Test question resource recommendation method and device, electronic device, and storage medium

Abstract

The application provides a test question resource recommendation method and device, electronic equipment and a storage medium. The method comprises the following steps: obtaining test question information of a target test question; performing feature coding on the target test question according to the test question information to obtain test question features of the target test question; determining a recommended test question matched with the target test question from all candidate test questions in a resource library according to the test question features and the test question information of the target test question and the test question features and the test question information of each candidate test question in the resource library; and performing test question resource recommendation on the target test question according to the recommended test question. The application can automatically complete the recommendation process while avoiding the influence of manual participation on the reliability of test question recommendation, thereby saving time and effort, and helping to improve the accuracy, reliability, rationality and recommendation efficiency of test question resource recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for recommending test question resources. Background Technology

[0002] To effectively alleviate the excessive homework and after-school tutoring burden on students in compulsory education, test question resource recommendations allow students to systematically learn and practice targeted exercises on learning devices, receiving feedback and explanations. The system also provides relevant explanations of the knowledge points covered in the questions and recommends similar questions with video explanations, enabling students to learn by analogy and master a type of question by solving one problem, thus consolidating and improving their understanding and mastery of related knowledge points. Therefore, accurately recommending similar test questions is crucial for improving students' learning efficiency.

[0003] Existing technologies typically require teaching and research teachers to manually select and recommend resources, which heavily relies on the professional capabilities of these teachers. The accuracy of the recommended test questions cannot be guaranteed, and all of these methods consume a significant amount of manpower and time, resulting in low efficiency. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for recommending test question resources, which solves the shortcomings of existing technologies that rely on teachers to manually select and recommend resources, resulting in unreliable accuracy and low efficiency in test question recommendations.

[0005] This invention provides a method for recommending test question resources, comprising:

[0006] Obtain the question information for the target test questions;

[0007] Based on the test question information, feature encoding is performed on the target test question to obtain the test question features of the target test question;

[0008] Based on the characteristics and information of the target test question, as well as the characteristics and information of each candidate test question in the resource library, a recommended test question that matches the target test question is determined from all candidate test questions in the resource library.

[0009] Based on the recommended test questions, test question resources are recommended for the target test questions.

[0010] According to a test question resource recommendation method provided by the present invention, the step of performing feature encoding on the target test question based on the test question information to obtain the test question features of the target test question includes:

[0011] Based on the test information of the target test question, determine the feature coding model corresponding to the target test question;

[0012] Based on the feature encoding model, the target test question is feature-encoded to obtain the test question features of the target test question.

[0013] According to a test question resource recommendation method provided by the present invention, the step of determining the feature coding model corresponding to the target test question based on the test question information of the target test question includes:

[0014] If the question information of the target question does not contain answer information and answer analysis information, the first feature encoding model shall be used as the feature encoding model corresponding to the target question.

[0015] The first feature encoding model is trained based on contrastive loss; the contrastive loss is obtained by contrastive learning training on a first sample test question that does not contain answer information and answer explanation information, and a second sample test question that includes answer information and answer explanation information.

[0016] According to a test question resource recommendation method provided by the present invention, the step of determining the feature coding model corresponding to the target test question based on the test question information of the target test question includes:

[0017] If the question information of the target question includes answer information and answer analysis information, the second feature encoding model is used as the feature encoding model corresponding to the target question.

[0018] The second feature encoding model is obtained by training based on error loss; the error loss is obtained by supervised training of the second sample test questions, which include answer information and answer analysis information, and the knowledge category labels corresponding to the second sample test questions.

[0019] According to a test question resource recommendation method provided by the present invention, the step of determining recommended test questions that match the target test question from all candidate test questions in the resource library based on the test question characteristics and test question information of the target test question, and the test question characteristics and test question information of each candidate test question in the resource library, includes:

[0020] Based on the characteristics of the target test question and the characteristics of each candidate test question, calculate the first similarity between the target test question and each candidate test question;

[0021] Based on the first similarity, a first target number of target candidate questions are selected from all candidate questions in the resource library;

[0022] Based on the question information of the target question and the question information of each target candidate question, calculate the second similarity between the target question and each target candidate question;

[0023] Based on the first similarity and the second similarity, recommended test questions that match the target test questions are determined from the target number of target candidate test questions.

[0024] According to a test question resource recommendation method provided by the present invention, the step of determining recommended test questions that match the target test questions from a first target number of target candidate test questions based on a first similarity and a second similarity includes:

[0025] The first similarity and the second similarity are fused together;

[0026] Based on the fusion results, a second target number of candidate test questions are selected from the first target number of target candidate test questions; the first target number is greater than the second target number.

[0027] Based on the candidate test questions to be recommended, the recommended test questions are determined.

[0028] According to a test question resource recommendation method provided by the present invention, the step of calculating a second similarity between the target test question and each target candidate test question based on the test question information of the target test question and the test question information of each target candidate test question includes:

[0029] The first knowledge point information, the first question difficulty information, and the first question type information are extracted from the question information of the target question; the second knowledge point information, the second question difficulty information, and the second question type information are extracted from the question information of each target candidate question.

[0030] Calculate the knowledge point similarity between the first knowledge point information of the target test question and the second knowledge point information of each target candidate test question;

[0031] Calculate the difficulty similarity between the first difficulty information of the target test question and the second difficulty information of each target candidate test question;

[0032] Calculate the type similarity between the first question type information of the target question and the second question type information of each target candidate question;

[0033] The similarity of knowledge points, the similarity of difficulty, and the similarity of type are fused to obtain the second similarity.

[0034] According to a test question resource recommendation method provided by the present invention, the resource library is pre-constructed based on the following steps:

[0035] For the current resource build, determine whether the resource coverage in the resource library obtained from the previous resource build is greater than a preset threshold;

[0036] If the resource coverage rate is determined to be less than or equal to the preset threshold, select multiple initial test questions from the database under the target chapter.

[0037] Based on the subject keyword database, data preprocessing is performed on each initial test question to obtain test questions to be screened; the data preprocessing includes data cleaning and data normalization.

[0038] Cluster multiple test questions to be screened, and based on the clustering results, identify key test questions among the multiple test questions to be screened that are less than a preset distance from the center cluster of each cluster and have complete test question information.

[0039] The key test questions are merged into the resource library obtained in the previous resource construction to obtain the resource library corresponding to the current resource construction;

[0040] The resource construction steps are executed iteratively until the resource coverage in the resource library is greater than the preset threshold.

