A method and system for generating test knowledge points based on artificial intelligence

Abstract

The present invention relates to a method and system for generating test question knowledge points based on artificial intelligence, and belongs to the field of data processing technology. The method comprises the following steps: test question text data preprocessing, test question text data segmentation, test question data set division, model pretraining, test question text vectorization, and test question text label word annotation; the output word vector of the Bert model is input into the convolution layer of the convolutional neural network, different convolution kernels are used for convolution operation, the maximum value of the column vector obtained by convolution is extracted by the pooling layer, and the final label word annotation result is obtained by full connection layer and Softmax processing; test question knowledge point generation: the knowledge points contained in the test question are generated according to the label words of the test question through effective association rules. The present invention improves the generalization ability of the model through multi-granularity word segmentation; improves the test question knowledge points by mining the association relationship between label words and knowledge points, and sets weights for label words, thereby improving the effectiveness and accuracy of generating knowledge points.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a method and system for generating test question knowledge points based on artificial intelligence. Background Art

[0002] Educational informatization has brought significant changes to the field of education. To address the challenges of utilizing massive educational data resources, educational data mining technology has emerged. However, the heterogeneous, implicit, and incomparable nature of educational data resources makes data mining very difficult. The field of educational data mining has made significant progress in certain areas. For example, machine learning has been used to address the widespread lack of annotations for cognitive verbs and knowledge points in massive test questions.

[0003] Despite this, the unannotated phenomenon caused by the word segmentation granularity setting during the question stem annotation process has resulted in a large amount of resources still not being fully utilized. At the same time, the importance of knowledge points is not distinguished during the annotation process, which affects the effectiveness of knowledge point generation. In addition, the potential relationships between knowledge points are not explored during data mining, resulting in incomplete knowledge points. Therefore, it is urgent to propose an artificial intelligence-based test knowledge point generation method and system. By annotating knowledge points in the test stem and mining the associated relationships, it can generate complete test knowledge points, providing a reference for applications such as test question recommendation. Summary of the Invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a method and system for generating test knowledge points based on artificial intelligence, which improves the generalization ability of the model through multi-granularity word segmentation; improves the test knowledge points by mining the correlation between label words and knowledge points, and sets weights for label words, thereby improving the effectiveness and accuracy of generating knowledge points.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The present invention provides a method for generating test knowledge points based on artificial intelligence, comprising the following steps:

[0007] Test question text data preprocessing: Collect large-scale test question data, label knowledge points, and establish a test question dataset;

[0008] Test question text data segmentation: The test question text in the test set is segmented iteratively using different granularities to generate new text;

[0009] Test data set division: divide the test data set into training set, validation set and test set;

[0010] Model pre-training: Map labels to label words and complete BERT model pre-training through prompt learning pre-training;

[0011] Test question text vectorization: Each word in the new text generated by word segmentation is vectorized using the Bert model;

[0012] Test question text label word annotation: The output word vector of the Bert model is input into the convolutional layer of the convolutional neural network. Different convolution kernels are used for convolution operations. Then, the maximum value of the column vector obtained by convolution is extracted through the pooling layer. The final label word annotation result is obtained through the fully connected layer and softmax processing.

[0013] Test question knowledge point generation: Generate the knowledge points contained in the test question through effective association rules based on the test question's label words;

[0014] Model indicator evaluation: Model evaluation is performed by selecting precision, recall rate and F1 value as the evaluation criteria for model knowledge point generation.

[0015] Furthermore, in the test question text data preprocessing, the test question text data annotation process is completed in a crowdsourcing mode, including the following steps:

[0016] Screening crowdsourcing users: Obtain test question sets based on the application scope, and recruit users with corresponding grade and subject categories to complete the labeling of test question knowledge points, and further screen them based on academic qualifications and work experience;

[0017] Training crowdsourcing users: Develop and distribute procedures and specifications for labeling test questions and knowledge points, and train users to master labeling requirements;

[0018] Crowdsourcing user evaluation: Conduct training evaluation on users who have completed the training, and select those with excellent evaluation results as users of the crowdsourcing project;

[0019] Carry out labeling work: assign labeling tasks to each crowdsourcing user and obtain the test question label annotated text data;

[0020] Label noun standardization: Create a label set by annotating text data with test question labels, and uniformly replace knowledge point nouns with the same label meaning.

