A knowledge graph-assisted test question recommendation method

Abstract

The present invention discloses a method for recommending test questions assisted by a knowledge graph. The steps of the present invention are as follows: 1. Construction of a subject knowledge graph; 2. Construction of a test question-knowledge point relationship matrix; 3. Diagnostic analysis of the mastery of knowledge points based on the learner's historical answer data; 4. Calculation of test question similarity and correction of the knowledge point-test question weight relationship; 5. User similarity analysis based on the diagnosis results; 6. Recommendation of test questions based on the current user's diagnosis and the similarity between users. The present invention solves the problem that the existing recommendation method does not consider the differences between individuals and the correlation between knowledge points. The present invention integrates the correlation between knowledge points as auxiliary information, which can effectively reduce the errors caused by hidden attributes between knowledge points. The present invention can effectively mine the knowledge that best suits the user's own mastery to check for omissions and fill in the gaps, continuously overcome dependent knowledge, and improve the user's personal ability.

Description

Technical Field

[0001] The present invention belongs to the fields of knowledge graphs, recommendation systems and smart education, and relates to a test question recommendation method assisted by knowledge graphs. Background Art

[0002] With the rise of online education, the resource data for online learning in various disciplines has exploded, and question bank websites, mini-programs, and mobile applications have emerged in an endless stream. How individuals can effectively and selectively use these questions within a limited time, and fill in knowledge gaps based on their own learning characteristics and improve personal abilities has become a hot topic of personalized recommendations in current smart education.

[0003] Faced with massive amounts of data resources, it is very difficult for information consumers to select useful data from this vast amount of data, which leads to information overload. Recommendation systems and search engines both help users quickly discover useful information. However, unlike search engines, recommendation systems profile users by analyzing user behavior data, user tag data, or contextual information, thereby proactively pushing information of interest to users. Smart education recommendation systems differ from traditional e-commerce recommendation systems. E-commerce recommendations mainly focus on how to increase user purchase conversion rates and how to discover long-tail products to increase corporate revenue. From the perspective of education and users, recommendation systems in education focus on improving users' academic performance, constantly stepping out of their comfort zone, and improving their personal abilities. In recent years, some scholars have referred to traditional e-commerce recommendation methods and used collaborative filtering methods to predict scores for users and test questions. However, such methods do not take into account the differences between individuals and ignore the fact that users may have different knowledge mastery. The recommendation effect is limited and the results lack rationality, making it difficult to meet the requirements of accurate test question recommendations in education. At the same time, with the rapid development of psychology, personalized recommendations based on cognitive diagnosis theory have been widely used by scholars. The existing cognitive diagnosis model models the user's cognition from the knowledge level while ensuring good interpretability. Therefore, it is widely used to evaluate the ability level of user candidates (i.e., the mastery of knowledge points). Although the method based on cognitive diagnosis makes full use of the user's personalization and uses the user's hidden attributes as the measurement standard, it is easy to cause large errors.

[0004] In existing question banks, the screening structure is mostly based on chapter categories, which indiscriminately delivers questions to users. With the introduction of knowledge graphs, the relationships between knowledge points are organized and the information in the graphs is used as auxiliary information for question recommendations, which can better help users find questions that suit their specific characteristics for training. Summary of the Invention

[0005] The purpose of this invention is to provide a test question recommendation method assisted by knowledge graphs to address the shortcomings of the existing technology.

[0006] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0007] Step 1: Construction of subject knowledge graph.

[0008] Step 2: Construct a question-knowledge point relationship matrix.

[0009] Step 3: Conduct diagnostic analysis of the learner’s knowledge mastery based on their historical answer data.

[0010] Step 4: Calculate the similarity of test questions and correct the weight relationship between knowledge point test questions.

[0011] Step 5: Conduct user similarity analysis based on the diagnosis results.

[0012] Step 6: Recommend test questions based on the current user's diagnosis and the similarity between users.

[0013] The beneficial effects of the present invention are as follows:

[0014] By utilizing knowledge graphs in cognitive diagnosis models and integrating the relationships between knowledge points as auxiliary information, we can effectively reduce errors caused by hidden attributes between knowledge points. By fully leveraging user personalization, we create user profiles and then use latent semantic models to discover similarities between users, thus compensating for errors that may be caused by personalized diagnosis. Compared to indiscriminate delivery, this invention can effectively mine the knowledge that best suits the user's own knowledge to fill gaps, continuously overcome dependent knowledge, and improve the user's personal abilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a test question recommendation method assisted by knowledge graph;

[0016] Figure 2 This is an example diagram of a knowledge graph built with knowledge points;

[0017] Figure 3 This is an example diagram of cognitive diagnosis results;

[0018] Figure 4 This is a code example diagram of the test question recommendation method based on cognitive diagnosis and knowledge graph. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and examples.

