize the prediction effect of the anchor prediction model obtained by training, it is also possible to use only the training data set of the test sample generation model of a single subject, and then train to obtain a more targeted anchor prediction model. For example, only the training data set of the mathematics test sample generation model is used to obtain an anchor prediction model that is more suitable for data test questions.

[0050] Exemplarily, the original test text includes a stem text. Among them, the stem text can also be called the question text. In addition to the stem text, the original test text can also include at least one of the following texts: a knowledge point name, an answer text, and an answer analysis text. The analyzed test text refers to the text obtained after the original test text is analyzed. The purpose of the analysis is to translate and analyze science-specific content such as science symbols and / or science-related professional terms, so that the model can more fully learn the content contained in the original test text. Among them, the specific implementation method of the analysis can be determined according to the actual situation, and the specific method of the analysis is not uniformly limited in the embodiments of the present disclosure.

[0051] Step S220, based on the parsed test question texts and anchor label data (i.e., anchor label) of each of the M science test question samples, an initial network model is trained to obtain an anchor prediction model.

[0052] Exemplarily, the anchor prediction model is used to predict the anchor data corresponding to the science test questions to be predicted based on the parsed test text of the science test questions to be predicted. The anchor data is data associated with the examination content of the science test questions to be predicted. In some embodiments, the model structure of the initial network model and the anchor prediction model is the same, and the difference lies in the model parameters. That is, the initial network model is trained based on the training data set of the model, and the model parameters of the initial network model are adjusted to obtain the anchor prediction model.

[0053] Exemplarily, in actual application, first, based on the original test texts of M science test samples, the analyzed test texts of M science test samples are generated, and then based on the analyzed test texts and anchor label data of M science test samples, the initial network model is trained to obtain the anchor prediction model.

[0054] The model training method provided in the embodiment of the present disclosure can train an anchor prediction model suitable for science test scenarios. That is, compared with the prior art, the anchor prediction model trained in the embodiment of the present disclosure is more suitable for science test scenarios. Based on this, the embodiment of the present disclosure provides a prerequisite for improving the accuracy of anchor prediction for science test questions and optimizing user learning effects with the help of the anchor prediction model obtained through training.

[0055] Combine the following Figure 3 The specific training method of the model is further illustrated with examples.

[0056] Figure 3 FIG. 1 is a flow chart of training an initial network model and obtaining an anchor prediction model according to an embodiment of the present disclosure. Figure 2 Based on the embodiment shown, Figure 3 The embodiment shown is described below in detail. Figure 3 The embodiment shown and Figure 2 The differences and similarities of the illustrated embodiments are not described in detail.

[0057] like Figure 3 As shown, in the embodiment of the present disclosure, the initial network model corresponds to a full-word masking loss function. In addition, based on the parsed test texts and anchor label data of the M science test samples, the initial network model is trained to obtain the anchor prediction model step, including the following steps.

[0058] Step S310, based on the parsed test texts and the full-word masking loss function of each of the M science test sample questions, pre-train the initial network model to obtain a pre-trained model.

[0059] Exemplarily, in the pre-trained model, the full word masking loss function satisfies the preset convergence condition. That is, the initial network model is pre-trained based on the parsed test texts of M science test questions, and the full word masking loss function is used as the training objective function to adjust the model parameters of the initial network model until a pre-trained model is obtained in which the full word masking loss function satisfies the preset convergence condition.

[0060] Step S320: Based on the analysis question texts of the M science question samples and the anchor point label data of the M science question samples, a pre-training model is trained to obtain an anchor point prediction model.

[0061] That is to say, after obtaining a pre-trained model whose full-word masking loss function satisfies the preset convergence condition, the pre-trained model is further trained based on the parsed test texts and anchor label data of the M science test samples, thereby obtaining an anchor prediction model that can predict anchor data. It can be understood that the anchor label data refers to the anchor labels of the samples used in the model training process, which can be regarded as reference data for adjusting the model parameters.

[0062] Since the full-word masking loss function can fully learn the contextual information of the text, the model training method provided by the embodiment of the present disclosure achieves the purpose of tuning the model based on a specific domain data set (i.e., continuing pre-training) based on the parsed test texts of the M science test samples. Based on this, the embodiment of the present disclosure can make the trained model suitable for science test scenarios, thereby improving the model's anchor point prediction accuracy for science test questions.

[0063] In some embodiments, the initial network model is an anchor prediction model trained based on a non-specific field data set. For example, the non-specific field data set does not distinguish between fields such as science and liberal arts, and is a general test question data set. In other words, the initial network model is an anchor prediction model that is common to all disciplines, and the present disclosure performs tuning operations for the science field on the general anchor prediction model, thereby generating an anchor prediction model that is more suitable for science test question scenarios. With such a setting, the model training process can be further simplified with the help of a general anchor prediction model, the amount of training calculations can be reduced, and training resources and training costs can be saved.

[0064] Combine the following Figure 4a and Figure 4b Examples are given to illustrate the model architecture and training process of the pre-trained model and anchor prediction model.

