A unit test planning method based on knowledge point fusion analysis

Abstract

The present invention discloses a unit test planning method for knowledge point fusion analysis, including: S1, collecting knowledge point data and preprocessing; S2, constructing a knowledge point graph model and dividing regions, calculating the partition center features of each partition and the global features of the knowledge point graph model; S3, generating an important graph index sequence according to the global features and the partition center features, calculating the knowledge point set corresponding to each node and the knowledge point set sequence according to the node feature similarity; S4, calculating the set historical average score features of each knowledge point set and completing the knowledge point difficulty classification, and sorting the difficulty categories of all knowledge point sets to obtain a difficulty sequence; S5, calculating the trimming quantile according to the difficulty sequence and trimming the knowledge point sets of different difficulties in the knowledge point set sequence, and generating sample questions according to the difficulty sequence and the knowledge point set sequence after the difficulty trimming. The present invention solves the problem that the traditional question setting method is difficult to fully capture the knowledge point association and the questions lack innovation.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and in particular to a unit test planning method for knowledge point fusion analysis. Background Art

[0002] Unit testing is an important part of educational assessment. It aims to conduct targeted assessments on students’ mastery of specific knowledge points by designing a series of test questions that cover the core content of the teaching course. Students can consolidate their knowledge and fill in gaps through unit testing, and teachers can provide timely feedback on teaching results through unit testing so that they can adjust their teaching strategies later. Therefore, the test results of unit testing must be accurate, and the key lies in how to select and plan test questions.

[0003] Traditional unit test planning methods mainly include:

[0004] 1. Empirical method: teachers select important knowledge points based on their own teaching experience and intuition, and design corresponding test questions;

[0005] 2. Statistical method: By conducting statistical analysis on the test data of previous years, we can determine the frequently tested knowledge points and difficult points, and design test questions accordingly;

[0006] 3. Fixed question type method: Teachers design a series of fixed question types and test points based on the course syllabus and textbooks to test students' mastery of knowledge points.

[0007] The above method has the following shortcomings:

[0008] It is difficult to fully capture the relationship and importance of knowledge points: Traditional methods can only focus on a single knowledge point or a simple combination of knowledge points. It is difficult to deeply explore the internal connections and mutual influences between knowledge points. This leads to the fact that important knowledge points may be ignored or omitted in unit test planning, and it is impossible to accurately evaluate students' mastery of all knowledge.

[0009] Test question generation lacks innovation and diversity: Traditional methods usually revolve around fixed question types and question-setting angles, resulting in a single test content and lack of variety. This not only fails to fully stimulate students' interest and enthusiasm, but may also limit students' thinking space and innovation ability. At the same time, fixed question types and test points may not be able to fully evaluate students' comprehensive abilities and adaptability. Summary of the invention

[0010] In view of this, the present invention aims to provide a unit test planning method for knowledge point fusion analysis to solve the problems in traditional methods that it is difficult to fully capture the association between knowledge points and the lack of innovation in questions.

[0011] A unit test planning method for knowledge point fusion analysis includes the following steps:

[0012] S1. Collect knowledge point data, including the name of the current knowledge point, the question, the solution, the name of the related knowledge point, and the historical average score of each question, and screen and classify the knowledge point data to obtain preprocessed knowledge point data;

[0013] S2. Construct a knowledge point graph model based on the preprocessed knowledge point data, partition the knowledge point graph model according to the node features, and calculate the partition center features of each partition and the global features of the knowledge point graph model;

[0014] S3, generating an important graph index sequence according to the global features and the partition center features, calculating the node feature similarity between each node and other nodes in the important graph index sequence, obtaining the combination of each node and its associated nodes as the corresponding knowledge point set according to the node feature similarity, and sorting all knowledge point sets to obtain a knowledge point set sequence;

[0015] S4, calculating the historical average score of each knowledge point set and the secondary interaction features of the historical average scores of the included knowledge points, further calculating the historical average score features of the set according to the historical average score and the secondary interaction features, and classifying the difficulty of the corresponding knowledge point set based on the historical average score and the secondary interaction features, and sorting the difficulty categories of all knowledge point sets to obtain a difficulty sequence;

[0016] S5. Calculate the trimming quantile according to the difficulty sequence, trim the knowledge point sets of different difficulty levels in the knowledge point set sequence according to the trimming quantile, and generate sample questions according to the difficulty sequence and the knowledge point set sequence after the difficulty trimming.

