In view of the problem that the distribution of knowledge points in the current knowledge system in the form of chapters can no longer meet the needs of students and teachers. This paper proposes to use SVM to identify knowledge points of tests, build a subject Knowledge Graph based on the identified knowledge points, and design and implement a subject Knowledge Graph system. This paper takes the mechanical movement of the first chapter of the eighth grade (upper) physics course as an example to conduct empirical research. By 10fold cross-validation, the average F1 value of the algorithm used in this paper is 89.66%.
In the existing teaching resources, knowledge points in the knowledge system are distributed in the form of chapter levels. However, in the face of fiercely competitive academic assessment, the difficulty of the assessment content of various tests is increasing day by day, focusing on the fusion of knowledge points. Therefore, the knowledge system distributed in the form of chapter levels can no longer meet the needs of students' learning and teachers' teaching. This paper argues that the Knowledge graph can show the complex correlation of knowledge points. At the same time, it also meets the needs of students and teachers better.
Knowledge Graph, as a form of structuring human knowledge, have attracted great attention from both academia and industry [1]. The Knowledge Graph describes the relationships between entities in real life in the form of triples. In view of the excellent form of Knowledge Graph, the research on subject Knowledge Graph has become a research hotspot of subject knowledge system in recent years. Chen et al. used TextRank to extract knowledge points from course introductions to construct a Knowledge Graph [2]. Cheng et al. used TF-IDF to retrieve teaching courseware and mine knowledge points to build a Knowledge Graph [3]. Su et al. evaluated the relationship between knowledge points by calculating the semantic similarity, PMI, and normalized Google distance between knowledge points [4]. Different from these studies, this paper believes that the relationship between knowledge points in the tests can better show the complex association of knowledge points.
Based on the above problems, this paper proposes to use the SVM [5] algorithm to identify the knowledge points of the tests, build the subject Knowledge Graph according to the identified knowledge point information, and design and implement the subject Knowledge Graph system based on test big data. The first chapter of the physics course in grade 8 of junior middle school is empirically studied to realize the knowledge point identification, the construction of a subject Knowledge Graph and the query of knowledge points, and the construction of a subject Knowledge Graph system. The method of knowledge point recognition and the construction of a subject Knowledge Graph are evaluated.
This paper aims to realize the identification of the knowledge points of the tests and the construction of the subject Knowledge Graph based on the teaching materials and tests, and to query the knowledge points according to the subject Knowledge Graph. The technology route is roughly divided into four steps. First, data acquisition is performed, including manual extraction of chapters, sections, and knowledge points information from the textbook and manual marking of knowledge points for the tests according to the extracted information. Then, this paper realizes the recognition algorithm of knowledge points, including preprocessing test points, extracting test points features by TF-IDF [6] method and realizing the recognition of test points by SVM classification algorithm. Next, using the chapters, sections, and knowledge points information extracted from the textbook and the tests constructs the subject Knowledge Graph. Finally, this paper uses the subject Knowledge Graph to query knowledge points.
It takes three steps to realize the identification of the knowledge points in the tests. First, the tests data is preprocessed. Secondly, the TF-IDF method is used to extract the features of the test data. Finally, the SVM classification algorithm is used to identify the knowledge points of the tests.
In order to improve the accuracy of tests identification, it is necessary to preprocess the data. The preprocessing steps of tests data include tests data modeling and tests word frequency matrix construction.
(
First, the pictures on the test are cleaned to obtain the tests data in plain text. When modeling the tests data, attributes such as "chapter", "test type", "knowledge point" of the tests are introduced to create the tests data model. That is, the tests data can be defined as ℚ =< |0 < ≤ >, where = , , , , represents the -th tests, represents the tests data after cleaning the picture, represents the word segmentation result, represents the chapter to which the tests belong, represents the type of the tests, represents the knowledge point of the tests, and represents the number of tests.