[0021] Furthermore, the word segmentation of the test question text data includes the following steps:

[0022] A coarse-grained word segmentation method is used to segment the text of the test question, thereby generating multiple words; a fine-grained word segmentation method is used to obtain a first input text at the character level for the generated words; each word generated by the coarse-grained word segmentation is randomly cut into several character strings to obtain a second input text; the first input text and the second input text are merged to obtain the word segmentation result.

[0023] Furthermore, the model pre-training includes the following steps:

[0024] Label word mapping: For the label set in the training set, each label in the label set is mapped to the corresponding label word by defining a label mapping function, and a label word set is established;

[0025] S42, prompt learning pre-training: given training samples Q ={q1,…,q h ,…,q z}, Q ∈A,q h Indicates the hth character in the sentence; define the prompt template "q h is a R ] entity", where [ R ] is a label slot, indicating q h The label word corresponding to the category;

[0026] For the original sentence Q After obtaining z prompt templates, connect them with commas to get the complete prompt P ={p1,…,p h ,…,p z}, p h Indicates that for q h Tips;

[0027] Will Q and P Splice them together as the input of Bert:

[0028] F =[CLS] Q [SEP] P ;

[0029] Among them, [CLS] indicates the beginning of a sentence, [SEP] is used to separate two sentences, and the new sequence obtained is F Comply with BERT's input specifications;

[0030] Will F Input into Bert for pre-training, so that the model can predict the entity label word corresponding to the [MASK] tag.

[0031] Furthermore, in the hint learning pre-training, the loss function of the model is calculated based on the difference between the predicted result and the true label:

[0032] L =-∑ h B h log(p(B h '| F ));

[0033] Among them, L represents the loss function of model prompt pre-training, B h represents the true label of the h-th character, B h ' represents the predicted label of the model.

[0034] Furthermore, in the test text vectorization, for a test text d∈D containing m word segments, D is a test text set, and the test text d={w1, w2, ..., w m}; participle w i The vector representation of = b i1 , b i2 ,…, b in ,in i =1, 2, ..., m; the test text d is expressed by the output d' of the Bert model as:

[0035] ;

[0036] In the formula, the dimension of d' is m×n, and m represents the length of the test text.

[0037] Furthermore, in the test question text label word annotation, the output obtained by the fully connected layer processing , where M is the weight matrix, b is the bias term;

[0038] Use the Softmax function to perform probability calculations, and the calculation formula is:

[0039] ;

[0040] in, y k 、 y j are the elements in vector y.

[0041] Furthermore, the generation of test knowledge points includes the following steps:

[0042] Tag word knowledge point generation: Based on all generated tag words, all knowledge points are traversed in the subject course database and test answer database to construct a knowledge point set;

[0043] Knowledge point association mining: Merge the labeled word set and knowledge point set of each test question into an item set to build a label word-knowledge point database;

[0044] Frequent item set mining: Use association analysis algorithm to mine frequent item sets that meet the support threshold in the label word-knowledge point database based on minimum support;

[0045] Generate association rules: Filter association rules from frequent item sets based on the set minimum support and minimum confidence;

[0046] Adjust screening parameters: Analyze the association rules generated after mining, and adjust the minimum support and minimum confidence according to the redundancy of the association rules;

[0047] Test question knowledge point generation: Set the weight of the test question label words according to the test syllabus, and set the minimum weight of the test question label words according to the learning level of the application group. Apply the generated effective association rules to generate knowledge points of the test question label words that meet the minimum weight requirements.

[0048] The present invention also provides a test question knowledge point generation system based on artificial intelligence, which is applied to the above-mentioned test question knowledge point generation method based on artificial intelligence, including a test question data acquisition unit, a test question text label word marking unit, a test question knowledge point generation unit, and a generated knowledge point evaluation unit;

[0049] The test question data collection unit is used to collect test question text data and complete the test question text label annotation through a crowdsourcing mode; the test question text label word annotation unit is configured as a Bert-CNN model, which vectorizes the test question text and inputs it into a convolutional neural network for secondary representation to complete the test question text label word annotation work.