[0020] like Figure 1-4 As shown in FIG, a method for recommending test questions assisted by knowledge graphs specifically includes the following steps:

[0021] Step 1: Construction of subject knowledge graph;

[0022] Determine the subject area and construct a knowledge point map based on authoritative books on the subject. The nodes constructed include chapters and sections. Chapters contain sections, and sections contain knowledge points. Chapters and sections share attributes such as unique identifiers, names, descriptions, and teaching objectives.

[0023] The knowledge points of relevant subjects are manually labeled, and after training, entities are extracted from all test questions and manually corrected to obtain subject knowledge points.

[0024] Domain experts construct the relationships between knowledge points, and integrate the relationships between knowledge points and chapters to obtain a knowledge graph.

[0025] Step 2: Construct a question-knowledge point relationship matrix;

[0026] Domain experts analyze the knowledge points tested in each test question and obtain the test question-knowledge point matrix Q, where Q ik =1 means that question i tests knowledge point k, Q ik =0 means that question i does not test knowledge point k.

[0027] Initialize the question-knowledge point weight matrix W, which represents the importance of the knowledge point in the question, W = Q, W ik ∈[0, 1]. W is initialized to Q, with a value of 0 or 1. The weight matrix is ​​subsequently modified based on the user's problem solving records, and its value is between 0 and 1. The initialized W weight matrix is ​​as follows:

[0028]

[0029] Where 0≤W ik ≤1, when W ik =0 means that question i does not test knowledge point k, W ik When !=0, it means that knowledge point k is tested in question i.

[0030] Step 3: Conduct diagnostic analysis of the learner’s knowledge mastery based on their historical answer data

[0031] The fuzzy cognitive diagnosis model (FuzzyCDF, Fuzzy Cognitive Diagnosis Framework) can estimate the user's mastery of knowledge points α, the user's discrimination of knowledge points a, the user's difficulty b, the probability of guessing the test questions g, the probability of error s, the user's mastery of the test questions X, and the user's potential ability θ through the user's answer records.

[0032] The knowledge point graph is used as auxiliary information to input into the fuzzy cognitive diagnosis model. This method believes that the dependency relationship of knowledge points has a similar impact on the difficulty of knowledge points. The more dependent a knowledge point is, the more basic it is. The formula for obtaining the user's mastery of knowledge points is:

[0033]

[0034] Among them, α uk Indicates user u’s mastery of knowledge point k; a uk Indicates the discrimination of knowledge points; b uk Indicates the difficulty of the knowledge point; θ u represents the potential ability level of user u; 1.7 is an empirical constant; r k represents the dependency influence of knowledge point k, r k The formula is as follows:

[0035]

[0036] triple(k, rely, r) indicates that knowledge point k in the knowledge graph depends on knowledge point r, and n indicates the total number of knowledge points in the section to which the current knowledge point belongs.

[0037] The mastery of objective questions is the same as the fuzzy cognitive diagnosis model, which is the fuzzy intersection of the user's mastery of knowledge points. In the present invention, it is believed that the mastery of subjective questions and the mastery of all knowledge points tested by the question are related to the weight of the question on the knowledge point.

[0038]

[0039] Among them, η ui Indicates user u’s understanding of question i, W ik is the weight of knowledge point k in question i, α uk is the user u’s mastery of knowledge point k.

[0040] The fuzzy cognitive diagnosis model uses the Monte Carlo (MCMC, Markov chain Monte Carlo) algorithm to uk 、a uk 、b uk and θ u The final model output is the mastery of knowledge points α, the discrimination of knowledge points to users a, the difficulty of knowledge points to users b, the probability of guessing the test questions g, the probability of error in the test questions s, the user's mastery of the test questions X, and the user's potential ability θ.

[0041] Step 4: Calculate the similarity of test questions and correct the weight relationship between knowledge points and test questions

[0042] By correcting the weight relationship between knowledge points and test questions, the similarity of test questions is calculated, and similar test questions among the user's wrong questions are found in the recommendation method to consolidate the knowledge points.

[0043] Using the weight matrix, calculate the distance of the test questions:

[0044]

[0045] The closer sim(i, o) is to 1, the higher the similarity.

[0046] The weight matrix W is continuously adjusted based on the user's test response. The adjustment method is:

[0047]

[0048]

[0049] Formula (5) indicates that when the user's mastery of the knowledge point is likely to answer the question, or is insufficient to answer the question, and the actual situation is the same as the prediction, i is the question index, k is the knowledge point index, and X is the number of questions. i is the predicted mastery of question i, A is the number of users, λ is the influence coefficient, λ∈[0,1]; in the special case, when knowledge point k is an isolated node, that is, knowledge point k is a basic knowledge point and is at the bottom of the dependent layer in the graph, then λ=1, that is, the influence of the dependent point is 0. Represents the knowledge point pair W that knowledge point k depends on ik The influence of r is the dependent knowledge point index, L is the number of dependent knowledge points, Q ir =0 means that question i tests knowledge point r.