[0041] According to a test question resource recommendation method provided by the present invention, the step of clustering multiple test questions to be screened includes:

[0042] Feature extraction is performed on the question stem and question option information of each question to be screened to obtain the first question feature of each question to be screened;

[0043] Feature extraction is performed on the question stem, question options, answer information, and answer explanation information of each question to be screened to obtain the second question features of each question to be screened;

[0044] Based on the first and second question features, the multiple questions to be screened are clustered.

[0045] The present invention also provides a test question resource recommendation device, comprising:

[0046] The acquisition unit is used to acquire the test question information of the target test question;

[0047] The encoding unit is used to perform feature encoding on the target test question based on the test question information to obtain the test question features of the target test question;

[0048] The matching unit is used to determine recommended questions that match the target question from all candidate questions in the resource library based on the question features and question information of the target question, as well as the question features and question information of each candidate question in the resource library;

[0049] The recommendation unit is used to recommend test question resources for the target test question based on the recommended test questions.

[0050] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the test question resource recommendation method as described above.

[0051] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the test question resource recommendation method as described above.

[0052] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the test question resource recommendation method as described above.

[0053] The test question resource recommendation method, apparatus, electronic device, and storage medium provided by this invention automatically encode the target test question based on its test question information to deeply mine test question features that represent the location and context information of the target test question. Then, by integrating the test question features and information of the target test question with the test question features and information of each candidate test question in the resource library, the invention automatically and accurately mines recommended test questions that match the target test question. The recommendation process requires no manual intervention, avoiding the impact of manual intervention on the reliability of test question recommendations. The fully automated recommendation process saves time and effort, and helps improve the accuracy, reliability, rationality, and efficiency of test question resource recommendations. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the existing technology-based method for recommending test question resources.

[0056] Figure 2 This is one of the flowcharts illustrating the test question resource recommendation method provided by the present invention;

[0057] Figure 3 This is a schematic diagram of the structure of the first initial feature encoding model in the test question resource recommendation method provided by the present invention;

[0058] Figure 4 This is a schematic diagram of the structure of the first feature encoding model in the test question resource recommendation method provided by the present invention;

[0059] Figure 5This is a schematic diagram of the structure of the second feature encoding model in the test question resource recommendation method provided by the present invention;

[0060] Figure 6 This is the second flowchart of the test question resource recommendation method provided by the present invention;

[0061] Figure 7 This is one of the flowcharts illustrating the process of constructing a resource library in the test question resource recommendation method provided by this invention;

[0062] Figure 8 This is the second flowchart illustrating the process of constructing a resource library in the test question resource recommendation method provided by this invention;

[0063] Figure 9 This is a schematic diagram of the test question resource recommendation device provided by the present invention;

[0064] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0066] like Figure 1 In existing technologies, it is common practice for teaching and research teachers to manually select recommended resources. Specifically, this involves manually selecting typical questions from a particular chapter and representing the question vectors based on the keywords appearing in the typical questions to obtain the TFTF-IDF (Term Frequency-Inverse Document Frequency) features of the typical questions. These TFTF-IDF features are then manually stored in a database. After obtaining the target question online, the TFTF-IDF features of the target question are represented, and questions with similar TFTF-IDF features to the target question are searched in the resource library and ranked to recommend multiple similar questions with the highest rankings based on business needs.

[0067] In summary, the existing test question recommendation methods have the following drawbacks:

[0068] First, the coverage of manually selected questions is too low. Relying on teaching and research teachers to manually select typical questions from a large number of test questions as recommended resources is highly dependent on human resources and has considerable subjectivity. The coverage of the selected questions is difficult to guarantee, which in turn makes it impossible to guarantee the accuracy of the test question recommendations.

[0069] Secondly, the test item representation method cannot express deep semantic information. TF-IDF mainly relies on keywords to represent feature vectors, which cannot capture deep semantic information or represent the differences in the order of different keywords. This results in a low accuracy rate of recommended similar test items and requires a lot of manpower and time, making it inefficient.

[0070] In summary, existing test question recommendation methods suffer from drawbacks such as unreliable accuracy and low efficiency.

[0071] To address the above problems, this embodiment provides a method for recommending test question resources. For example... Figure 2 The diagram shown is one of the flowcharts of the test question resource recommendation method provided in this embodiment. The method includes the following steps:

[0072] Step 201: Obtain the question information for the target question;

[0073] The target test question refers to the test question that needs to be pushed as a test question resource. It can be test question text, test question image or test question audio, etc. This implementation does not make specific limitations on this.

[0074] The test question information includes, but is not limited to, first sub-information and second sub-information; the first sub-information is used to represent the question information, including but not limited to one or more of the following: question stem information, question option information, answer information, and answer explanation information; the second sub-information is used to represent the knowledge information of the test question, including but not limited to one or more of the following: knowledge point information, question difficulty information, and question type information. This embodiment does not specifically limit this.

[0075] It should be noted that after obtaining the test information of the target test questions, the test information can be preprocessed, including but not limited to data cleaning and data standardization based on the subject keyword database, in order to improve the quality of the data information and facilitate subsequent test question feature extraction and test question recommendation, so as to achieve test question resource recommendation more quickly and accurately.

[0076] Step 202: Based on the test question information, perform feature encoding on the target test question to obtain the test question features of the target test question;

[0077] Optionally, in order to make accurate resource recommendations for target test questions, feature encoding can be performed on the target test questions based on the test question information, thereby deeply mining the test question features that represent the location information and context information of the target test questions.

[0078] Here, the extraction of test item features can be achieved through one or more pre-trained feature encoding models, or through one or more pre-configured feature encoding rules. When multiple pre-trained feature encoding models are configured for different test item information, the corresponding feature encoding model can be determined first based on the information content of the test item information, so as to achieve the extraction of test item features of the target test item according to the corresponding feature encoding model; when only one pre-trained feature encoding model is configured for different test item information, the extraction of test item features of the target test item can be achieved directly according to the feature encoding model. This embodiment does not specifically limit this.