[0050] As a preferred technical solution of the present invention, the test question knowledge point generation unit establishes a set of knowledge points corresponding to the label words based on the annotated label words of the test question text, and generates the knowledge points of the test question through the setting of label word weights and weight requirements; the knowledge point generation evaluation unit is used to evaluate the results of knowledge point generation through selected evaluation indicators.

[0051] The beneficial effects of the present invention are:

[0052] The present invention ensures the model training requirements by building its own data set through a crowdsourcing model, increases the sample capacity and improves the generalization ability of the model through multi-granularity word segmentation before labeling the model; constructs a comprehensive knowledge point set based on the annotated label words, and further improves the knowledge points corresponding to the annotated label words by mining the association between the annotated label words and the knowledge points, while also ensuring the validity of the annotated label words; in the process of generating knowledge points in the test questions, weights are set according to the importance of the knowledge points and the application object's mastery of the knowledge points, further ensuring the validity and accuracy of the generated knowledge points. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0054] Figure 1A schematic diagram of the steps of a method for generating test knowledge points based on artificial intelligence provided by an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the steps for generating knowledge points for a test question provided by an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of the structure of the CNN model in an embodiment of the present invention;

[0057] Figure 4 A schematic structural diagram of an artificial intelligence-based test question knowledge point generation system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0059] like Figure 1 As shown, this embodiment proposes a method for generating test knowledge points based on artificial intelligence, including the following steps:

[0060] S1. Test question text data preprocessing: Collect and obtain large-scale test question data, and perform knowledge point labeling to establish a test question dataset. It should be noted that the test question dataset includes a test question set and a label set. In order to enable the model to better adapt to the labeling of test question text knowledge point labels, the test question text data labeling process in this embodiment is completed in a crowdsourcing mode, including the following steps:

[0061] S11. Screen crowdsourcing users: Obtain the test question set according to the application scope, and recruit users with grade and subject categories corresponding to the test question set to complete the labeling of test question knowledge point labels. The recruited users can also be further screened by academic qualifications and work experience to ensure the quality of knowledge point labeling.

[0062] It is understandable that the test question set can be obtained through existing question banks on educational websites, such as Answer Network, Kaoyan Help, etc. In this embodiment, teachers are selected as the main group of recruits to improve the accuracy and efficiency of annotation.

[0063] S12. Training crowdsourcing users: To ensure the standardization and high quality of data, formulate and issue the process and specifications for labeling test questions and knowledge points, and train users to master the requirements for labeling.

[0064] S13. Crowdsourcing user evaluation: Conduct training evaluation on users who have completed the training, and select those with excellent evaluation results as users of the crowdsourcing project.

[0065] S14. Carry out labeling work: assign labeling tasks to each crowdsourcing user and obtain the test question label annotated text data.

[0066] S15. Label noun standardization: Create a label set by annotating text data with test question labels, and uniformly replace knowledge point nouns with the same label meaning to ensure the uniqueness of the knowledge point labels.

[0067] After completing the collection of test text data and annotation labels, the annotation labels are unified to ensure the accuracy of later model training and knowledge point generation. In this embodiment, by adopting a crowdsourcing model to obtain the data set, the format and quality of the generated data can be controlled, thereby ensuring that the data meets the requirements of later BERT model training.

[0068] S2. Test question text data segmentation: The test question text in the test set is segmented iteratively using different granularities to generate new text;

[0069] To ensure that all knowledge points in each test question are marked, word representations of different granularities are required when segmenting the test text to increase the vocabulary. The specific generation steps are as follows: First, the test text is segmented using a coarse-grained word segmentation method to generate multiple words; then, a fine-grained word segmentation method is used on the generated words to obtain the first character-level input text; each word generated by the coarse-grained word segmentation is randomly cut into several strings to obtain the second input text; and then the first and second input texts are combined to obtain the word segmentation results. For example, the word "power function" is segmented into six words: "power, function, number, power function, function, power function." Therefore, a word of length O can be segmented into (O+1)*O / 2 words after word segmentation of different granularities.