[0050] Formula (6) indicates that when the user has sufficient knowledge to answer the question, but fails to answer correctly, it is possible that the answer is a mistake, and the probability of answering correctly is deducted from the error rate; when the user has insufficient knowledge to answer the question, but answers correctly, it is possible that the answer is a guess. i represents the probability of guessing question i, s i represents the probability of failure on question i.

[0051] Step 5: Conduct user similarity analysis based on the diagnosis results

[0052] Select the target knowledge point t as the knowledge point to be recommended.

[0053] Source: 1. Mapping the position of user u in the entire graph, i.e. G u ∈G, where G u All the knowledge points in G belong to the points involved in the questions that user u has already done. uAs the input of the LFM model (Latent Factor Model), the knowledge points that the user may master at the current cognitive level are obtained.

[0054] 2. The user selects the target knowledge point t. Assume that the user can only select knowledge points that do not contain any child nodes (base nodes). The LFM model is used for prediction. The steps are:

[0055] The result of cognitive diagnosis, user-knowledge point mastery, is preprocessed and the matrix after preprocessing is recorded as α c . Consider the user's understanding of the knowledge points α uk When it is less than the median, it is considered that it has not been mastered, and as the value to be predicted in the recommendation system, it is recorded as 0, that is, α uk <median(k), then

[0056] Perform matrix decomposition on the processed user-knowledge point mastery matrix: α c =p·q T , randomly initialize matrices p and q, and compare the diagnosis results of the user-knowledge point mastery matrix with the corresponding p u and The value obtained by dot multiplication establishes the squared error loss function, where the loss function is defined as:

[0057]

[0058] λ is the regularization coefficient, r ui is the actual value, is the predicted value, Use the gradient descent method to optimize the loss function, and after a certain number of iterations, obtain the matrices p and q, as well as the prediction matrix α p .

[0059] Step 6: Recommend test questions based on the current user’s diagnosis and similarity between users

[0060] According to the prediction matrix α p Get the knowledge point t that is most suitable for the current user u to learn. Search for related questions for the target knowledge point t:

[0061] The target user is set as u. If the target user u does not master the dependent knowledge point k (the number of knowledge points is less than the median of the knowledge point and is considered to be not mastered), that is, α uk < median(k), then knowledge point k is added to the queue because it is the knowledge required by the target user u to learn knowledge point t. After determining knowledge point k, the most relevant questions for knowledge point k are found according to the weight matrix W. The measurement method is:

[0062]

[0063] i is the question with the largest proportion of this knowledge point in the examination questions, i is entered into the QR team, Indicates the weight sum of all knowledge points tested by question i. In the wrong question record of user u, find question N related to the current knowledge point k and similar to the wrong question, N is added to the queue QR, hat k =QR. Here, hat is the set of predicted knowledge points and questions, and QR is the recommended questions.

[0064] User u’s recommended test questions hat ki Correct answer, find hat according to W matrix ki Knowledge points to be examined i , and check the mastery of each knowledge point, P i is the set of knowledge points tested by question i. If all knowledge points are mastered, it is considered that the knowledge tested by the relevant questions is mastered, and the knowledge point P is obtained. i Chapter C, (P i ∈C), get the next knowledge point to be learned in this chapter and make a prediction of t. If all the knowledge points in this chapter are mastered, then move on to the next chapter.

[0065] User u’s recommended test questions hat ki Wrong answer. Find the knowledge points that knowledge point k depends on, as well as the previous knowledge points of knowledge point k in section C. The reason for the wrong answer may be that the dependent knowledge points are not mastered or the previous knowledge points of this knowledge point are not mastered. Iterate through the mastery of each knowledge point set in turn. If the mastery is greater than the threshold, skip it and assume it is mastered; if it is less than the threshold, it is considered not mastered and related questions are recommended.

[0066] like Figure 1 As shown in the figure, a flow chart of a test question recommendation method assisted by a knowledge graph is shown. A subject knowledge graph is constructed from the original data. In the cognitive diagnosis process, personalized diagnosis is performed based on the entity relationships in the knowledge graph. Then, a latent semantic model is used to perform similarity analysis. Finally, the test questions that are most suitable for the user to learn are obtained based on the analysis results and the graph.