[0017] Furthermore, the step S1 includes:

[0018] S11, collecting knowledge point data, wherein the knowledge point data includes the current knowledge point name, question, solution, related knowledge point names, and historical average scores of each question;

[0019] S12, screening the knowledge point data, deleting the knowledge points that have no corresponding questions, solutions or historical average scores, and obtaining screened knowledge point data;

[0020] S13. Classify the filtered knowledge point data, using the current knowledge point name as an index, so that each current knowledge point name corresponds to at least 3 questions, each question corresponds to 1 historical average score and at least 1 solution, to obtain the preprocessed knowledge point data.

[0021] Furthermore, the step S2 includes:

[0022] S21. Constructing knowledge point graph model:

[0023] The name of each current knowledge point is used as the node name, the question, solution and the historical average score of each question are used as the initial features of the node, and the edges of the graph model are constructed with the names of the associated knowledge points;

[0024] S22. Calculate the node features of the knowledge point graph model:

[0025] After the word embedding operation is performed on the initial node features, the RoBERTa model is input for calculation, and the output result of the RoBERTa model is used as the node feature. The expression is as follows:

[0026] ;

[0027] in, is the node feature of the i-th node in the knowledge point graph model, i is the node index, For RoBERTa model operation, is the word embedding operation, is the initial node feature of the i-th node;

[0028] S23. Partition the knowledge point graph model:

[0029] The number of partitions is calculated based on the node characteristics of the knowledge point graph model. The expression is as follows:

[0030] ;

[0031] in, is the number of partitions, For multi-layer perceptron operation, is the dimension reduction matrix;

[0032] According to the number of partitions, the K-Means clustering algorithm is used to cluster the knowledge point graph model to obtain the divided knowledge point graph model;

[0033] The PageRank algorithm is used to calculate the central node of each partition, and the partition number is added to the node features of all nodes in each partition to indicate the partition to which the node belongs;

[0034] S24. Calculate the partition center features of each partition and the global features of the knowledge point graph model:

[0035] The partition center feature is calculated based on the central nodes of each partition. The expression is as follows:

[0036] ;

[0037] in, is the partition center feature of the jth partition, j is the partition index, is the number of nodes in the jth partition, It is a graph convolutional network operation;

[0038] The global features of the knowledge point graph model are calculated based on the partition center features. The expression is as follows:

[0039] ;

[0040] in, is the global feature of the knowledge point graph model, is the weight of the jth partition, For splicing operation, Score the number of nodes in the jth partition, Score the score of the jth partition, is the mean of the historical average scores of all nodes in the jth partition, is the similarity score of the jth partition, is the cosine similarity operation, It is the average value of all node features in the knowledge point graph model.

[0041] Since there are many knowledge points and they are text information, the amount of information is large and the vector dimension is large, when faced with massive and complex knowledge point data, traditional deep learning methods often find it difficult to effectively capture the relationship between knowledge points. The present invention introduces the idea of ​​partitioning, first uses a multi-layer perceptron to adaptively calculate the number of partitions, then uses the K-Means algorithm for clustering partitions, and then uses the PageRank algorithm to calculate the central nodes of each partition; on this basis, the present invention calculates the central features of each partition through operations such as graph convolutional networks (GCN), that is, the core information of the knowledge points in the area, and then combines these features with the number of nodes, scores, and similarity scores of the partitions. The global features of the entire knowledge point graph model are calculated by weighted summation, that is, the overall correlation between all knowledge points, which provides an important basis for the subsequent difficulty classification of knowledge point sets and sample question generation.

[0042] Compared with traditional methods: the present invention can capture the correlation between knowledge points more comprehensively through partition representation and extraction of regional center features, avoiding the information loss problem caused by large amount of information and high dimension in traditional methods; partition representation reduces the dimension of a single vector, thereby reducing the amount of calculation and improving the operation efficiency of the algorithm; combining the global features of the knowledge point graph model with the partition center features is also conducive to more accurate subsequent judgment of the difficulty category of each knowledge point set, thereby providing a more targeted knowledge point set for sample question generation.