[0070] S3. Test data set division: divide the test data set into training set, validation set and test set;

[0071] S4. Model pre-training: Map labels to label words and complete the BERT model pre-training through prompt learning pre-training. Specifically, it includes the following steps:

[0072] S41. Label word mapping: For the label set in the training set, each label in the label set is mapped to the corresponding label word by defining a label mapping function, and a label word set is established.

[0073] For example, the label set T={t1,…,t h ,…,t g}, where g represents the number of labels, t h Represents the hth label, mapping function Φ (t h= "TP") = "time point", indicating that "time point" is the tag word "TP". It can be understood that the semantics of the tag information is improved through the mapping of labels and tag words.

[0074] S42, prompt learning pre-training:

[0075] Given a training sample Q ={q1,…,q h ,…,q z}, Q ∈A,q h Indicates the hth character in the sentence; define the prompt template "q h is a R ] entity", where [ R ] is a label slot, indicating q h The label word corresponding to the category.

[0076] It should be noted that during the training process, the size of each label slot is unified based on the number of Chinese characters contained in the longest label word in the training data.

[0077] For the original sentence Q After obtaining z prompt templates, connect them with commas to get the complete prompt P ={p1,…,p h ,…,p z}, p h Indicates that for q h Then, Q and P Splice them together as the input of Bert:

[0078] F =[CLS] Q [SEP] P ;

[0079] Among them, [CLS] indicates the beginning of a sentence, and [SEP] is used to separate two sentences. The new sequence obtained is F It complies with BERT’s input specifications. F The input is sent to Bert for pre-training. The goal of the task is to let the model predict the entity label word corresponding to the [MASK] tag.

[0080] The loss function of the model can be calculated as the difference between the predicted results and the true labels:

[0081] L =-∑ h B h log(p(B h '| F ));

[0082] Among them, L represents the loss function of model prompt pre-training, B h represents the true label of the h-th character, B h ' represents the predicted label of the model.

[0083] It is understandable that by pre-training the BERT model with prompt learning, the model's performance in identifying test text labels can be more effectively improved, and the model's ability to understand the data set can be improved, enabling it to better adapt to the test data set.

[0084] S5. Test question text vectorization: Use the Bert model to vectorize each word in the new text generated by word segmentation;

[0085] For a test text d∈D containing m words, it is expressed as:

[0086] d={w1, w2, ..., w m};

[0087] Participle w i The vector representation of = bi 1, bi 2,…, bi n, where i =1, 2, …, m; therefore, the test text d is expressed by the output d' of the Bert model as:

[0088] ;

[0089] In the formula, d' has a dimension of m × n, meaning that each test text in the test text set D has an m × n dimension. m represents the length of the test text (i.e., the number of tokens). Texts longer than m are truncated, and texts shorter than m are padded. n represents the length of the vector, i.e., the hidden_size of the last hidden layer in the BERT model. Because the test text is tokenized at different granularities, conversion and other processing are required to ensure that the input format meets the model requirements.

[0090] It should be noted that the BERT (Bidirectional encoder representations from transformers) model is an unsupervised pre-trained language model proposed by Google. This model uses a pre-training and then fine-tuning approach for natural language processing.

[0091] S6. Test question text label word annotation: The output word vector of the last layer encoder of the Bert model is input into the convolutional layer of the convolutional neural network (CNN). Different convolution kernels are used to perform convolution operations through CNN to obtain text features at different levels.

[0092] It should be noted that the CNN model performs convolution and pooling through three different convolution kernels, and finally concats the three results. The sizes of the three different convolution kernels are [2, 3, 4], and the number of each convolution kernel is 256. Then, the pooling layer is used to extract the maximum value of the column vector obtained by convolution.

[0093] Convolutional neural networks were first used in the field of image processing research. They obtain local features of images through convolution operations. Their sparse links and parameter sharing characteristics make them highly efficient. Figure 3 As shown in the figure, a convolutional neural network consists of an input layer, a convolution layer, a pooling layer, and an output layer. The convolution layer performs convolution operations to extract different features of the input information. The convolution layer is composed of several convolution kernel modules, and the parameters of each kernel are optimized using the backpropagation algorithm. After each convolution kernel is calculated, a column vector is generated, representing the features extracted from the sentence by the kernel. The number of features corresponds to the number of kernels. The pooling layer compresses and extracts the feature map matrix generated by the convolution layer, extracting the key local features of the text and reducing the dimensionality of the vector, thereby improving classification accuracy and model efficiency.