[0065] Figure 4a FIG. 1 is a schematic diagram of the training process of the pre-training model provided by an embodiment of the present disclosure. Figure 4a As shown, in the disclosed embodiment, the pre-trained model obtained by training corresponds to multiple encoder layers, and the pre-trained model corresponds to a whole word masking loss function (Whole Word Masking LM Lose). It can also be understood that the initial network model corresponding to the pre-trained model corresponds to multiple encoder layers and a whole word masking loss function. That is, the difference between the initial network model and the pre-trained model obtained by training is only the difference in model parameters.

[0066] For example, Figure 4a The training process shown can correspond to Figure 3Step S310 mentioned in the illustrated embodiment. Specifically, in the disclosed embodiment, for each science test question sample, the original test question text includes a question stem text, a knowledge point name, an answer text, and an answer analysis text. The question stem text, the knowledge point name, the answer text, and the answer analysis text are spliced ​​to obtain a spliced ​​text, and then the spliced ​​text is parsed to obtain the parsed test question text of the science test question sample. Then, after obtaining the parsed test question texts of the M science test question samples, the initial network model is pre-trained based on the parsed test question texts of the M science test question samples, and the full-word masking loss function is used as the objective function of the training, and the model parameters of the initial network model are adjusted until a pre-trained model whose full-word masking loss function meets the preset convergence conditions is obtained.

[0067] Exemplarily, the model architecture of the pre-trained model provided in the embodiments of the present disclosure is the model architecture of the Roberta model.

[0068] Based on Figure 4a After training the pre-trained model, it is necessary to Figure 4b The training process shown continues to train the obtained pre-trained model and fine-tune the model to finally obtain an anchor prediction model that can predict anchor data. Figure 4b This article gives an example to illustrate the specific implementation method of model fine-tuning.

[0069] Figure 4b FIG. 1 is a schematic diagram of the training process of the anchor prediction model provided by an embodiment of the present disclosure. Figure 4b As shown in the figure, in the training process of the anchor prediction model, the classification token (CLS Token) layer, the fully connected (FC) layer and the normalization (softmax) function layer in the model are involved. It can be understood that the CLSToken layer, the FC layer and the softmax function layer exist in the initial network model, the pre-trained model and the anchor prediction model, but are only used in the process of training the anchor prediction model. In the actual application process, the vector obtained by the CLS token layer realizes the final anchor prediction through the FC layer and the softmax function layer.

[0070] For example, Figure 4b The training process shown can correspond to Figure 3Step S320 mentioned in the illustrated embodiment. Specifically, in the training process of the anchor prediction model, the input of the pre-trained model is still the analytical test text corresponding to each of the M science test question samples mentioned above, and the output of the pre-trained model is the predicted anchor labels 1 to n. Then, based on the predicted anchor labels 1 to n (which can be regarded as the anchor labels predicted by the model), and the difference between the anchor label data corresponding to each of the M science test question samples (which can be regarded as the correct anchor labels), the model parameters of the pre-trained model can be adjusted, and finally an anchor prediction model capable of predicting the anchor data can be obtained.

[0071] Combine the following Figure 5 The following example illustrates the specific implementation method of generating analytical test texts for M science test sample questions.

[0072] Figure 5 FIG. 1 is a flow chart of generating the analysis text of M science test questions samples provided by an embodiment of the present disclosure. Figure 2 Based on the embodiment shown, Figure 5 The embodiment shown is described below in detail. Figure 5 The embodiment shown and Figure 2 The differences and similarities of the illustrated embodiments are not described in detail.

[0073] like Figure 5 As shown, in the embodiment of the present disclosure, the steps of generating the analyzed test texts of M science test samples respectively based on the original test texts of M science test samples include the following steps.

[0074] Step S510, for each science question sample in the M science question samples, regularization is performed on the original question text of the science question sample to obtain a regularized question text.

[0075] Exemplarily, the original test question text is regularized based on the preset regularization processing rules to obtain the regularized test question text. The purpose of the regularization processing can be to pre-clean the original test question text so that the model can better learn and utilize the information contained in the original test question text. Among them, the construction method of the preset regularization processing rules can include: firstly, a large number of science test question samples are collected for analysis as examples of the regularization processing rules, and a regularization library for science test questions including the preset regularization processing rules is formed.

[0076] Step S520, performing formula parsing processing on the regularized test question text to obtain the parsed test question text of the science test question sample.

[0077] Exemplarily, the regularized test text is subjected to Latex formula parsing (i.e., the Latex format content in the regularized test text is parsed) to generate a parsed test text. For example, the Latex formula parsing includes processing the fraction symbol in the Latex format into "\", processing the exponentiation symbol in the Latex format into "^", and so on.

[0078] For example, the science test sample A is a mathematics test sample. The stem (also called the text or title) of the mathematics test sample is: Then the value of x is.

[0079] If the science test sample A is processed by word segmentation (existing technical solution), the word segmentation result is: NUM VAR_QUA NUM EQUAL NUM VARx value TARGET PM NUM. If the science test sample A is processed based on the above steps S510 and S520, the parsed test text is: -5 ∧ x ∧ 2-1=1 then the value of x.