[0043] Furthermore, the step S3 includes:

[0044] S31. Generate an important graph index sequence based on the global features of the knowledge point graph model and the partition center features of each partition. The expression is as follows:

[0045] ;

[0046] in, is an important graph index sequence, wherein the important graph index sequence includes the sequence numbers of the nodes in the knowledge point graph model;

[0047] S32. For each node in the important graph index sequence, the node feature is used as input to calculate the knowledge point set:

[0048] S321, using the cosine similarity method to calculate the node feature similarity between the current node and other nodes in the knowledge point graph model, and using the Z-score method to normalize the obtained node feature similarity to obtain the normalized node feature similarity;

[0049] S322, arranging the normalized node feature similarities in descending order and accumulating them from large to small, calculating the accumulated probability, selecting all nodes whose sum of accumulated probabilities is less than a preset threshold as associated nodes of the current node, combining the current node with its associated nodes, and obtaining a corresponding knowledge point set;

[0050] S323, traverse each node in the important graph index sequence, and repeat steps S321 and S322;

[0051] S33. According to the order of nodes in the important graph index sequence, each knowledge point set is concatenated to obtain a knowledge point set sequence.

[0052] The traditional method of only selecting the node with the highest probability or a fixed number of k nodes will result in the generated questions being too monotonous and lacking in diversity, covering only the most likely or fixed combinations of knowledge points. The present invention, however, uses a method of summing up the probabilities, which is dynamically adjusted based on the actual association relationship and probability distribution between the nodes. It does not require a preset fixed number of nodes, making the generated questions more diverse and covering more different combinations of knowledge points and association methods. Moreover, by selecting associated nodes through the cumulative probabilities, more novel questions can be generated, which is conducive to examining students' flexible use of knowledge points and their comprehensive understanding ability.

[0053] Furthermore, the length of the important graph index sequence in step S31 does not exceed 20.

[0054] Furthermore, the threshold in step S322 is 0.4.

[0055] Further, the step S4 includes:

[0056] S41. Calculate the historical average score feature of the knowledge point set:

[0057] First, calculate the historical average score of the knowledge point set, the expression is as follows:

[0058] ;

[0059] in, is the historical average score of the kth knowledge point set, k is the knowledge point set index, is the number of knowledge points in the kth knowledge point set, is the historical average score of the i-th knowledge point in the k-th knowledge point set, is the number of questions corresponding to the i-th knowledge point in the k-th knowledge point set, is the historical average score of the qth question corresponding to the ith knowledge point in the kth knowledge point set, where q is the question index;

[0060] Then, the quadratic interaction feature of the historical average score of each knowledge point in the knowledge point set is calculated, and the expression is as follows:

[0061] ;

[0062] in, is the quadratic interaction feature of the historical average score of each knowledge point in the kth knowledge point set, is the historical average score of the rth knowledge point in the kth knowledge point set, r is the knowledge point count index, ;

[0063] Finally, the historical average feature of the knowledge point set is calculated, and the expression is as follows:

[0064] ;

[0065] in, is the historical average score feature of the k-th knowledge point set, is the first set weight, is the second set weight;

[0066] S42. Classify the difficulty of the knowledge point set according to the historical average score feature of the set. The expression is as follows:

[0067] ;

[0068] in, is the difficulty probability distribution of the kth knowledge point set. The difficulty probability distribution is divided into three categories: difficult questions, medium questions, and easy questions. The category with the largest probability is selected as the difficulty category of the knowledge point set. It is the softmax function operation;

[0069] S43. According to the order of nodes in the important graph index sequence, the difficulty categories of each knowledge point set are concatenated to obtain a difficulty sequence, which is recorded as .

[0070] In step S41, the product of the historical average scores of each knowledge point in the knowledge point set is calculated and used as association information to reflect the mutual influence and dependence between knowledge points, enriching the expressive power of the historical average score features of the set, and participating in the difficulty classification of the knowledge point set in subsequent steps to more comprehensively consider the association relationship and mutual influence between knowledge points.