[0079] Step 203: Based on the question characteristics and question information of the target question, and the question characteristics and question information of each candidate question in the resource library, determine the recommended question that matches the target question from all candidate questions in the resource library;

[0080] The resource library can be automatically constructed in advance by the recommendation device based on resource recommendation requirements. The resource library contains multiple candidate questions. These candidate questions can be questions adapted to the teaching resources for the target question. The candidate questions can cover all questions required by the resource recommendation requirements, or they can be typical questions selected from all questions in the required chapters. For example, based on the required chapters of the teaching resources, all questions in that chapter can be located, and then questions that meet the recommendation requirements of that teaching resource can be selected as candidate questions.

[0081] Here, the question information for each candidate question in the resource library may also include first sub-information and second sub-information.

[0082] Feature extraction for each candidate question can involve deep mining of features representing the location and context information of each candidate question. Similar to feature extraction for the target question, feature extraction for each candidate question can be achieved using one or more pre-trained feature encoding models or one or more pre-configured feature encoding rules. The feature extraction for both candidate and target questions can be achieved using the same feature encoding model or rules, or different feature encoding models and rules; this embodiment of the invention does not impose specific limitations on this approach.

[0083] It should be noted that the extraction of test feature of candidate test questions can be performed after the test feature of the target test question is determined, or the feature extraction can be performed on all candidate test questions before that. Only after the test feature of the target test question is determined can the test feature of each candidate test question be retrieved.

[0084] Optionally, after obtaining the question features and information of the target question, as well as the question features and information of each candidate question in the resource database, the question features and information can be treated as a whole feature to calculate the overall matching degree between the target question and each candidate question. Alternatively, a first matching degree can be calculated based on the question features, and a second matching degree can be calculated based on the question information. The first and second matching degrees can then be merged to obtain the overall matching degree between the target question and each candidate question. Therefore, based on the overall matching degree between the target question and each candidate question, the candidate question that best matches the target question can be selected from all candidate questions as the recommended question.

[0085] Here, there can be one or more recommended test questions. The method for selecting recommended test questions based on the degree of feature matching can be as follows: sort the overall matching degree between the target test question and each candidate test question in descending order, and select a preset number of candidate test questions with the highest feature matching degree as recommended test questions. The preset number here can be a pre-set 3 or 4. Alternatively, it can be that, based on the first matching degree, select the first preset number of candidate test questions with the highest first feature matching degree in descending order as target candidate test questions. Then, sort the overall matching degree between the target test question and each candidate test question in descending order, and select a second preset number of target candidate test questions with the highest overall matching degree as recommended test questions. Here, the first preset number is greater than the second preset number.

[0086] Step 204: Based on the recommended test questions, recommend test question resources for the target test questions.

[0087] Optionally, after obtaining the recommended questions for the target test, the system can accurately recommend more recommended questions to the user, so that the user can improve their learning efficiency by studying the recommended questions.

[0088] The test question resource recommendation method provided in this embodiment automatically encodes the target test question based on its test question information to deeply mine test question features that represent the location and context information of the target test question. Then, by combining the test question features and information of the target test question with the test question features and information of each candidate test question in the resource library, it automatically and accurately mines recommended test questions that match the target test question. The recommendation process does not require manual intervention. While avoiding the impact of manual intervention on the reliability of test question recommendations, the recommendation process is completed fully automatically, saving time and effort, and helping to improve the accuracy, reliability, rationality and efficiency of test question resource recommendations.

[0089] In some embodiments, feature encoding is performed on the target test question based on the test question information to obtain the test question features of the target test question, including:

[0090] Based on the test information of the target test question, determine the feature coding model corresponding to the target test question;

[0091] Based on the feature encoding model, the target test question is feature-encoded to obtain the test question features of the target test question.

[0092] Optionally, to ensure that question features representing the location and context information of the target question can be deeply extracted, multiple pre-trained feature encoding models can be configured for different question information. For example, different feature encoding models can be configured based on whether the question information contains answer information and answer parsing information.

[0093] For the extraction of test feature of target test questions, the information content contained in the test information of the target test questions is first determined. Based on the information content contained in the test information, an appropriate feature coding model is selected to extract the test feature of the target test questions. This enables targeted feature coding for target test questions under different test information, thereby making the test feature obtained by feature coding more representative of the test question characteristics and further improving the accuracy of test question recommendation.

[0094] Here, the feature extraction for the target test question and the feature extraction for the candidate test questions can be the same feature coding model or different feature coding models. Specifically, it is adaptively determined based on the test information of the target test question and the test information of the candidate test questions. This embodiment does not make specific limitations on this.

[0095] In this embodiment, for different test question information, the corresponding feature coding model is adaptively selected to classify and code the target test questions. This improves the efficiency of feature coding, and because the feature coding model is more targeted, it can also improve the accuracy of feature coding, uncover more representative question features, and further improve the accuracy of test question recommendation.

[0096] In some embodiments, if the question information of the target question does not include answer information and answer analysis information, the first feature encoding model is used as the feature encoding model corresponding to the target question.

[0097] The first feature encoding model is trained based on contrastive loss; the contrastive loss is obtained by contrastive learning training on a first sample test question that does not contain answer information and answer explanation information, and a second sample test question that includes answer information and answer explanation information.

[0098] Optionally, if the test information of the target test question is analyzed and it is determined that the test information of the target test question does not contain answer information and answer explanation information, in order to learn the features of answer information and answer explanation information during the feature encoding process, so that the test question features contain richer features and improve the test question recommendation accuracy, the first feature encoding model obtained by comparative learning training can be used as the feature encoding model corresponding to the target test question.

[0099] Furthermore, before extracting test item features based on the first feature encoding model, the first feature encoding model needs to be pre-trained. The specific training steps include:

[0100] First, a large amount of sample data is collected in advance. The sample data here includes the first sample test questions and the second sample test questions. The test information of the first sample test questions includes the question stem and the test option information, but does not include the answer information and the answer explanation information. The test information of the second sample test questions includes both the question stem and the test option information, as well as the answer information and the answer explanation information.

[0101] During the training process, a batch of sample questions can be randomly selected from a large amount of sample data, including first sample questions and second sample questions. This embodiment of the invention does not limit the proportion of first sample questions and second sample questions in the sample questions.

[0102] From the same batch of first and second sample test questions, one first sample test question and one second sample test question are randomly selected to form a sample test question pair. Sample test question pairs in which the stem information of the first sample test question matches the answer information and answer explanation information of the second sample test question are marked as positive sample test pairs. Sample test question pairs in which the stem information of the first sample test question does not match the answer information and answer explanation information of the second sample test question are marked as negative sample test pairs, thus obtaining contrastive learning labels.