[0080] In summary, after the word segmentation processing of the science test sample A, the numbers in the question stem are converted to NUM, the unknown number (also known as the variable) x in the question stem is converted to VARx, and some mathematical symbols are lost. It can be seen that the word segmentation processing of the science test sample and then generating the training data set for model training will greatly affect the model prediction accuracy. Figure 5 The illustrated embodiment obtains the analyzed test text after processing the science test sample A, which largely retains the content covered by the original test text, thereby providing a data basis for improving the prediction accuracy of the trained anchor point prediction model.

[0081] The model training method provided in the embodiment of the present disclosure generates analyzed test text that is more suitable for model training by performing a series of preprocessing (including regularization processing and / or formula parsing processing) on ​​the original test text, thereby providing a data basis for training an anchor point prediction model with higher prediction accuracy.

[0082] Combination of the above Figures 2 to 5 , describes in detail the embodiment of the model training method disclosed in the present invention, and the following is combined with Figures 6 to 12 , an embodiment of the test question processing method disclosed in the present invention is described in detail.

[0083] Figure 6 FIG. 1 is a flow chart of a method for processing test questions provided by an embodiment of the present disclosure. Figure 6 As shown, the test question processing method provided by the embodiment of the present disclosure includes the following steps.

[0084] Step S610, determining the original test text of the target science test question.

[0085] For example, the target science test questions refer to test questions that require anchor data prediction, such as test questions that users input by taking photos.

[0086] Step S620, using the anchor prediction model, based on the original test text of the target science test question, determine the anchor data corresponding to the target science test question.

[0087] Exemplarily, the anchor prediction model mentioned in step 620 refers to an anchor prediction model that can be applied to science test scenarios. Further, the anchor prediction model mentioned in step 620 is trained based on the model training method mentioned in any of the above embodiments.

[0088] Exemplarily, in actual application, the original test text of the target science test question is first determined, and then the anchor point prediction model is used to determine the anchor point data corresponding to the target science test question based on the original test text of the target science test question.

[0089] The test question processing method provided by the embodiment of the present disclosure can use the anchor point prediction model to accurately predict the anchor point data corresponding to the target science test questions, thereby improving the user's learning effect and experience favorability.

[0090] It can be understood that there are many specific implementation methods for using the anchor prediction model to predict the anchor data corresponding to the target science test questions. For example, the original test text of the target science test questions is parsed to obtain the corresponding parsed test text, and then the parsed test text is input into the anchor prediction model so that the anchor prediction model can make predictions. In addition, in order to further improve the prediction accuracy, it is also possible to use Figure 7 Predictions are made in the manner shown.

[0091] Figure 7 The figure is a flow chart of determining the anchor point data corresponding to the target science test question provided by an embodiment of the present disclosure. Figure 6 Based on the embodiment shown, Figure 7 The embodiment shown is described below in detail. Figure 7 The embodiment shown and Figure 6 The differences and similarities of the illustrated embodiments are not described in detail.

[0092] like Figure 7 As shown, in the embodiment of the present disclosure, the anchor point prediction model is used to determine the anchor point data corresponding to the target science test question based on the original test question text of the target science test question, including the following steps.

[0093] Step S710, based on the original test text of the target science test question, determine N related science test questions corresponding to the target science test question.

[0094] Exemplarily, N is a positive integer greater than 1, such as N=10. For example, based on the original test text of the target science test question, 50 science test questions related to the target science test question are searched, and then the 50 science test questions are sorted by relevance, and the top 10 science test questions (i.e., top 10) are determined as relevant science test questions. Optionally, based on the original test text of the target science test question, 50 science test questions related to the target science test question are searched in a preset test question library using a text similarity algorithm.

[0095] Step S720, determining the analysis texts of the N related science test questions.

[0096] Exemplarily, the original test text of each relevant science test question is parsed to obtain the parsed test text of the relevant science test question. It can be understood that, relative to the anchor prediction model, the relevant science test question can be regarded as the science test question to be predicted mentioned in the above embodiment.

[0097] Step S730, using the anchor point prediction model, based on the analysis question texts of the N related science questions, predict the anchor point data corresponding to each of the N related science questions.

[0098] Step S740, based on the anchor point data corresponding to each of the N related science test questions, determine the anchor point data corresponding to the target science test question.

[0099] Exemplarily, among the anchor point data corresponding to each of N related science test questions, the anchor point data that best fits the target science test question is selected as the anchor point data corresponding to the target science test question.

[0100] That is to say, in the disclosed embodiment, the anchor prediction model is not directly used to predict the anchor point of the target science test question, but the anchor prediction model is used to predict the anchor points of N related science test questions respectively, and then the anchor point data corresponding to the target science test question is determined based on the anchor point data corresponding to each of the N related science test questions. It can be seen that the disclosed embodiment can fully take into account the knowledge point information that may be involved in the target science test question, thereby achieving the purpose of determining more accurate anchor point data for the target science test question.

[0101] It can be understood that during the training process of the anchor prediction model, the data richness of the training data set can not only affect the prediction effect of the anchor prediction model finally trained, but also determine the data input of the anchor prediction model in the actual application process.