[0094] After the full connection layer and Softmax processing, the final labeling result is obtained. The output obtained after the full connection layer processing , where M is the weight matrix, b is the bias term.

[0095] Use the Softmax function to perform probability calculations, and the calculation formula is:

[0096] ;

[0097] in, y k 、 y j are the elements in vector y.

[0098] According to the calculation, Softmax( y k ) value, and obtain the annotated label words of the test text d.

[0099] It's understandable that the convolutional layer captures local features, while the fully connected layer combines all local features into global features, which are used to calculate the final score for each category. Softmax is the primary function that implements the fully connected layer. It functions as a classifier, outputting the final probability value for each category.

[0100] S7, Test question knowledge point generation: Generate the knowledge points contained in the test question through effective association rules based on the test question’s label words. Specifically, Figure 2 As shown in Figure 2, the generation of test knowledge points includes the following steps:

[0101] S71. Generation of knowledge points with label words: all knowledge points are traversed in the subject course database and the test question answer database based on all the generated label words to construct a knowledge point set.

[0102] It is understandable that the label words of the test questions are the knowledge point keywords. The label words are used to find matching knowledge points from the corresponding subject knowledge points in the existing learning platform database. At the same time, in order to ensure the comprehensiveness and attention of the knowledge points, the knowledge points mentioned in the test answers are also matched.

[0103] S72. Knowledge point association mining: Merge the annotated label word set and knowledge point set of each test question into an item set to build a label word-knowledge point database.

[0104] S73. Frequent item set mining: Use association analysis algorithms (such as Apriori algorithm or FP-Growth algorithm) to mine frequent item sets that meet the support threshold in the label word-knowledge point database based on minimum support.

[0105] S74. Generate association rules: Filter association rules from the frequent item set according to the set minimum support and minimum confidence.

[0106] S75. Adjust screening parameters: Analyze the association rules generated after mining, and adjust the minimum support and minimum confidence according to the redundancy of the association rules.

[0107] It should be noted that in the process of frequent item set mining, the initial minimum support and minimum confidence are set according to the characteristics of the data set and experience. After the association rules are generated, the parameters are adjusted and mining is performed again according to whether there are too few association rules or whether they are redundant, until reasonable and effective association rules that meet the requirements of the test knowledge points are generated.

[0108] It is understandable that whether there are too few association rules or whether they are redundant, that is, whether the knowledge points generated by different label words in the same test question overlap, etc., by resetting the support and confidence, the correlation between the labeled label words and the corresponding knowledge points is guaranteed, the uniqueness of the generated label knowledge points is guaranteed, and the validity of the labeled label words is guaranteed.

[0109] S76. Test question knowledge point generation: Set the weight of the test question label words according to the test syllabus, and set the minimum weight of the test question label words according to the learning level of the application group. Apply the generated effective association rules to generate knowledge points of the test question label words that meet the minimum weight requirements.

[0110] It is understandable that all the knowledge point label words of the test questions are not necessarily the knowledge points tested by the test questions or that certain knowledge points are not required in the problem-solving process. Therefore, the weights of the test question label words are set according to the test syllabus. At the same time, since the generated knowledge points have different requirements for different objects, in order to improve the personalization of the generated test question knowledge points, the minimum weight of the test question label word selection is set for different objects. That is, for objects with fewer knowledge points, as many knowledge points as possible are generated. For objects with more knowledge points, the knowledge points in the syllabus are mainly generated. Therefore, this step is mainly to set the weights of the test question label words, that is, to adjust the importance and number of the generated knowledge points.

[0111] It should be noted that association rules are of the form The implication of , , X and Y are different itemsets and are non-empty. In a transaction, if the antecedent X appears, then the consequent Y will appear with a certain probability. Assume that there is an association rule , then the support Indicates the possibility of a certain item set X∪Y appearing in the transaction database; confidence It represents the probability of the consequent occurring under the condition that the antecedent occurs, and is the possibility of a conditional probability occurring.