(
Jieba determines the association probability between Chinese characters through a Chinese thesaurus and forms phrases with high probability between Chinese characters to form word segmentation results. This paper uses Jieba to segment the test to get . After completing the word segmentation, the word segmentation results are cleaned, and the irrelevant phrases such as prepositions and adjectives are screened out, so that . = − . The complete set of phrases after word segmentation and screening of all tests data is denoted as = { |0 < ≤ }, where represents the number of phrases, and represents the -th phrase in the complete set of phrases. Then construct a word frequency matrix Μ =< > according to the complete set of phrases, and represents whether the -th phrase appears in the -th tests. 2.2.2.
Since the word frequency matrix obtained from the preprocessing results still has a large number of word segmentations, this paper uses the TF-IDF method to extract the features of the tests and optimize the word frequency matrix.
The TF-IDF method is often used to evaluate the importance of a word or phrase to a document set or one of the documents in a corpus. Therefore, this paper uses the TF-IDF method to extract the more important feature words from the tests corresponding to each knowledge point. TF-IDF consists of two parts, TF and IDF. TF refers to word frequency, which indicates the frequency of the word or phrase appearing in the knowledge point corresponding to the tests. The calculation of TF is shown in equation 1. knowledge point.
Among them, represents the number of tests in which the word or phrase appears in the tests corresponding to the knowledge point, and refers to the total number of tests corresponding to the
IDF refers to the inverse document frequency, the value of IDF is inversely proportional to the frequency of the word or phrase in the tests bank. The calculation of IDF is shown in equation 2. denominator, the denominator in the equation needs to be added by 1.
Among them, represents the total number of tests in the tests bank. To avoid having a 0 in the Among them, is the Lagrange multiplier vector, and the value represents the degree of penalty for misclassified points. The larger the
value, the greater the penalty for misclassified points. ② For ∗ calculated by ①, select a positive component 0 < ∗ < of ∗ and calculate ∗ as shown in equation 7.
∗ = − ∑ =1 ∗ ( , ), ③ Finally, the classification decision function of SVM is obtained as shown in equation 8, and the classification hyperplane of SVM is shown in equation 9.
( ) = (∑ =1 ∗
( , ) + ∗), ∑ =1 ∗ ( , ) + ∗ = 0,
In this paper, the SVM algorithm is used to classify each knowledge point label into two categories so as to realize the identification of the knowledge points in the tests. The basic idea of the SVM algorithm is to find the hyperplane that can correctly divide the training set and has the largest interval.
①
Define the training set
= {( 1, 1), ( 2, 2), ⋯ , ( , )}, where is an -dimensional feature vector, and the
value takes the form of 1 or -1. Then select the kernel function ( , ) and
the penalty coefficient , calculate the hyperplane when the interval is the largest, so that the closest
point to the hyperplane is as far away from the hyperplane as possible. Next, use Lagrange to solve
optimization problem, the result is shown in equation (
∑ =1 = 0,
0 ≤ ≤ ,
,
,
(
The subject Knowledge Graph this paper constructed in this paper defines three concept: "chapter": , "section": and "knowledge point": , and defines the subject Knowledge Graph this paper constructed as ℤ =< ℎ, , >=<< , 1, >, < , 2, >, < , 3, >>, Where ℎ represents the head entity, represents the relationship, represents the tail entity, 1 represents the relationship between chapters and sections at the chapter level, 2 represents the relationship between sections and knowledge points at the chapter level, and 3 represents the relationship between knowledge points and knowledge points. In this paper, the entities "chapter", "section" and "knowledge point" are extracted from the textbook through artificial mode, and the relationship between entities "chapter" and "section", "section" and "knowledge point" is defined in the form of the chapter level of knowledge point in the textbook. The knowledge point information is identified from the tests according to this paper proposed test knowledge point identification algorithm: the co-occurrence of knowledge points and the inspection frequency of knowledge points. Then judge the dependence between knowledge points and the strength of knowledge points in the test paper. Finally, the chapters, sections, knowledge points, and their relationships are stored in the Neo4j graph database in the form of triples, and the thickness of the relationship indicates the strength of the relationship between knowledge points, and the size of nodes indicates the degree of knowledge points.