[0102] Specifically, if during the training process of the anchor prediction model, the original test text used to generate the training data set only includes the text of the question stem, then Figure 7 In the embodiment shown, in the actual application of the anchor prediction model, if the anchor prediction model is used to predict the anchor data corresponding to the relevant science test question A, the data input is the stem text of the relevant science test question A (for example, the stem text is parsed and then input into the anchor prediction model). The anchor prediction model trained in this case can be called a stem-based anchor prediction model. If, during the training process of the anchor prediction model, the original test question text that generates the training data set includes not only the stem text, but also the knowledge point name, answer text, and answer parsing text, then, based on Figure 7 In the embodiment shown, in the actual application process of the anchor prediction model, if the anchor prediction model is used to predict the anchor data corresponding to the relevant science test question A, the data input is the question stem text, knowledge point name, answer text and answer analysis text of the relevant science test question A (for example, the question stem text, knowledge point name, answer text and answer analysis text are spliced, and the spliced ​​text is parsed and then input into the anchor prediction model). The anchor prediction model trained in this case can be called an anchor prediction model based on the entire question.

[0103] Therefore, in view of the difference in the data richness of the original test text, the anchor prediction model mentioned in the above embodiment can be two types, namely, an anchor prediction model based on the question stem and an anchor prediction model based on the entire question. Figure 8 Examples are given to illustrate the specific applications of the two models.

[0104] Figure 8 The figure is a flow chart of determining the anchor point data corresponding to each of N related science test questions provided by an embodiment of the present disclosure. Figure 7 Based on the embodiment shown, Figure 8 The embodiment shown is described below in detail. Figure 8 The embodiment shown and Figure 7 The differences and similarities of the illustrated embodiments are not described in detail.

[0105] like Figure 8 As shown, in the embodiment of the present disclosure, the anchor point prediction model is used to predict the anchor point data corresponding to each of the N related science test questions based on the analyzed test texts of each of the N related science test questions, including the following steps.

[0106] Step S810, for each relevant science question in the N relevant science questions, based on the question source information of the relevant science question, determine the anchor point prediction model corresponding to the relevant science question.

[0107] The source information of the relevant science questions (also called the source tag of the questions) can represent the level of detail of the original text of the relevant science questions. For example, the source information of the questions is "Aibiaokepaisou questions", "Kuaihuowang questions", etc.

[0108] Exemplarily, the anchor prediction model corresponding to the relevant science test questions is an anchor prediction model based on the question stem or an anchor prediction model based on the entire question. That is, the anchor prediction model corresponding to the relevant science test questions is one of the above two models.

[0109] Step S820: input the analysis question text of the relevant science question into the anchor point prediction model corresponding to the relevant science question to obtain the anchor point data corresponding to the relevant science question.

[0110] Since the source information of the test questions can characterize the level of detail of the original test question text, the disclosed embodiment can select an anchor point prediction model that is more suitable for the test question based on the source information of the test questions, thereby further improving the accuracy of the anchor point prediction.

[0111] As mentioned above, anchor data can include not only anchor label data, but also label confidence data corresponding to the anchor label. For example, if the anchor label is "a linear equation with one variable, interest rate problem, involving the calculation method of bank annual interest rate", the label confidence corresponding to the anchor label is 0.8. Fig. 9 An example is given to illustrate how to use voting to determine the anchor data corresponding to the target science test question based on the anchor data corresponding to each of N related science test questions.

[0112] Fig. 9 The figure is a flow chart of determining the anchor point data corresponding to the target science test question provided by another embodiment of the present disclosure. Figure 7 Based on the embodiment shown, Fig. 9 The embodiment shown is described below in detail. Fig. 9 The embodiment shown and Figure 7 The differences and similarities of the illustrated embodiments are not described in detail.

[0113] like Fig. 9 As shown, in the embodiment of the present disclosure, based on the anchor point data corresponding to each of N related science test questions, the step of determining the anchor point data corresponding to the target science test question includes the following steps.

[0114] Step S910, based on the anchor prediction models corresponding to the N related science test questions, determine P related science test questions corresponding to the anchor prediction model based on the whole test question.

[0115] That is to say, from N relevant science test questions, P relevant science test questions corresponding to the anchor prediction model based on the full question are selected, where P is less than or equal to N.

[0116] Step S920, determine whether P is a positive integer or zero.

[0117] Exemplarily, in actual application, if P is a positive integer, steps S930 and S940 are executed, and if P is zero, step S950 is executed.

[0118] Step S930: Based on the label confidence data corresponding to each of the P related science test questions and the preset confidence conditions, Q related science test questions are selected from the P related science test questions, where Q is less than or equal to P.

[0119] It is understood that the preset reliability condition can be determined according to actual conditions, and the embodiments of the present disclosure do not impose a uniform limitation on this. For example, the preset reliability condition is that the preset reliability is greater than 0.5. If Q is zero, the preset reliability condition can be modified so that Q is not zero.

[0120] That is to say, for the P relevant science test questions obtained in step S910, further selection is performed based on the label confidence data, and Q relevant science test questions that meet the preset confidence conditions are selected from the P relevant science test questions.

[0121] Step S940, based on the anchor point label data corresponding to each of the Q related science questions, determine the anchor point data corresponding to the target science question.

[0122] After multiple layers of screening, relevant science questions with poor quality and low label confidence have been eliminated, thereby achieving the purpose of determining the anchor data corresponding to the target science questions based on the anchor label data corresponding to Q relevant science questions with high quality and high label confidence.