[0071] Furthermore, the step S5 comprises:

[0072] S51. Calculate the trimming quantile according to the difficulty sequence. The expression is as follows:

[0073] ;

[0074] in, is the clipping ratio feature, is a gated recurrent unit operation, They are the difficult question cutting quantile, medium question cutting quantile, and simple question cutting quantile of the k-th knowledge point set respectively;

[0075] S52, cutting the knowledge point sets of different difficulty levels in the knowledge point set sequence according to the cutting quantile, and obtaining the knowledge point set sequence after difficulty cutting, denoted as ;

[0076] S53. Generate sample questions according to the difficulty sequence and the knowledge point set sequence after difficulty trimming. The expression is as follows:

[0077] ;

[0078] in, For sample questions, Operates for the LLaMA2 model.

[0079] The LLaMA2 model used in this invention is an open source artificial intelligence model released by Meta in July 2023.

[0080] In step S51, a gated recurrent unit is combined to calculate the cropping ratio feature, which can dynamically adjust the cropping ratio according to the characteristics and distribution of the difficulty sequence, avoiding the limitations and inflexibility of the traditional fixed cropping ratio; in addition, the dynamic calculation of the cropping percentile can adapt to the diverse needs of questions of different difficulty levels. Whether it is a difficult question, a medium question or a simple question, it can be appropriately cropped according to its difficulty characteristics and the importance of the knowledge point, so as to generate sample questions that better meet actual needs in subsequent steps.

[0081] Compared with the prior art, the advantages of the present invention are:

[0082] (1) The present invention performs feature extraction and analysis from the bottom up. At the knowledge point level, the RoBERTa model is used for feature extraction. At the partition level, the number of partitions and the central node features are adaptively calculated to analyze knowledge points in a larger range. At the entire data level, the global features of the knowledge point graph model are calculated by combining the central node features of each partition to perform a macroscopic representation of all knowledge point features. After obtaining the global features of the knowledge point graph model, a knowledge point sequence is first generated, and then related knowledge points are supplemented to the sequence, and finally sample questions are generated, which solves the problem that the traditional unit test planning method is difficult to fully capture the knowledge point association and the questions lack innovation.

[0083] (2) Taking into account the characteristics of a large number of knowledge points and a large amount of information, the present invention introduces the idea of ​​partitioning to address the problem of insufficient vector representation capabilities, calculates the central nodes and features of the central nodes of each partition, and then obtains the global features of the knowledge point graph model; avoids the problem of information loss caused by the large amount of information and high dimension in traditional methods; at the same time, the partition representation reduces the dimension of a single vector, thereby reducing the amount of calculation and improving the operating efficiency of the algorithm.

[0084] (3) In terms of innovative test point generation, instead of directly generating text like traditional methods, the present invention first uses a generative model to generate an index of important knowledge points, then enriches each important knowledge point based on the index to obtain a knowledge point set, and finally performs adaptive knowledge point cutting based on the difficulty category of the knowledge point set to ensure that difficult questions contain more knowledge points and simple questions contain fewer knowledge points. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A flowchart of a unit test planning method for knowledge point fusion analysis provided in Example 1 of the present invention.

[0086] Figure 2 This is a flow chart for obtaining partition center features in Example 1 of the present invention.

[0087] Figure 3 This is a flow chart for obtaining global features in Example 1 of the present invention.

[0088] Figure 4 This is a flowchart of obtaining a knowledge point set sequence in Example 1 of the method of the present invention. DETAILED DESCRIPTION

[0089] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0090] Embodiment 1:

[0091] See also Figure 1 , a unit test planning method for knowledge point fusion analysis, comprising the following steps:

[0092] S1. Collect knowledge point data and filter and classify them to obtain preprocessed knowledge point data:

[0093] S11, collecting knowledge point data, wherein the knowledge point data includes the current knowledge point name, question, solution, related knowledge point names, and historical average scores of each question;

[0094] S12, screening the knowledge point data, deleting the knowledge points that have no corresponding questions, solutions or historical average scores, and obtaining screened knowledge point data;

[0095] S13. Classify the filtered knowledge point data, using the current knowledge point name as an index, so that each current knowledge point name corresponds to at least 3 questions, each question corresponds to 1 historical average score and at least 1 solution, to obtain the preprocessed knowledge point data.