[0103] Furthermore, a first initial feature encoding model is created. For example... Figure 3 As shown, the first initial feature encoding model here includes a feature extraction branch model corresponding to the first sample test question and a feature extraction branch model corresponding to the second sample test question, a concatenation layer, and a classification layer; both branch models include a BERT module and / or an attention encoding module, and may also include a pooling layer, etc., but this embodiment does not specifically limit this.

[0104] Subsequently, by inputting the question information of the first sample question and the question information of the second sample question in the sample question pair into the corresponding feature extraction branch models in the first initial feature encoding model, the question features U of the first sample question and the question features V of the second sample question are obtained. Then, the question features of the first sample question and the question features of the second sample question are concatenated and subtracted. Then, the distance between the question features of the first sample question and the question features of the second sample question is calculated through the classification layer to obtain the comparative learning result between the first sample question and the second sample question.

[0105] The contrastive loss is determined based on the contrastive learning results and contrastive learning labels. The model signal corresponding to the contrastive loss is used to incentivize the first initial feature encoding model to narrow the distance between positive sample pairs and widen the distance between negative sample pairs. This means that the distance between the feature representations of sample questions with similar semantic content is closer in the representation space, while the distance between the feature representations of sample questions with dissimilar semantic content is greater. For a sample sample pair (x1, x2), its corresponding label set is y ∈ {0, 1}, where y = 1 represents a positive sample pair and y = 0 represents a negative sample pair. Therefore, the formula for calculating the contrastive loss is:

[0106] L((x1,x2),y)=yL p (D(f θ (x1),f θ (x2)))+(1-y)L n (D(f θ (x1),f θ (x2)));

[0107] Where L(·) is the contrastive loss, D(·) is the distance function, and L p (·) is a first-order increasing function, L n (·) is a first-order decreasing function, f θ (x1) and f θ (x2) represents the item features of the first and second sample items in the sample item pair, respectively. It should be noted that when the sample item pair is a positive sample pair, the greater the distance between the item features of the first and second sample items, the higher the contrast loss; conversely, when the sample item pair is a negative sample pair, the closer the distance between the item features of the first and second sample items, the higher the contrast loss.

[0108] Therefore, the first initial feature encoding model is trained using the contrastive loss described above, so as to construct the first feature encoding model based on the feature extraction branch model corresponding to the first sample test item in the trained first initial feature encoding model.

[0109] like Figure 4 The diagram shows the structure of the first feature encoding model obtained through contrastive learning training, which includes the Bert module, the attention encoding module, and the pooling layer.

[0110] After obtaining the first feature encoding model, the question information of the target question can be processed through the Bert module, attention encoding module, and pooling layer in the first feature encoding model to extract the question features that integrate answer information and answer analysis information.

[0111] Compared to existing technologies that only use question stem information, especially when answer explanations are missing or question stem information is limited, which can lead to difficulties in effectively recalling similar recommended questions, this embodiment inputs the question information of the target question into the first feature encoding model. This allows the first feature encoding model to integrate features containing answer information and answer explanation information into the question features of the target question that do not contain answer information and answer explanation information. This results in richer and more complete feature representations in the question features, thereby improving the accuracy of question recommendation.

[0112] In some embodiments, determining the feature coding model corresponding to the target test question based on the test question information of the target test question includes:

[0113] If the question information of the target question includes answer information and answer analysis information, the second feature encoding model is used as the feature encoding model corresponding to the target question.

[0114] The second feature encoding model is obtained by training based on error loss; the error loss is obtained by supervised training of the second sample test questions, which include answer information and answer parsing information, and the test question features corresponding to the second sample test questions.

[0115] Optionally, when the test information of the target test question includes answer information and answer explanation information, that is, when the test information of the target test question includes not only the stem information and the option information, but also the answer information and answer explanation information, then the second feature encoding model trained based on the error loss can be used to directly perform deep feature extraction on the stem information, option information, answer information and answer explanation information contained in the test information to obtain test features that can represent the position information and context information of the target test question.

[0116] Furthermore, before extracting test item features based on the second feature encoding model, the second feature encoding model needs to be pre-trained. The specific training steps include:

[0117] First, a large amount of sample data is collected in advance. This sample data includes a second sample of test questions, where the question information includes the question stem, answer choices, and answer explanations. Based on this, the knowledge category labels for the second sample of test questions are annotated. These knowledge category labels can be obtained through manual annotation or predicted using a pre-trained network model for knowledge category recognition.

[0118] Based on this, the second sample test question is first input into the second initial feature encoding model of the knowledge recognition model. The second initial feature encoding model extracts the encoded features of the second sample test question, obtaining its test question features. These features are then input into the recognition layer of the knowledge recognition model to obtain the predicted knowledge category of the second sample test question. The error loss between the predicted knowledge category and the knowledge category label is calculated, and the knowledge recognition model is trained based on this error loss, resulting in the trained knowledge recognition model. Finally, based on the second initial feature encoding model of the trained knowledge recognition model, a second feature encoding model is constructed.

[0119] like Figure 5 As shown, the second feature encoding model includes a BERT module and an attention-based encoding module; wherein, the BERT module, as a feature embedding layer, is used to map the input information of the second feature encoding model into a matrix of dimension R. m×n The vector representation of is used to represent the embedded features that have positional and contextual information and contain deep connections with each positional vector. The specific calculation formula is as follows:

[0120] E = Bert Emb (T),E∈R m×n ;

[0121] Where E is the embedded feature output by the Bert module, T is the input of the second feature encoding model, and m×n is the feature dimension of the embedded feature.

[0122] Subsequently, an attention-based encoding module is used to encode the embedded features. For long sequence data, RNN (Recurrent Neural Network) models often suffer from the vanishing gradient problem, causing the model to ignore the beginning of the long sequence data. However, the Transformer in the attention-based encoding module can solve this problem. The Transformer encoding layer outputs joint encoded information containing contextual information, i.e., the test item feature T. C The specific calculation formula for feature encoding is as follows:

[0123] T C=Transformer Encoder (E),T C ∈R m×h ;

[0124] Among them, T C The output of the coding module represents the test item features, where m×h is the feature dimension of the test item features.

[0125] Therefore, based on the constructed second feature encoding model, deep feature extraction is performed on the question stem information, question option information, answer information, and answer analysis information contained in the question information to obtain question features that can represent the positional information and context information of the target question. This makes the question features contain rich surface features, thereby improving the accuracy of question recommendation.