[0112] S8. Model indicator evaluation: This embodiment uses precision (P), recall (R), and F1-score as evaluation criteria for model knowledge point generation. The calculation formula is:

[0113] ;

[0114] Where, T P Indicates the number of positive class samples judged as positive classes, F PIndicates the number of negative class samples judged as positive class, T N Indicates the number of negative class samples judged as negative classes, F N Indicates the number of positive samples that are classified as negative. It is understandable that the calculation process requires the use of a confusion matrix to record the classifier's judgment results.

[0115] This embodiment also provides a test question knowledge point generation system based on artificial intelligence, such as Figure 4 As shown, it includes a test question data collection unit, a test question text label word marking unit, a test question knowledge point generation unit, and a generated knowledge point evaluation unit.

[0116] The test question data collection unit is used to collect test question text data and complete test question text labeling through a crowdsourcing mode.

[0117] The test question text label word marking unit is configured as a Bert-CNN model, which vectorizes the test question text and inputs it into a convolutional neural network for secondary representation to complete the label word marking work of the test question text.

[0118] The test question knowledge point generating unit establishes a knowledge point set corresponding to the label words based on the annotated label words of the test question text, and generates the knowledge points of the test question by setting the label word weights and weight requirements.

[0119] The generated knowledge point evaluation unit is used to evaluate the results of knowledge point generation using selected evaluation indicators.

[0120] The present invention ensures the model training requirements by building its own data set through a crowdsourcing model, increases the sample capacity and improves the generalization ability of the model through multi-granularity word segmentation before labeling the model; constructs a comprehensive knowledge point set based on the annotated label words, and further improves the knowledge points corresponding to the annotated label words by mining the association between the annotated label words and the knowledge points, while also ensuring the validity of the annotated label words; in the process of generating knowledge points in the test questions, weights are set according to the importance of the knowledge points and the application object's mastery of the knowledge points, further ensuring the validity and accuracy of the generated knowledge points.

[0121] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for generating test knowledge points based on artificial intelligence, characterized by: The following steps are involved: Test question text data preprocessing: Collect large-scale test question data, label knowledge points, and establish a test question dataset; Test question text data segmentation: The test question text in the test set is segmented iteratively using different granularities to generate new text; Test data set division: Divide the test data set into training set, validation set and test set; Model pre-training: Map labels to label words and complete BERT model pre-training through prompt learning pre-training; Test question text vectorization: Each word in the new text generated by word segmentation is vectorized using the Bert model; Test question text label word annotation: The output word vector of the Bert model is input into the convolutional layer of the convolutional neural network. Different convolution kernels are used for convolution operations. Then, the maximum value of the column vector obtained by convolution is extracted through the pooling layer. After passing through the fully connected layer and Softmax processing, the label word annotation result is obtained. Test question knowledge point generation: Generate the knowledge points contained in the test question through effective association rules based on the test question's label words; Model indicator evaluation: Model evaluation is performed by selecting precision, recall, and F1 value as evaluation criteria for model knowledge point generation; The test question text data segmentation includes the following steps: The test text is segmented using a coarse-grained word segmentation method to generate multiple words; a fine-grained word segmentation method is used on the generated words to obtain a character-level first input text; each word generated by the coarse-grained word segmentation is randomly cut into several character strings to obtain a second input text; the first input text and the second input text are merged to obtain a word segmentation result; The model pre-training includes the following steps: Label word mapping: For the label set in the training set, each label in the label set is mapped to the corresponding label word by defining a label mapping function, and a label word set is established; Tips for learning pre-training: given training samples Q ={q1,…,q h ,…,q z }, Q ∈A,q h Indicates the hth character in the sentence; the prompt template is defined as: q h is a R ] entity, where [ R ] is a label slot, indicating q h The label word corresponding to the category; For the original sentence Q After obtaining z prompt templates, connect them with commas to get the complete prompt P ={p1,…,p h ,…,p z }, p h Indicates that for q h Tips; Will Q and P Splice them together as the input of Bert: F =[CLS] Q [SEP] P ; Among them, [CLS] indicates the beginning of a sentence, [SEP] is used to separate two sentences, and the new sequence obtained is F Comply with BERT's input specifications; will new sequence F Input into Bert for pre-training, so that the model can predict the entity label word corresponding to the [MASK] tag; In the hint learning pre-training, the loss function of the model is calculated based on the difference between the predicted results and the true labels: L =-∑ h B h log(p(B h ’| F )); Among them, L represents the loss function of model prompt pre-training, B h represents the true label of the h-th character, B h ' represents the predicted label of the model.