[0123] Step S950, for each of the N related science test questions corresponding to the anchor tag data, based on the anchor tag data with the highest number of occurrences, determine the anchor data corresponding to the target science test question.

[0124] Exemplarily, in actual application, based on the anchor prediction models corresponding to each of N related science questions, P related science questions corresponding to the anchor prediction model based on the full question are determined. If P is a positive integer (i.e., there are related science questions corresponding to the anchor prediction model based on the full question), then based on the label confidence data corresponding to each of the P related science questions and preset confidence conditions, Q related science questions are selected from the P related science questions, and then based on the anchor label data corresponding to each of the Q related science questions, the anchor data corresponding to the target science question is determined. If P is zero (i.e., there is no related science question corresponding to the anchor prediction model based on the full question), then for the anchor label data corresponding to each of the N related science questions, the anchor data corresponding to the target science question is determined based on the anchor label data with the highest number of occurrences.

[0125] That is to say, if there are relevant science questions corresponding to the anchor prediction model based on the full question, the anchor data corresponding to the target science question is determined based on the relevant science questions corresponding to the anchor prediction model based on the full question; if there are no relevant science questions corresponding to the anchor prediction model based on the full question, the N relevant science questions are no longer screened, and the anchor data corresponding to the target science question is determined directly based on the anchor label data with the highest number of occurrences (for example, the anchor label data with the highest number of occurrences is directly determined as the anchor data corresponding to the target science question).

[0126] The disclosed embodiment can use different models to predict anchor points for questions from different sources, and perform multi-layer screening of source priority and label confidence based on the source of the question, thereby obtaining the anchor point data corresponding to the target science question. In other words, the disclosed embodiment can accurately and flexibly determine the anchor point data corresponding to the target science question based on the actual situation of the relevant science question.

[0127] It can be understood that, relative to the Q related science questions, there may be two situations where the number of related science questions with the highest label confidence is one or more. Fig.10 Give examples to illustrate how to deal with these two situations.

[0128] Fig.10 The figure is a schematic diagram of a process for determining anchor point data corresponding to a target science test question provided by another embodiment of the present disclosure. Fig. 9 Based on the embodiment shown, Fig.10 The embodiment shown is described below in detail. Fig.10 The embodiment shown and Fig. 9 The differences and similarities of the illustrated embodiments are not described in detail.

[0129] like Fig.10As shown, in the embodiment of the present disclosure, based on the anchor point label data corresponding to each of Q related science test questions, the step of determining the anchor point data corresponding to the target science test question includes the following steps.

[0130] Step S1010, for the Q relevant science questions, determine the number of relevant science questions with the highest label confidence.

[0131] Step S1020, determining whether the number of relevant science test questions with the highest label confidence is one or more.

[0132] Exemplarily, in actual application, if the number of relevant science questions with the highest label confidence is one, step S1030 is executed; if the number of relevant science questions with the highest label confidence is multiple, step S1040 is executed.

[0133] Step S1030, determining the anchor data corresponding to the target science question based on the anchor label data corresponding to the related science question with the highest label confidence.

[0134] Step S1040, for each of the anchor label data corresponding to a plurality of related science test questions with the highest label confidence, based on the anchor label data with the highest number of occurrences, determine the anchor data corresponding to the target science test question.

[0135] For example, the anchor point label data with the highest number of occurrences is directly determined as the anchor point data corresponding to the target science test question. It can be understood that the anchor point data corresponding to the target science test question may only include the anchor point label data.

[0136] Exemplarily, in the actual application process, for Q related science questions, the number of related science questions with the highest label confidence is determined. If the number of related science questions with the highest label confidence is one (i.e., there is no tie), then based on the anchor label data corresponding to the related science question with the highest label confidence, the anchor data corresponding to the target science question is determined. If the number of related science questions with the highest label confidence is multiple (i.e., there is a tie), then for the anchor label data corresponding to each of the multiple related science questions with the highest label confidence, based on the anchor label data with the highest number of occurrences, the anchor data corresponding to the target science question is determined. It can be seen that the disclosed embodiment can greatly improve the accuracy of the anchor data corresponding to the target science question.

[0137] Fig.11 FIG. 1 is a flow chart of determining N related science test questions provided by an embodiment of the present disclosure. Figure 7 Based on the embodiment shown, Fig.11 The embodiment shown is described below in detail. Fig.11 The embodiment shown and Figure 7 The differences and similarities of the illustrated embodiments are not described in detail.

[0138] like Fig.11 As shown, in the embodiment of the present disclosure, based on the original test text of the target science test question, the steps of determining N related science test questions corresponding to the target science test question include the following steps.

[0139] Step S1110, obtaining a preset science test question bank.

[0140] Step S1120, determining N relevant science questions based on the original question text of the target science question and the preset science question bank.

[0141] That is, in the disclosed embodiment, the N related science questions corresponding to the target science question are selected from the preset science question bank. This arrangement can not only ensure the diversification of related science questions, but also reduce the difficulty of obtaining related science questions.