[0096] S2. Build a knowledge point graph model based on the preprocessed knowledge point data, partition the knowledge point graph model according to the node features, and calculate the partition center features of each partition and the global features of the knowledge point graph model:

[0097] like Figure 2 , Figure 3 As shown;

[0098] S21. Constructing knowledge point graph model:

[0099] The name of each current knowledge point is used as the node name, the question, solution and the historical average score of each question are used as the initial features of the node, and the edges of the graph model are constructed with the names of the associated knowledge points;

[0100] S22. Calculate the node features of the knowledge point graph model:

[0101] After the word embedding operation is performed on the initial node features, the RoBERTa model is input for calculation, and the output result of the RoBERTa model is used as the node feature. The expression is as follows:

[0102] ;

[0103] in, is the node feature of the i-th node in the knowledge point graph model, i is the node index, For RoBERTa model operation, is the word embedding operation, is the initial node feature of the i-th node;

[0104] S23. Partition the knowledge point graph model:

[0105] The number of partitions is calculated based on the node characteristics of the knowledge point graph model. The expression is as follows:

[0106] ;

[0107] in, is the number of partitions, For multi-layer perceptron operation, is the dimension reduction matrix;

[0108] It should be noted that when there are a large number of knowledge points, a stratification mechanism can be added on the basis of partitioning. First, calculate whether stratification is needed. If stratification is needed, calculate the number of regions in each layer. To ensure that the extraction depth is not too deep, the maximum number of stratification layers is 3. The calculation method of stratification probability is:

[0109] ;

[0110] in, is the stratification probability, when , stratification is required, and the number of stratification layers is calculated as follows:

[0111] ;

[0112] in, is the number of stratified layers, The number of partitions at the 1st, 2nd, and 3rd layers respectively;

[0113] According to the number of partitions, the K-Means clustering algorithm is used to cluster the knowledge point graph model to obtain the divided knowledge point graph model;

[0114] The PageRank algorithm is used to calculate the central node of each partition, and the partition number is added to the node features of all nodes in each partition to indicate the partition to which the node belongs;

[0115] S24. Calculate the partition center features of each partition and the global features of the knowledge point graph model:

[0116] The partition center feature is calculated based on the central nodes of each partition. The expression is as follows:

[0117] ;

[0118] in, is the partition center feature of the jth partition, j is the partition index, is the number of nodes in the jth partition, It is a graph convolutional network operation;

[0119] The global features of the knowledge point graph model are calculated based on the partition center features. The expression is as follows:

[0120] ;

[0121] in, is the global feature of the knowledge point graph model, is the weight of the jth partition, For splicing operation, Score the number of nodes in the jth partition, Score the score of the jth partition, is the mean of the historical average scores of all nodes in the jth partition, is the similarity score of the jth partition, is the cosine similarity operation, It is the average value of all node features in the knowledge point graph model.

[0122] S3. Generate an important graph index sequence based on the global features and the partition center features, calculate the node feature similarity between each node and other nodes in the important graph index sequence, obtain the combination of each node and its associated nodes as the corresponding knowledge point set based on the node feature similarity, and further obtain the knowledge point set sequence:

[0123] like Figure 4 As shown;

[0124] S31. Generate an important graph index sequence based on the global features of the knowledge point graph model and the partition center features of each partition. The expression is as follows:

[0125] ;

[0126] in, is an important graph index sequence, wherein the important graph index sequence includes the serial numbers of the nodes in the knowledge point graph model, and the length of the important graph index sequence does not exceed 20;

[0127] S32. For each node in the important graph index sequence, the node feature is used as input to calculate the knowledge point set:

[0128] S321, using the cosine similarity method to calculate the node feature similarity between the current node and other nodes in the knowledge point graph model, and using the Z-score method to normalize the obtained node feature similarity to obtain the normalized node feature similarity;

[0129] S322, arranging the normalized node feature similarities in descending order and accumulating them from large to small, calculating the accumulated probability, selecting all nodes whose sum of accumulated probabilities is less than a preset threshold (0.4 in this embodiment) as associated nodes of the current node, combining the current node with its associated nodes, and obtaining a corresponding knowledge point set;

[0130] S323, traverse each node in the important graph index sequence, and repeat steps S321 and S322;

[0131] S33. According to the order of nodes in the important graph index sequence, each knowledge point set is concatenated to obtain a knowledge point set sequence.