[0126] like Figure 6 The diagram shown is a second schematic flowchart of the test question resource recommendation method provided in this embodiment; in some embodiments, step 203 further includes:

[0127] Step 2031: Calculate the first similarity between the target test question and each candidate test question based on the test question characteristics of the target test question and the test question characteristics of each candidate test question;

[0128] Step 2032: Based on the first similarity, select a first target number of target candidate questions from all candidate questions in the resource library;

[0129] Step 2033: Calculate the second similarity between the target test question and each target candidate test question based on the test question information of the target test question and the test question information of each target candidate test question;

[0130] Step 2034: Based on the first similarity and the second similarity, determine recommended test questions that match the target test questions from the target candidate test questions of the first target number.

[0131] Optionally, after obtaining the test features and test information of the target test question, as well as the test features and test information of each candidate test question, the similarity between the test features of the target test question and the test features of each candidate test question can be calculated, thereby obtaining the first similarity between the target test question and each candidate test question.

[0132] The similarity can be calculated using formulas such as cosine distance and Pearson similarity. When using cosine distance to calculate similarity, the angle between the vectors representing the features of the target question and the features of each candidate question is considered. The result is the product of the inner product of the feature vectors (the sum of the product of corresponding elements) and the product of the magnitudes of the two vectors. The specific formula is as follows:

[0133]

[0134] Where similarity is the similarity score; θ is the vector angle between the feature of the target question and the feature of each candidate question, A i and B i These are the i-th sub-feature in the test features of the target test question and the i-th sub-feature in the test features of the candidate test question, respectively, where n is the number of sub-features in the test features of the target test question or the candidate test question.

[0135] Subsequently, based on the first similarity between the target question and each candidate question, ranked from highest to lowest, the top-ranked candidate questions in the resource library are coarsely recalled as target candidate questions, and then recalled sequentially from highest to lowest similarity. The first target number can be set according to recommendation requirements, such as 100.

[0136] For the calculation of the second similarity, information such as knowledge points, difficulty, and type can be obtained from the test question information. Then, the knowledge points, difficulty, and type information can be used as a whole as a feature vector, and the second similarity can be calculated according to the calculation method of the first similarity. Alternatively, the knowledge point similarity between the target test question and each target candidate test question can be calculated separately according to the calculation method of the first similarity, the difficulty similarity between the target test question and each target candidate test question can be calculated separately according to the knowledge point information, and the difficulty similarity between the target test question and each target candidate test question can be calculated separately according to the difficulty information, and the difficulty similarity between the target test question and each target candidate test question can be calculated separately according to the test type information. The second similarity can be calculated by combining the knowledge point similarity, difficulty similarity, and type similarity. This embodiment does not make specific limitations on this.

[0137] Subsequently, by combining the first similarity and the second similarity, a preset number or a preset proportion of target candidate questions are selected from the first target number of target candidate questions as recommended questions.

[0138] In this embodiment, by using the test question features extracted from the test question surface by an online deep learning-based feature encoding model, as well as the knowledge point information, test difficulty information, and test type information in the test question information, multiple and multi-dimensional recalls can be performed to accurately recommend the most similar test questions for the target test question, thereby improving the accuracy of test question recommendations.

[0139] In some embodiments, step 2034, which involves determining recommended test questions matching the target test questions from the first target number of target candidate test questions based on the first similarity and the second similarity, further includes:

[0140] The first similarity and the second similarity are fused together;

[0141] Based on the fusion results, a second target number of candidate test questions are selected from the first target number of target candidate test questions; the first target number is greater than the second target number.

[0142] Based on the candidate test questions to be recommended, the recommended test questions are determined.

[0143] Optionally, the first and second similarities are weighted and fused to obtain an overall similarity. Based on the overall similarity between the target question and each candidate target question, ranked from highest to lowest, the top-ranked second number of candidate target questions are selected as recommended questions, and then recalled sequentially from highest to lowest similarity. The second number of candidate target questions can be set according to recommendation requirements, but it must be less than the first number, such as 3 or 5.

[0144] The weights in the weighted fusion can be adaptively set according to the contribution of test feature and test information to the recommendation of test resource. The higher the contribution of the test feature, the greater the weight of the similarity. The lower the contribution of the test feature, the smaller the weight of the similarity.

[0145] In this embodiment, by recalling similar test questions multiple times and in multiple dimensions, the most similar recommended test questions can be accurately recommended for the target test question, thereby improving the accuracy of test question recommendations.

[0146] In some embodiments, calculating the second similarity between the target test question and each target candidate test question based on the test question information of the target test question and the test question information of each target candidate test question includes:

[0147] The first knowledge point information, the first question difficulty information, and the first question type information are extracted from the question information of the target question; the second knowledge point information, the second question difficulty information, and the second question type information are extracted from the question information of each target candidate question.

[0148] Calculate the knowledge point similarity between the first knowledge point information of the target test question and the second knowledge point information of each target candidate test question;

[0149] Calculate the difficulty similarity between the first difficulty information of the target test question and the second difficulty information of each target candidate test question;

[0150] Calculate the type similarity between the first question type information of the target question and the second question type information of each target candidate question;

[0151] The similarity of knowledge points, the similarity of difficulty, and the similarity of type are fused to obtain the second similarity.

[0152] Optionally, for the second similarity calculation, information such as knowledge points, difficulty levels, and types can be obtained from the question information of the target question and the question information of each target candidate question. Then, according to the calculation method of the first similarity, the knowledge point similarity between the target question and each target candidate question is calculated based on the knowledge point information, the difficulty similarity between the target question and each target candidate question is calculated based on the difficulty information, and the difficulty similarity between the target question and each target candidate question is calculated based on the question type information. The knowledge point similarity, difficulty similarity, and type similarity are weighted and fused to calculate the second similarity. This embodiment does not specifically limit this.

[0153] The weights of the weighted fusion of knowledge point similarity, difficulty similarity, and type similarity can also be adaptively set according to their contribution to the recommendation of test question resources. For example, the weights of the weighted fusion of the first similarity, knowledge point similarity, difficulty similarity, and type similarity are set to 0.3, 0.3, 0.2, and 0.2, respectively.

[0154] Then, the overall similarity is obtained by weighting and summing the first similarity, knowledge point similarity, difficulty similarity, and type similarity. Based on the overall similarity between the target question and each target candidate question from high to low, the second largest number of target candidate questions are selected as recommended questions for sequential recall.