[0142] Fig.12 The figure shows a flow chart of determining the original test text of a target science test question provided by an embodiment of the present disclosure. Figure 6 Based on the embodiment shown, Fig.12 The embodiment shown is described below in detail. Fig.12 The embodiment shown and Figure 6 The differences and similarities of the illustrated embodiments are not described in detail.

[0143] like Fig.12 As shown, in the embodiment of the present disclosure, the step of determining the original test text of the target science test question includes the following steps.

[0144] Step S1210, obtaining the test question image corresponding to the target science test question.

[0145] For example, the target science test questions are test questions in a book. The user can take a photo of the test questions in the book to obtain the test question image.

[0146] Step S1220, performing character recognition operation on the test question image to obtain the original test question text of the target science test question.

[0147] Exemplarily, an optical character recognition (OCR) operation is performed on the test question image to obtain the original test question text of the target science test question.

[0148] The disclosed embodiments can simplify the method of obtaining the original test question text, facilitate user operation, and thus improve user experience.

[0149] Combination of the above Figures 2 to 12 , describes in detail the method embodiment of the present disclosure, and the following is combined with Figures 13 to 15 , describes the device embodiment of the present disclosure in detail. In addition, it should be understood that the description of the method embodiment corresponds to the description of the device embodiment, so the parts not described in detail can refer to the previous method embodiment.

[0150] Fig.13 FIG. 1 is a schematic diagram of the structure of a model training device provided by an embodiment of the present disclosure. Fig.13 As shown, the model training device provided by the embodiment of the present disclosure includes a generation module 1310 and a training module 1320. Specifically, the generation module 1310 is used to generate the respective parsed test texts of M science test question samples based on the respective original test texts of M science test question samples, wherein the original test texts include the question stem texts, and M is a positive integer. The training module 1320 is used to train the initial network model based on the respective parsed test texts and anchor label data of the M science test question samples to obtain an anchor prediction model, wherein the anchor prediction model is used to predict the anchor data corresponding to the science test question to be predicted based on the parsed test texts of the science test question to be predicted, and the anchor data is data associated with the examination content of the science test question to be predicted.

[0151] In some embodiments, the initial network model corresponds to a full-word masking loss function. In addition, the training module 1320 is also used to pre-train the initial network model based on the respective parsed test texts and the full-word masking loss function of the M science test question samples to obtain a pre-trained model, wherein in the pre-trained model, the full-word masking loss function satisfies a preset convergence condition; based on the respective parsed test texts of the M science test question samples and the respective anchor point label data of the M science test question samples, the pre-trained model is trained to obtain an anchor point prediction model.

[0152] In some embodiments, the generation module 1310 is also used to, for each science question sample in the M science question samples, perform regularization on the original question text of the science question sample to obtain a regularized question text; and perform formula parsing on the regularized question text to obtain an parsed question text of the science question sample.

[0153] In some embodiments, the original test question text also includes at least one of a knowledge point name, an answer text, and an answer analysis text.

[0154] In some embodiments, the M science test question samples include at least one of mathematics test question samples, physics test question samples and chemistry test question samples; and / or, the anchor point data includes anchor point label data, and the anchor point data also includes label confidence data.

[0155] Fig.14FIG. 1 is a schematic diagram of the structure of a test question processing device provided by an embodiment of the present disclosure. Fig.14 As shown, the test question processing device provided by the embodiment of the present disclosure includes a first determination module 1410 and a second determination module 1420. Specifically, the first determination module 1410 is used to determine the original test question text of the target science test question. The second determination module 1420 is used to use the anchor prediction model to determine the anchor point data corresponding to the target science test question based on the original test question text of the target science test question.

[0156] In some embodiments, the second determination module 1420 is also used to determine N related science questions corresponding to the target science question based on the original question text of the target science question, where N is a positive integer; obtain the original question text of each of the N related science questions; use the anchor point prediction model to predict the anchor point data corresponding to each of the N related science questions based on the original question text of each of the N related science questions; and determine the anchor point data corresponding to the target science question based on the anchor point data corresponding to each of the N related science questions.

[0157] In some embodiments, the anchor prediction model includes an anchor prediction model based on the question stem and an anchor prediction model based on the entire question. The second determination module 1420 is also used to, for each of the N relevant science questions, determine the anchor prediction model corresponding to the relevant science question based on the question source information of the relevant science question, wherein the anchor prediction model corresponding to the relevant science question is an anchor prediction model based on the question stem or an anchor prediction model based on the entire question; and predict the anchor data corresponding to the relevant science question based on the original question text of the relevant science question using the anchor prediction model corresponding to the relevant science question.

[0158] In some embodiments, the anchor data includes anchor label data and label confidence data. The second determination module 1420 is also used to determine P related science questions corresponding to the anchor prediction model based on the whole question based on the anchor prediction model corresponding to each of the N related science questions, where P is less than or equal to N; if P is a positive integer, based on the label confidence data corresponding to each of the P related science questions and the preset confidence conditions, select Q related science questions from the P related science questions, where the label confidence data corresponding to each of the Q related science questions meets the preset confidence conditions, and Q is less than or equal to P; based on the anchor label data corresponding to each of the Q related science questions, determine the anchor data corresponding to the target science question.