[0132] For example, suppose you are generating unit test questions for junior high school mathematics, and you have built a knowledge point graph model that includes the main knowledge points of junior high school mathematics, such as "algebraic equations", "geometry introduction", "function foundation", etc., and these knowledge points are divided into different partitions;

[0133] First, based on the global features of the knowledge point graph model and the partition center features of each partition, the RoBERTa model is used to generate an important graph index sequence. , including the serial numbers of some of the most important knowledge points in the current test: [1, 3, 5, 7, 9], and the corresponding knowledge point names are: ["Algebraic Equations", "Introduction to Geometry", "Linear Function", "Quadratic Function", "Introduction to Probability"];

[0134] After calculation, the knowledge point set sequence is obtained as follows:

[0135] [“Algebraic equations + linear functions + quadratic functions”, “Introduction to geometry + linear functions”, …, “Introduction to probability + quadratic functions”].

[0136] S4. Classify the difficulty of the knowledge point set according to the historical average score characteristics of the knowledge point set, and further obtain the difficulty sequence:

[0137] S41. Calculate the historical average score feature of the knowledge point set:

[0138] First, calculate the historical average score of the knowledge point set, the expression is as follows:

[0139] ;

[0140] in, is the historical average score of the kth knowledge point set, k is the knowledge point set index, is the number of knowledge points in the kth knowledge point set, is the historical average score of the i-th knowledge point in the k-th knowledge point set, is the number of questions corresponding to the i-th knowledge point in the k-th knowledge point set, is the historical average score of the qth question corresponding to the ith knowledge point in the kth knowledge point set, where q is the question index;

[0141] Then, the quadratic interaction feature of the historical average score of each knowledge point in the knowledge point set is calculated, and the expression is as follows:

[0142] ;

[0143] in, is the quadratic interaction feature of the historical average score of each knowledge point in the kth knowledge point set, is the historical average score of the rth knowledge point in the kth knowledge point set, r is the knowledge point count index, ;

[0144] Finally, the historical average feature of the knowledge point set is calculated, and the expression is as follows:

[0145] ;

[0146] in, is the historical average score feature of the k-th knowledge point set, is the first set weight, is the second set weight;

[0147] S42. Classify the difficulty of the knowledge point set according to the historical average score feature of the set. The expression is as follows:

[0148] ;

[0149] in, is the difficulty probability distribution of the kth knowledge point set. The difficulty probability distribution is divided into three categories: difficult questions, medium questions, and easy questions. The category with the largest probability is selected as the difficulty category of the knowledge point set. It is the softmax function operation;

[0150] S43. According to the order of nodes in the important graph index sequence, the difficulty categories of each knowledge point set are concatenated to obtain a difficulty sequence, which is recorded as .

[0151] S5. Generate sample questions based on difficulty sequence and knowledge point set sequence:

[0152] S51. Calculate the trimming quantile according to the difficulty sequence. The expression is as follows:

[0153] ;

[0154] in, is the clipping ratio feature, is a gated recurrent unit operation, They are the difficult question cutting quantile, medium question cutting quantile, and simple question cutting quantile of the k-th knowledge point set respectively;

[0155] For example, calculate , then for difficult questions, cut off 0~10% of the knowledge points in the knowledge point set, for medium questions, cut off 10~30% of the knowledge points, and for simple questions, cut off 30~60% of the knowledge points;

[0156] S52, cutting the knowledge point sets of different difficulty levels in the knowledge point set sequence according to the cutting quantile, and obtaining the knowledge point set sequence after difficulty cutting, denoted as ;

[0157] S53. Generate sample questions according to the difficulty sequence and the knowledge point set sequence after difficulty trimming. The expression is as follows:

[0158] ;

[0159] in, For sample questions, Operates for the LLaMA2 model.

[0160] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0161] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in multiple embodiments of the present invention.