[0155] As shown in Table 1, in order to further verify the effectiveness of the test question recommendation provided in this embodiment, various recall methods for recommending test questions were compared. The comparison shows that the recall method for recommending test questions based on the weighted fusion result of the first similarity, knowledge point similarity, difficulty similarity and type similarity provided in this embodiment has better accuracy in test question recommendation.

[0156] Table 1 Comparison of various recall methods used for recommended test items

[0157]

[0158] In this embodiment, multi-dimensional features are used to recall recommended test questions, which can accurately recommend the most similar test questions to the target test question, thereby further improving the accuracy of test question recommendations.

[0159] In some embodiments, the repository is pre-built based on the following steps:

[0160] For the current resource build, determine whether the resource coverage in the resource library obtained from the previous resource build is greater than a preset threshold;

[0161] If the resource coverage rate is determined to be less than or equal to the preset threshold, select multiple initial test questions from the database under the target chapter.

[0162] Based on the subject keyword database, data preprocessing is performed on each initial test question to obtain test questions to be screened; the data preprocessing includes data cleaning and data normalization.

[0163] Cluster multiple test questions to be screened, and based on the clustering results, identify key test questions among the multiple test questions to be screened that are less than a preset distance from the central cluster of each cluster and have complete test question information.

[0164] The key test questions are merged into the resource library obtained in the previous resource construction to obtain the resource library corresponding to the current resource construction;

[0165] The resource construction steps are executed iteratively until the resource coverage in the resource library is greater than the preset threshold.

[0166] It should be noted that the resource repository is constructed on a chapter-by-chapter basis. For example... Figure 7 The diagram shown is one of the flowcharts for building a resource repository. The specific methods for building the resource repository for each target chapter include:

[0167] Step 701: Verify the resource coverage of the target chapter obtained in the previous resource construction.

[0168] Step 702: Determine whether the resource coverage rate of the target chapter obtained in the previous resource construction is greater than a preset threshold. Specifically, obtain the test dataset of the target chapter from high-frequency teaching materials, and verify whether the similar question coverage of existing resources in the resource library based on the previous resource construction is greater than the preset threshold, that is, verify whether the similar question coverage of the current resource meets the standard determined by the business. If it is determined that the resource coverage rate is greater than the preset threshold, proceed to step 604; if it is determined that the resource coverage rate is less than or equal to the preset threshold, proceed to step 603.

[0169] Step 703: Select multiple initial questions from online user return questions and pre-stored questions in the database for re-clustering to improve the resource library. Specifically, based on the existing question resources in the resource library from the previous resource construction, verify whether there are similar questions in the online user return questions and pre-stored questions in the database that are similar to the existing question resources. Select questions under the target chapter that are not similar to the existing question resources as initial questions. Based on multiple initial questions, select resources and add them to the library, thereby realizing the construction of the resource library. The specific steps are as follows: Figure 8 As shown, it includes the following steps:

[0170] Step 7031: Select multiple questions from the target chapter that are not similar to existing question resources as initial questions;

[0171] Step 7032: Based on the subject keyword database, perform data preprocessing on each initial test question to obtain test questions to be screened. Specifically, based on the subject keywords in the subject keyword database, perform data cleaning and data normalization on each initial test question to ensure that the subject keywords are fully preserved in the test question information after preprocessing and are not lost due to preprocessing. Subject keywords include, but are not limited to, keywords such as angle, function, and linear equation in one variable.

[0172] Step 7033: Extract the coding features of each question to be screened to obtain the feature vector of each question;

[0173] Step 7034: Cluster the multiple questions to be screened based on their feature vectors to form multiple clusters. Among the multiple questions to be screened, identify key questions whose distance to the centroid of each cluster is less than a preset distance and whose question information is complete. Complete question information means that the question includes at least knowledge point information, question difficulty information, and question type information. The preset distance can be set according to actual needs.

[0174] Step 7035: Merge the key test questions into the resource library obtained in the previous resource construction to obtain the resource library corresponding to the current resource construction; and return to step 701 to continue the resource library construction steps until the resource coverage in the resource library is greater than the preset threshold, and return to step 704.

[0175] Step 704: The deployment of test question resources under the target chapter is complete.

[0176] In this embodiment, during the construction of the resource library, a subject keyword library is also constructed to ensure that subject information is not lost during the preprocessing stage. The resources are clustered by the semantic feature vectors of each initial test question, and the initial test questions with complete information near the center point are selected and recorded into videos to form the resource library. This effectively avoids the need for manual experience in constructing the resource library. While improving the construction efficiency of the resource library, it can also ensure that the resource library has good coverage of test question resources, thereby helping to improve the efficiency and accuracy of test question recommendation.

[0177] In some embodiments, the step of clustering multiple test questions to be screened further includes:

[0178] Feature extraction is performed on the question stem and question option information of each question to be screened to obtain the first question feature of each question to be screened;

[0179] Feature extraction is performed on the question stem, question options, answer information, and answer explanation information of each question to be screened to obtain the second question features of each question to be screened;

[0180] Based on the first and second question features, the multiple questions to be screened are clustered.

[0181] For question clustering, other clustering algorithms such as K-means clustering can be used.

[0182] Optionally, considering that some test questions may not include answer and answer explanation information during the resource database creation process and in practical applications, to increase the resource coverage of the resource database, feature extraction can be performed on the question stem and option information of each test question to obtain the first test question feature. Feature extraction can then be performed on the question stem, option information, answer information, and answer explanation information of each test question to obtain the second test question feature. Multiple test questions to be screened can then be clustered based on the first and second test question features to achieve accurate clustering of the test questions to be screened, thereby selecting more representative key test questions to construct the resource database.

[0183] The test question resource recommendation device provided by the present invention is described below. The test question resource recommendation device described below can be referred to in correspondence with the test question resource recommendation method described above.

[0184] like Figure 9 As shown, this embodiment provides a test question resource recommendation device, which includes:

[0185] Unit 901 is used to obtain the test question information of the target test question;

[0186] The encoding unit 902 is used to perform feature encoding on the target test question based on the test question information to obtain the test question features of the target test question;

[0187] The matching unit 903 is used to determine recommended questions that match the target question from all candidate questions in the resource library based on the question features and question information of the target question, as well as the question features and question information of each candidate question in the resource library;

[0188] The recommendation unit 904 is used to recommend test question resources for the target test question based on the recommended test questions.