[0159] In some embodiments, the second determination module 1420 is also used to determine the number of relevant science questions with the highest label confidence for Q relevant science questions; if the number of relevant science questions with the highest label confidence is one, then based on the anchor label data corresponding to the relevant science question with the highest label confidence, the anchor data corresponding to the target science question is determined; if the number of relevant science questions with the highest label confidence is multiple, then for the anchor label data corresponding to each of the multiple relevant science questions with the highest label confidence, based on the anchor label data with the highest number of occurrences, the anchor data corresponding to the target science question is determined.

[0160] In some embodiments, the second determination module 1420 is further used to, if P is zero, determine the anchor data corresponding to the target science question based on the anchor label data with the highest number of occurrences for each of the N related science questions.

[0161] In some embodiments, the second determination module 1420 is further used to obtain a preset science question bank; and determine N relevant science questions based on the original question text of the target science question and the preset science question bank.

[0162] In some embodiments, the first determination module 1410 is further used to obtain a test question image corresponding to the target science test question; perform character recognition operations on the test question image to obtain the original test question text of the target science test question.

[0163] It can be understood that the modules in the above-mentioned model training device can also be placed in the test question processing device, that is, the test question processing device includes a generation module 1310 and a training module 1320. Alternatively, the existing modules in the test question processing device directly execute the steps to be executed by the modules in the model training device, such as the second determination module 1420 directly executes the steps to be executed by the generation module 1310 and the training module 1320. In other words, those skilled in the art can set it according to the actual situation, and the embodiments of the present disclosure do not make a unified limitation on this.

[0164] Fig.15 Shown is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. Fig.15 The electronic device 1500 shown (the electronic device 1500 may be a computer device) includes a memory 1501, a processor 1502, a communication interface 1503 and a bus 1504. The memory 1501, the processor 1502 and the communication interface 1503 are connected to each other through the bus 1504.

[0165] The memory 1501 may be a read-only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM). The memory 1501 may store a program. When the program stored in the memory 1501 is executed by the processor 1502, the processor 1502 and the communication interface 1503 are used to execute the various steps of the model training method and / or the test question processing method of the embodiment of the present disclosure.

[0166] Processor 1502 can adopt a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), a graphics processing unit (GPU) or one or more integrated circuits to execute relevant programs to implement the functions required to be performed by the units in the model training device and / or the test question processing device of the embodiment of the present disclosure.

[0167] The processor 1502 may also be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the model training method and / or the test question processing method disclosed in the present invention may be completed by an integrated logic circuit of hardware or software instructions in the processor 1502. The above-mentioned processor 1502 may also be a general-purpose processor, a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FieldProgrammable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in conjunction with the embodiments of the present invention may be directly embodied as being executed by a hardware decoding processor, or may be executed by a combination of hardware and software modules in a decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 1501, and the processor 1502 reads the information in the memory 1501, and combines its hardware to complete the functions required to be performed by the units included in the model training device and / or the test question processing device of the embodiment of the present disclosure, or executes the model training method and / or the test question processing method of the method embodiment of the present disclosure.

[0168] The communication interface 1503 uses a transceiver such as but not limited to a transceiver to implement communication between the electronic device 1500 and other devices or a communication network. For example, the original test text of the target science test can be obtained through the communication interface 1503.

[0169] The bus 1504 may include a path for transmitting information between various components of the electronic device 1500 (eg, the memory 1501 , the processor 1502 , and the communication interface 1503 ).

[0170] It should be understood that the training module 1320 in the model training device can be equivalent to the processor 1502.

[0171] It should be noted that although Fig.15 The electronic device 1500 shown only shows a memory, a processor, and a communication interface, but in the specific implementation process, those skilled in the art should understand that the electronic device 1500 also includes other devices necessary for normal operation. At the same time, according to specific needs, those skilled in the art should understand that the electronic device 1500 may also include hardware devices for implementing other additional functions. In addition, those skilled in the art should understand that the electronic device 1500 may also only include the devices necessary to implement the embodiments of the present disclosure, and does not necessarily include Fig.15 All devices shown in .

[0172] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this disclosure.

[0173] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0174] In the several embodiments provided in the present disclosure, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the 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, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0175] 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 solution of this embodiment.

[0176] In addition, each functional unit in each embodiment of the present disclosure 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.

[0177] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk.

[0178] The above is only a specific embodiment of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present disclosure, which should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for processing test questions, characterized in that: include: Determine the original test texts of the target science test questions; Determine the anchor point data corresponding to the target science test question based on the original test question text of the target science test question by using the anchor point prediction model, wherein the anchor point data is data associated with the examination content of the target science test question, and the anchor point prediction model is trained based on the respective analytic test question texts and anchor point label data of M science test question samples, where M is a positive integer; The method of using the anchor prediction model to determine the anchor data corresponding to the target science test question based on the original test question text of the target science test question includes: Based on the original test text of the target science test question, determine N related science test questions corresponding to the target science test question, where N is a positive integer; Determine the analytical question texts of the N relevant science questions; Using the anchor point prediction model, based on the analysis question texts of the N related science questions, predict the anchor point data corresponding to each of the N related science questions; Based on the anchor point data corresponding to each of the N related science test questions, the anchor point data corresponding to the target science test question is determined.