[0162] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A unit test planning method for knowledge point fusion analysis, characterized in that: The following steps are involved: S1. Collect knowledge point data, including the name of the current knowledge point, the question, the solution, the name of the related knowledge point, and the historical average score of each question, and screen and classify the knowledge point data to obtain preprocessed knowledge point data; S2. Construct a knowledge point graph model based on the preprocessed knowledge point data, partition the knowledge point graph model according to the node features, and calculate the partition center features of each partition and the global features of the knowledge point graph model; The construction of the knowledge point graph model includes: using each current knowledge point name as a node name, using the question, solution and the historical average score of each question as the initial node features, and constructing the edges of the graph model with the associated knowledge point names; calculating the node features according to the initial node features; Partitioning the knowledge point graph model includes: calculating the number of partitions according to node features, and clustering the knowledge point graph model using a K-Means clustering algorithm according to the number of partitions to obtain a divided knowledge point graph model; S3. Generate an important graph index sequence based on the global features and the partition center features, calculate the node feature similarity between each node in the important graph index sequence and other nodes, obtain the combination of each node and its associated nodes as the corresponding knowledge point set according to the node feature similarity, and sort all knowledge point sets to obtain a knowledge point set sequence, including: S31. Generate an important graph index sequence based on the global features of the knowledge point graph model and the partition center features of each partition. The expression is as follows: ； in, is an important graph index sequence, wherein the important graph index sequence includes the sequence numbers of the nodes in the knowledge point graph model, For RoBERTa model operation, is the global feature of the knowledge point graph model, For splicing operation, For the j The partition center characteristics of the partitions, is the number of partitions; S32. For each node in the important graph index sequence, the node feature is used as input to calculate the knowledge point set: S321, using the cosine similarity method to calculate the node feature similarity between the current node and other nodes in the knowledge point graph model, and using the Z-score method to normalize the obtained node feature similarity to obtain the normalized node feature similarity; S322, arranging the normalized node feature similarities in descending order and accumulating them from large to small, calculating the accumulated probability, selecting all nodes whose sum of accumulated probabilities is less than a preset threshold as associated nodes of the current node, combining the current node with its associated nodes, and obtaining a corresponding knowledge point set; S323, traverse each node in the important graph index sequence, and repeat steps S321 and S322; S33, according to the node order in the important graph index sequence, splicing each knowledge point set to obtain a knowledge point set sequence; S4, calculating the historical average score of each knowledge point set and the secondary interaction features of the historical average scores of the included knowledge points, further calculating the historical average score features of the set according to the historical average score and the secondary interaction features, and classifying the difficulty of the corresponding knowledge point set based on the historical average score and the secondary interaction features, and sorting the difficulty categories of all knowledge point sets to obtain a difficulty sequence; The calculation formula of the secondary interaction feature is as follows: ； in, For the k The quadratic interaction feature of the historical average score of each knowledge point in the knowledge point set, k is the knowledge point set index, For the k The number of knowledge points in a knowledge point set, For the k The first i The historical average score of knowledge points, For the k The first r The historical average score of each knowledge point; S5. Calculate the trimming quantile according to the difficulty sequence, trim the knowledge point sets of different difficulty levels in the knowledge point set sequence according to the trimming quantile, and generate sample questions according to the difficulty sequence and the knowledge point set sequence after the difficulty trimming.

2. The unit test planning method for knowledge point fusion analysis according to claim 1 is characterized in that: The step S1 comprises: S11, collecting knowledge point data, wherein the knowledge point data includes the current knowledge point name, question, solution, related knowledge point names, and historical average scores of each question; S12, screening the knowledge point data, deleting the knowledge points that have no corresponding questions, solutions or historical average scores, and obtaining screened knowledge point data; S13. Classify the filtered knowledge point data, using the current knowledge point name as an index, so that each current knowledge point name corresponds to at least 3 questions, each question corresponds to 1 historical average score and at least 1 solution, to obtain the preprocessed knowledge point data.