[0189] The test question resource recommendation device provided in this embodiment automatically encodes the target test question based on its test question information to deeply mine test question features that represent the location and context information of the target test question. Then, by combining the test question features and information of the target test question with the test question features and information of each candidate test question in the resource library, it automatically and accurately mines recommended test questions that match the target test question. The recommendation process does not require manual intervention. While avoiding the impact of manual intervention on the reliability of test question recommendations, the recommendation process is completed fully automatically, saving time and effort, and helping to improve the accuracy, reliability, rationality and efficiency of test question resource recommendations.

[0190] In some embodiments, the encoding unit 902 is specifically used for:

[0191] Based on the test information of the target test question, determine the feature coding model corresponding to the target test question; based on the feature coding model, perform feature coding on the target test question to obtain the test question features of the target test question.

[0192] In some embodiments, the encoding unit 902 is further configured to:

[0193] If the target test question does not contain answer information and answer explanation information, the first feature encoding model is used as the feature encoding model corresponding to the target test question. The first feature encoding model is trained based on contrastive loss. The contrastive loss is obtained by contrastive learning training on a first sample test question that does not contain answer information and answer explanation information, and a second sample test question that includes answer information and answer explanation information.

[0194] In some embodiments, the encoding unit 902 is further configured to:

[0195] When the question information of the target question includes answer information and answer explanation information, the second feature encoding model is used as the feature encoding model corresponding to the target question; the second feature encoding model is obtained by training based on error loss; the error loss is obtained by supervised training on a second sample question including answer information and answer explanation information and the knowledge category label corresponding to the second sample question.

[0196] In some embodiments, the matching unit 903 is specifically used for:

[0197] Based on the characteristics of the target test question and the characteristics of each candidate test question, a first similarity is calculated between the target test question and each candidate test question; based on the first similarity, a first target number of target candidate test questions are selected from all candidate test questions in the resource library; based on the test information of the target test question and the test information of each target candidate test question, a second similarity is calculated between the target test question and each target candidate test question; based on the first similarity and the second similarity, recommended test questions that match the target test question are determined from the first target number of target candidate test questions.

[0198] In some embodiments, the matching unit 903 is further configured to:

[0199] The first similarity and the second similarity are fused; based on the fusion result, a second target number of candidate questions to be recommended are selected from the first target number of candidate questions; the first target number is greater than the second target number; based on the candidate questions to be recommended, the recommended questions are determined.

[0200] In some embodiments, the matching unit 903 is further configured to:

[0201] The first knowledge point information, the first question difficulty information, and the first question type information are extracted from the question information of the target question. The second knowledge point information, the second question difficulty information, and the second question type information are extracted from the question information of each target candidate question. The knowledge point similarity between the first knowledge point information of the target question and the second knowledge point information of each target candidate question is calculated. The difficulty similarity between the first question difficulty information of the target question and the second question difficulty information of each target candidate question is calculated. The type similarity between the first question type information of the target question and the second question type information of each target candidate question is calculated. The knowledge point similarity, the difficulty similarity, and the type similarity are fused to obtain the second similarity.

[0202] In some embodiments, the apparatus further includes a building unit, specifically used for:

[0203] For the current resource construction, determine whether the resource coverage rate in the resource library obtained from the previous resource construction is greater than a preset threshold; if the resource coverage rate is less than or equal to the preset threshold, select multiple initial test questions under the target chapter from the database; perform data preprocessing on each initial test question according to the subject keyword library to obtain test questions to be screened; the data preprocessing includes data cleaning and data normalization; cluster the multiple test questions to be screened, and based on the clustering results, determine the key test questions among the multiple test questions to be screened that are less than a preset distance from the centroid of each cluster and have complete test question information; merge the key test questions into the resource library obtained from the previous resource construction to obtain the resource library corresponding to the current resource construction; iteratively execute the resource construction steps until the resource coverage rate in the resource library is greater than the preset threshold.

[0204] In some embodiments, the building unit is further configured to:

[0205] Feature extraction is performed on the question stem and question option information of each question to be screened to obtain the first question feature of each question to be screened;

[0206] Feature extraction is performed on the question stem, question options, answer information, and answer explanation information of each question to be screened to obtain the second question features of each question to be screened;

[0207] Based on the first and second question features, the multiple questions to be screened are clustered.

[0208] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10 As shown, the electronic device may include a processor 1001, a communications interface 1002, a memory 1003, and a communication bus 1004, wherein the processor 1001, the communications interface 1002, and the memory 1003 communicate with each other via the communication bus 1004. The processor 1001 can call logical instructions in the memory 1003 to execute a test question resource recommendation method. This method includes: acquiring test question information of a target test question; performing feature encoding on the target test question based on the test question information to obtain test question features of the target test question; determining recommended test questions matching the target test question from all candidate test questions in the resource library based on the test question features and test question information of the target test question, as well as the test question features and test question information of each candidate test question in the resource library; and recommending test question resources for the target test question based on the recommended test questions.

[0209] Furthermore, the logical instructions in the aforementioned memory 1003 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the test question resource recommendation method provided by the above methods. The method includes: obtaining test question information of a target test question; performing feature encoding on the target test question based on the test question information to obtain test question features of the target test question; determining recommended test questions that match the target test question from all candidate test questions in the resource library based on the test question features and test question information of the target test question, as well as the test question features and test question information of each candidate test question in the resource library; and recommending test question resources for the target test question based on the recommended test questions.

[0211] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a test question resource recommendation method provided by the methods described above. The method includes: acquiring test question information of a target test question; performing feature encoding on the target test question based on the test question information to obtain test question features of the target test question; determining recommended test questions matching the target test question from all candidate test questions in the resource library based on the test question features and test question information of the target test question, as well as the test question features and test question information of each candidate test question in the resource library; and recommending test question resources for the target test question based on the recommended test questions.