2. The test question processing method according to claim 1, characterized in that: Before determining the anchor point data corresponding to the target science test question based on the original test question text of the target science test question by using the anchor point prediction model, the method further includes: Based on the original test texts of the M science test question samples, generate the parsed test texts of the M science test question samples, wherein the original test question texts include the question stem texts; Based on the respective analyzed test question texts and anchor point label data of the M science test question samples, an initial network model is trained to obtain the anchor point prediction model.

3. The test question processing method according to claim 2, characterized in that: The initial network model corresponds to a full-word masking loss function, and the initial network model is trained based on the respective parsed test texts and anchor label data of the M science test samples to obtain the anchor prediction model, including: Based on the parsed test texts of the M science test samples and the full-word masking loss function, pre-training the initial network model to obtain a pre-trained model, wherein in the pre-trained model, the full-word masking loss function satisfies a preset convergence condition; Based on the analyzed test question texts and anchor point label data of each of the M science test question samples, the pre-trained model is trained to obtain the anchor point prediction model.

4. The test question processing method according to claim 2 or 3, characterized in that: The generating of the respective parsed test texts of the M science test samples based on the respective original test texts of the M science test samples comprises: For each science test question sample in the M science test question samples, Regularizing the original test text of the science test sample to obtain a regularized test text; The regularized test question text is processed by formula analysis to obtain the analyzed test question text of the science test question sample.

5. The test question processing method according to claim 1, characterized in that: The anchor prediction model includes an anchor prediction model based on the question stem and an anchor prediction model based on the whole question. The anchor prediction model is used to predict the anchor data corresponding to each of the N related science questions based on the respective analytic question texts of the N related science questions, including: For each of the N relevant science questions, Based on the question source information of the relevant science questions, determining the anchor prediction model corresponding to the relevant science questions, wherein the anchor prediction model corresponding to the relevant science questions is the anchor prediction model based on the question stem or the anchor prediction model based on the entire question; The analyzed test question text of the relevant science test question is input into the anchor point prediction model corresponding to the relevant science test question to obtain the anchor point data corresponding to the relevant science test question.

6. The test question processing method according to claim 5, characterized in that: The anchor point data includes anchor point label data and label confidence data, and the step of determining the anchor point data corresponding to the target science test question based on the anchor point data corresponding to each of the N related science test questions includes: Based on the anchor prediction models corresponding to the N relevant science test questions, determine P relevant science test questions corresponding to the anchor prediction model based on the full question, where P is less than or equal to N; If P is a positive integer, based on the label confidence data corresponding to each of the P related science questions and the preset confidence condition, Q related science questions are selected from the P related science questions, wherein the label confidence data corresponding to each of the Q related science questions satisfies the preset confidence condition, and Q is less than or equal to P; Determine the anchor point data corresponding to the target science question based on the anchor point label data corresponding to each of the Q related science questions; If P is zero, then for the anchor point label data corresponding to each of the N related science test questions, the anchor point data corresponding to the target science test question is determined based on the anchor point label data with the highest number of occurrences.

7. The test question processing method according to claim 6, characterized in that: The determining the anchor point data corresponding to the target science test question based on the anchor point label data corresponding to each of the Q related science test questions includes: For the Q relevant science questions, determine the number of relevant science questions with the highest label confidence; If the number of the relevant science test questions with the highest label confidence is one, then determining the anchor point data corresponding to the target science test question based on the anchor point label data corresponding to the relevant science test question with the highest label confidence; If there are multiple relevant science questions with the highest label confidence, then for the anchor label data corresponding to each of the multiple relevant science questions with the highest label confidence, the anchor data corresponding to the target science question is determined based on the anchor label data with the highest number of occurrences.

8. The test question processing method according to any one of claims 1 to 3, characterized in that: The original test text of the target science test questions includes: Obtaining a test question image corresponding to the target science test question; A character recognition operation is performed on the test question image to obtain the original test question text of the target science test question.

9. The test question processing method according to any one of claims 1 to 3, characterized in that: The M science test question samples include at least one of mathematics test question samples, physics test question samples and chemistry test question samples; and / or, the anchor point data includes anchor point label data, and the anchor point data also includes label confidence data; and / or, the original test question text also includes at least one of the knowledge point name, answer text and answer analysis text.

10. An anchor point prediction device, characterized in that: include: The first determination module is used to determine the original test text of the target science test question; A second determination module is configured to determine the anchor data corresponding to the target science test question based on the original test question text of the target science test question by using an anchor prediction model, wherein the anchor data is data associated with the examination content of the target science test question, and the anchor prediction model is trained based on the respective analytic test question texts and anchor label data of M science test question samples, where M is a positive integer; Wherein, the second determining module is specifically used for: Based on the original test text of the target science test question, determine N related science test questions corresponding to the target science test question, where N is a positive integer; Determine the analytical question texts of the N relevant science questions; Using the anchor point prediction model, based on the analysis question texts of the N related science questions, predict the anchor point data corresponding to each of the N related science questions; Based on the anchor point data corresponding to each of the N related science test questions, the anchor point data corresponding to the target science test question is determined.

11. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor, The processor is used to execute the method described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is used to execute the method according to any one of claims 1 to 9.

Images