3. The unit test planning method for knowledge point fusion analysis according to claim 2 is characterized in that: The step S2 comprises: S21. Constructing knowledge point graph model: The name of each current knowledge point is used as the node name, the question, solution and the historical average score of each question are used as the initial features of the node, and the edges of the graph model are constructed with the names of the associated knowledge points; S22. Calculate the node features of the knowledge point graph model: After the word embedding operation is performed on the initial node features, the RoBERTa model is input for calculation, and the output result of the RoBERTa model is used as the node feature. The expression is as follows: ； in, is the node feature of the i-th node in the knowledge point graph model, i is the node index, For RoBERTa model operation, is the word embedding operation, is the initial node feature of the i-th node; S23. Partition the knowledge point graph model: The number of partitions is calculated based on the node characteristics of the knowledge point graph model. The expression is as follows: ； in, is the number of partitions, For multi-layer perceptron operation, is the dimension reduction matrix; According to the number of partitions, the K-Means clustering algorithm is used to cluster the knowledge point graph model to obtain the divided knowledge point graph model; The PageRank algorithm is used to calculate the central node of each partition, and the partition number is added to the node features of all nodes in each partition to indicate the partition to which the node belongs; S24. Calculate the partition center features of each partition and the global features of the knowledge point graph model: The partition center feature is calculated based on the central nodes of each partition. The expression is as follows: ； in, is the partition center feature of the jth partition, j is the partition index, is the number of nodes in the jth partition, It is a graph convolutional network operation; The global features of the knowledge point graph model are calculated based on the partition center features. The expression is as follows: ； in, is the global feature of the knowledge point graph model, is the weight of the jth partition, For splicing operation, Score the number of nodes in the jth partition, Score the score of the jth partition, is the mean of the historical average scores of all nodes in the jth partition, is the similarity score of the jth partition, is the cosine similarity operation, It is the average value of all node features in the knowledge point graph model.

4. The unit test planning method for knowledge point fusion analysis according to claim 3 is characterized in that: The length of the significance graph index sequence in step S31 does not exceed 20.

5. The unit test planning method for knowledge point fusion analysis according to claim 3 is characterized in that: The threshold in step S322 is 0.

4.

6. The unit test planning method for knowledge point fusion analysis according to claim 3 is characterized in that: The step S4 comprises: S41. Calculate the historical average score feature of the knowledge point set: First, calculate the historical average score of the knowledge point set, the expression is as follows: ； in, is the historical average score of the kth knowledge point set, k is the knowledge point set index, is the number of knowledge points in the kth knowledge point set, is the historical average score of the i-th knowledge point in the k-th knowledge point set, is the number of questions corresponding to the i-th knowledge point in the k-th knowledge point set, is the historical average score of the qth question corresponding to the ith knowledge point in the kth knowledge point set, where q is the question index; Then, the quadratic interaction feature of the historical average score of each knowledge point in the knowledge point set is calculated, and the expression is as follows: ； in, is the quadratic interaction feature of the historical average score of each knowledge point in the kth knowledge point set, is the historical average score of the rth knowledge point in the kth knowledge point set, r is the knowledge point count index, ; Finally, the historical average feature of the knowledge point set is calculated, and the expression is as follows: ； in, is the historical average score feature of the k-th knowledge point set, is the first set weight, is the second set weight; S42. Classify the difficulty of the knowledge point set according to the historical average score feature of the set. The expression is as follows: ； in, is the difficulty probability distribution of the kth knowledge point set. The difficulty probability distribution is divided into three categories: difficult questions, medium questions, and easy questions. The category with the largest probability is selected as the difficulty category of the knowledge point set. It is the softmax function operation; S43. According to the order of nodes in the important graph index sequence, the difficulty categories of each knowledge point set are concatenated to obtain a difficulty sequence, which is recorded as .

7. The unit test planning method for knowledge point fusion analysis according to claim 6 is characterized in that: The step S5 comprises: S51. Calculate the trimming quantile according to the difficulty sequence. The expression is as follows: ； in, is the clipping ratio feature, is a gated recurrent unit operation, They are the difficult question cutting quantile, medium question cutting quantile, and simple question cutting quantile of the k-th knowledge point set respectively; S52, cutting the knowledge point sets of different difficulty levels in the knowledge point set sequence according to the cutting quantile, and obtaining the knowledge point set sequence after difficulty cutting, denoted as ; S53. Generate sample questions according to the difficulty sequence and the knowledge point set sequence after difficulty trimming. The expression is as follows: ； in, For sample questions, Operates for the LLaMA2 model.

Images