Conduct knowledge point inquiry, first connect the Neo4j database, and then judge whether the knowledge point 1 exists in the database, if the knowledge point exists in the database, then query the relevant information when 1 is the head entity and the tail entity, and build Knowledge Graph ( 1) and relation list ( 1) related to knowledge point, if the knowledge point does not exist in the database, the output "the entity has not been added to the database yet".
This paper takes the mechanical movement of the first chapter of the eighth grade (upper) physics course as an example of how to conduct empirical research. Using the domestic authoritative primary and secondary education resource website-subject network (zxxk.com), the eighth grade (upper) physics course in the first chapter of the mechanical movement part of the sample tests, simulated 13740 tests.
In this paper, the SVM classification is performed on each type of knowledge point separately, so as to realize the multi-label classification of the knowledge points in the tests. First, preprocess the tests data, select the Chinese word segmentation component Jieba to segment the tests data, filter the irrelevant phrases such as adjectives and prepositions, and construct the tests word frequency matrix. Then, the TF-TDF method is used to extract the tests characteristics that can completely cover all the tests of the knowledge point. For the tests of different knowledge points, the word frequency matrix is optimized by the tests features of all knowledge points. The SVM model is trained by the optimized word frequency matrix to realize the classification of a single knowledge point in the tests and finally realize the identification of the knowledge points in the tests.
As shown in Figure 1, the SVM model was optimized by adjusting the proportion of the training set, and the model performance was evaluated by the F1 value of the test set.
According to the chapter level of the knowledge points in the textbook and combined with the knowledge point information identified by the knowledge point recognition algorithm of the tests, the subject Knowledge Graph is constructed. The subject Knowledge Graph constructed in this paper contains 17 nodes: 1 "chapter" node, 4 "section" nodes, and 12 "knowledge point" nodes, and the relationships between them. As shown in Figure 3, the Knowledge Graph constructed in this paper not only retains the chapter level of knowledge points, but also includes the frequency of knowledge points in the test paper and the degree of relevance of knowledge points in the test paper.
According to the method proposed in this paper, a disciplinary knowledge atlas system is constructed. The system includes three functional interfaces: test knowledge point recognition, Knowledge Graph display, and knowledge point query interface.
In the test knowledge point identification interface, enter the tests. The system will display the time of identification, the identified tests, and the identification result. It is convenient for users to learn correspondingly according to the recognition results.
In the Knowledge Graph display interface, user can clearly observe the distribution of knowledge points in chapters, the frequency of inspection in the test paper, and the dependencies between knowledge points in the tests.
In the knowledge point query interface, enter the knowledge point name. The system will display the relevant information about the knowledge point: the relationship diagram and the relationship table, which is convenient for learning about a certain knowledge point.
In order to solve the problem that the knowledge points in the current knowledge system are distributed in the form of chapter levels. This paper uses the SVM algorithm to identify the knowledge points of the tests. By 10-fold cross-validation, the average F1 value of the algorithm used in this paper is 89.66%, which can effectively identify the knowledge points in the test set. The subject knowledge graph is constructed by identifying the knowledge points information and manually extracting knowledge points at the textbook chapter level. According to the method this paper proposed, this paper designed and implemented the subject Knowledge Graph system and realized three interfaces in the system: test knowledge point identification, Knowledge Graph display, and a knowledge point query interface. The next work will be carried out in two aspects: optimization of the method for identifying knowledge points in tests and adding functions such as question-and-answer search and test recommendation to the system.
This work was supported by the Educational Science Planning of Henan Province, China under Grant [2021YB0037]; the Open Fund of Scientific Research Laboratory of Henan University for undergraduates under Grant [No.71].