2. The method for generating test knowledge points based on artificial intelligence according to claim 1, characterized in that: In the test question text data preprocessing, the test question text data annotation process is completed in a crowdsourcing mode, including the following steps: Screening crowdsourcing users: Obtain test question sets based on the application scope, and recruit users with corresponding grade and subject categories to complete the labeling of test question knowledge points, and further screen them based on academic qualifications and work experience; Training crowdsourcing users: Develop and distribute procedures and specifications for labeling test questions and knowledge points, and train users to master labeling requirements; Crowdsourcing user evaluation: Conduct training evaluation on users who have completed training, and select users for crowdsourcing projects based on the training evaluation results; Labeling: assign labeling tasks to each crowdsourcing user and obtain the test question labeling text data; Standardize label nouns: Create a label set by annotating text data with test question labels, and uniformly replace knowledge point nouns with the same label meaning.

3. The method for generating test knowledge points based on artificial intelligence according to claim 1, characterized in that: In the test text vectorization, for a test text d∈D containing m word segments, D is the test text set, and the test text d={w1, w2, ..., w m }; participle w i The vector representation of = b i1 , b i2 ,…, b in ,in i =1, 2, …, m; The test text d is expressed as the output d' of the Bert model: ； In the formula, the dimension of d' is m×n, and m represents the length of the test text.

4. The method for generating test knowledge points based on artificial intelligence according to claim 3, characterized in that: In the test question text label word annotation, the fully connected layer processes the output obtained , where M is the weight matrix, b is the bias term; Use the Softmax function to perform probability calculations, and the calculation formula is: ； in, y k 、 y j are the elements in vector y.

5. The method for generating test knowledge points based on artificial intelligence according to claim 1, characterized in that: The generation of test knowledge points includes the following steps: Tag word knowledge point generation: Based on all generated tag words, all knowledge points are traversed in the subject course database and test answer database to construct a knowledge point set; Knowledge point association mining: Merge the labeled word set and knowledge point set of each test question into an item set to build a label word-knowledge point database; Frequent item set mining: Use association analysis algorithm to mine frequent item sets that meet the support threshold in the label word-knowledge point database based on minimum support; Generate association rules: Filter association rules from frequent item sets based on the set minimum support and minimum confidence; Adjust screening parameters: Analyze the association rules generated after mining, and adjust the minimum support and minimum confidence according to the redundancy of the association rules; Test question knowledge point generation: Set the weight of the test question label words according to the test syllabus, and set the minimum weight of the test question label words according to the learning level of the application group. Apply the generated effective association rules to generate knowledge points of the test question label words that meet the minimum weight requirements.

6. An artificial intelligence-based test question knowledge point generation system, applied to the artificial intelligence-based test question knowledge point generation method according to any one of claims 1 to 5, characterized in that: It includes a test question data collection unit, a test question text label word marking unit, a test question knowledge point generation unit, and a generated knowledge point evaluation unit; The test question data collection unit is used to collect test question text data and complete the test question text label annotation through a crowdsourcing mode; the test question text label word annotation unit is configured as a Bert-CNN model, which vectorizes the test question text and inputs it into a convolutional neural network for secondary representation to complete the test question text label word annotation work.

7. The artificial intelligence-based test question knowledge point generation system according to claim 6, characterized in that: The test question knowledge point generation unit establishes a set of knowledge points corresponding to the label words based on the annotated label words of the test question text, and generates the knowledge points of the test question by setting the label word weights and weight requirements; the generated knowledge point evaluation unit is used to evaluate the results of the knowledge point generation through the selected evaluation indicators.

Images