[0212] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0213] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recommending test question resources, characterized in that, include: Obtain the question information for the target test questions; Based on the test question information, feature encoding is performed on the target test question to obtain the test question features of the target test question; Based on the characteristics and information of the target test question, and the characteristics and information of each candidate test question in the resource library, a recommended test question that matches the target test question is determined from all candidate test questions in the resource library; Based on the recommended test questions, test question resources are recommended for the target test questions; The steps for obtaining the test item features of the target test item include: When the target test question does not contain answer information and answer explanation information, the target test question is feature-encoded based on a first feature encoding model to obtain the test question features. The first feature encoding model is used to fuse the answer information and answer explanation information into the test question features of the target test question. The first feature encoding model is obtained by training a first initial feature encoding model based on contrastive loss. The contrastive loss is used to shorten the distance between positive sample pairs and widen the distance between negative sample pairs. The stem information of the first sample test question in the positive sample pair matches the answer information and answer explanation information of the second sample test question. The stem information of the first sample test question in the negative sample pair does not match the answer information and answer explanation information of the second sample test question. The first sample test question does not contain answer information and answer explanation information, while the second sample test question contains answer information and answer explanation information. The first initial feature encoding model includes feature extraction branch models corresponding to the first sample test question and the second sample test question, respectively; the first feature encoding model is constructed based on the feature extraction branch model corresponding to the first sample test question in the trained first initial feature encoding model.

2. The test question resource recommendation method according to claim 1, characterized in that, The step of encoding the target test question based on the test question information to obtain the test question features of the target test question includes: Based on the test information of the target test question, determine the feature coding model corresponding to the target test question; Based on the feature encoding model, the target test question is feature-encoded to obtain the test question features of the target test question.

3. The test question resource recommendation method according to claim 2, characterized in that, The step of determining the feature coding model corresponding to the target test question based on the test question information of the target test question includes: If the question information of the target question includes answer information and answer analysis information, the second feature encoding model is used as the feature encoding model corresponding to the target question. The second feature encoding model is obtained by training based on error loss; the error loss is obtained by supervised training of the second sample test questions, which include answer information and answer analysis information, and the knowledge category labels corresponding to the second sample test questions.

4. The test question resource recommendation method according to any one of claims 1-3, characterized in that, The step of determining recommended test questions that match the target test question from all candidate test questions in the resource library, based on the test question characteristics and test question information of the target test question and the test question characteristics and test question information of each candidate test question in the resource library, includes: Based on the characteristics of the target test question and the characteristics of each candidate test question, calculate the first similarity between the target test question and each candidate test question; Based on the first similarity, a first target number of target candidate questions are selected from all candidate questions in the resource library; Based on the question information of the target question and the question information of each target candidate question, calculate the second similarity between the target question and each target candidate question; Based on the first similarity and the second similarity, recommended test questions that match the target test questions are determined from the target number of target candidate test questions.

5. The test question resource recommendation method according to claim 4, characterized in that, The step of determining recommended test questions that match the target test questions from the first target number of target candidate test questions based on the first similarity and the second similarity includes: The first similarity and the second similarity are fused together; Based on the fusion results, a second target number of candidate test questions are selected from the first target number of target candidate test questions; the first target number is greater than the second target number. Based on the candidate test questions to be recommended, the recommended test questions are determined.

6. The test question resource recommendation method according to claim 4, characterized in that, The step of calculating the second similarity between the target test question and each target candidate test question based on the test question information of the target test question and the test question information of each target candidate test question includes: The first knowledge point information, the first question difficulty information, and the first question type information are extracted from the question information of the target question; the second knowledge point information, the second question difficulty information, and the second question type information are extracted from the question information of each target candidate question. Calculate the knowledge point similarity between the first knowledge point information of the target test question and the second knowledge point information of each target candidate test question; Calculate the difficulty similarity between the first difficulty information of the target test question and the second difficulty information of each target candidate test question; Calculate the type similarity between the first question type information of the target question and the second question type information of each target candidate question; The similarity of knowledge points, the similarity of difficulty, and the similarity of type are fused to obtain the second similarity.

7. The test question resource recommendation method according to any one of claims 1-3, characterized in that, The resource repository is pre-built based on the following steps: For the current resource build, determine whether the resource coverage in the resource library obtained from the previous resource build is greater than a preset threshold; If the resource coverage rate is determined to be less than or equal to the preset threshold, select multiple initial test questions from the database under the target chapter. Based on the subject keyword database, data preprocessing is performed on each initial test question to obtain test questions to be screened; the data preprocessing includes data cleaning and data normalization. Cluster multiple test questions to be screened, and based on the clustering results, identify key test questions among the multiple test questions to be screened that are less than a preset distance from the central cluster of each cluster and have complete test question information. The key test questions are merged into the resource library obtained in the previous resource construction to obtain the resource library corresponding to the current resource construction; The resource construction steps are executed iteratively until the resource coverage in the resource library is greater than the preset threshold.

8. The test question resource recommendation method according to claim 7, characterized in that, The clustering of multiple test questions to be screened includes: Feature extraction is performed on the question stem and question option information of each question to be screened to obtain the first question feature of each question to be screened; Feature extraction is performed on the question stem, question options, answer information, and answer explanation information of each question to be screened to obtain the second question features of each question to be screened; Based on the first and second question features, the multiple questions to be screened are clustered.

9. A test question resource recommendation device, characterized in that, include: The acquisition unit is used to acquire the test question information of the target test question; The encoding unit is used to perform feature encoding on the target test question based on the test question information to obtain the test question features of the target test question; The matching unit is used to determine recommended questions that match the target question from all candidate questions in the resource library based on the question features and question information of the target question, as well as the question features and question information of each candidate question in the resource library; The recommendation unit is used to recommend test question resources for the target test question based on the recommended test questions; The encoding unit is specifically used for: If the target test question does not contain answer information and answer explanation information, the target test question is feature-encoded based on a first feature encoding model to obtain the test question features. The first feature encoding model is obtained by training a first initial feature encoding model based on contrastive loss. The contrastive loss is used to shorten the distance between positive sample pairs and widen the distance between negative sample pairs. The stem information of the first sample test question in the positive sample pair matches the answer information and answer explanation information of the second sample test question. The stem information of the first sample test question in the negative sample pair does not match the answer information and answer explanation information of the second sample test question. The first sample test question does not contain answer information and answer explanation information, while the second sample test question contains answer information and answer explanation information. The first feature encoding model is used to integrate answer information and answer analysis information into the question features of the target question; the first initial feature encoding model includes feature extraction branch models corresponding to the first sample question and the second sample question respectively; the first feature encoding model is constructed based on the feature extraction branch model corresponding to the first sample question in the trained first initial feature encoding model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the test question resource recommendation method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the test question resource recommendation method as described in any one of claims 1 to 8.

Images