[0067] like Figure 2 The figure shows an example of a knowledge graph built with knowledge points. Chapters, sections, and knowledge points are entities in the knowledge graph. The relationships between entities include inclusion and dependency. The entity chapter contains the entity section, which contains basic knowledge points. There are dependencies between basic knowledge points. For example, the garbage collector section contains knowledge points such as G1 and CMS. CMS relies on the knowledge point mark-compression algorithm, which in turn relies on the knowledge point reachability analysis algorithm.

[0068] like Figure 3The following is an example of cognitive diagnosis results, showing a user's mastery of various knowledge points. The internal nodes represent the user's mastery of each knowledge point. As shown in the figure, the user has the highest mastery of knowledge point 2 and the lowest mastery of knowledge point 4.

[0069] like Figure 4 The following is a code example of a test question recommendation method based on cognitive diagnosis and knowledge graphs. It indicates that the target knowledge point t is determined, and first the knowledge points on which the knowledge point t depends are found to be mastered. If not, the question with the greatest correlation with the dependent knowledge point and similar test questions in the historical question set are searched, and finally the test questions that meet the user's learning requirements are output.

Claims

1. A knowledge graph-assisted test question recommendation method, characterized by The steps include: Step 1: Construction of subject knowledge graph; Step 2: Construct a question-knowledge point relationship matrix; Step 3: Conduct diagnostic analysis of the learner's knowledge mastery based on their historical answer data; Step 4: Calculate the similarity of test questions and correct the weight relationship between knowledge points and test questions; Step 5: Conduct user similarity analysis based on the diagnosis results; Step 6: Recommend test questions based on the current user's diagnosis and the similarity between users; The calculation of question similarity and the correction of knowledge point question weight relationships are specifically implemented as follows: By correcting the weight relationship between knowledge points and test questions, the similarity of test questions is calculated, and similar test questions among the user's wrong questions are found in the recommendation method to consolidate the knowledge points; Using the weight matrix, calculate the distance of the test questions: When sim(i,o) is closer to 1, it means the similarity is higher; The weight matrix W is continuously adjusted based on the user's test response. The adjustment method is: Formula (5) indicates that when the user's mastery of the knowledge point is likely to answer the question, or is insufficient to answer the question, and the actual situation is the same as the prediction, i is the question index, k is the knowledge point index, and X is the number of questions. i is the predicted mastery of question i, A is the number of users; λ is the influence coefficient, λ∈[0,1]. In the special case, when knowledge point k is an isolated node, that is, knowledge point k is a basic knowledge point and is at the bottom of the dependent layer in the graph, then λ=1, that is, the influence of the dependent point is 0; Represents the knowledge point pair W that knowledge point k depends on ik The influence of r is the dependent knowledge point index, L is the number of dependent knowledge points, Q ir =0 means that question i examines knowledge point r; W ik is the weight of knowledge point k in test question i; Formula (6) indicates that when the user has sufficient knowledge to answer the question, but fails to answer correctly, it is possible that the answer is a mistake, and the probability of answering correctly is deducted from the error rate; when the user has insufficient knowledge to answer the question, but answers correctly, it is possible that the answer is a guess; where g i represents the probability of guessing question i, s i represents the probability of failure on question i.

2. A method for recommending test questions assisted by knowledge graph according to claim 1, characterized in that The construction of the test question-knowledge point relationship matrix is ​​specifically implemented as follows: Domain experts analyze the knowledge points tested in each test question and obtain the test question-knowledge point matrix Q, where Q ik =1 means that question i tests knowledge point k, Q ik =0 means that question i does not test knowledge point k; Initialize the question-knowledge point weight matrix W, which represents the importance of the knowledge point in the question, W = Q, W ik ∈[0,1]; W is initialized to Q, with a value of 0 or 1. The weight matrix is ​​subsequently modified based on the user's problem-solving records, and its value is between 0 and 1. The initialized W weight matrix is ​​as follows: Where 0≤W ik ≤1, when W ik =0 means that question i does not test knowledge point k, W ik When !=0, it means that knowledge point k is tested in question i.

3. The method for recommending test questions using knowledge graphs according to claim 2, characterized in that The diagnostic analysis of the knowledge point mastery based on the learner's historical answer data is specifically implemented as follows: The knowledge point graph is used as auxiliary information to input into the fuzzy cognitive diagnosis model. This method believes that the dependency relationship of knowledge points has a similar impact on the difficulty of knowledge points. The more dependent a knowledge point is, the more basic it is. The formula for obtaining the user's mastery of knowledge points is: Among them, α uk Indicates user u’s mastery of knowledge point k; a uk Indicates the discrimination of knowledge points; b uk Indicates the difficulty of the knowledge point; θ u represents the potential ability level of user u; r k represents the dependency influence of knowledge point k, r k The formula is as follows: triple(k,rely,r) indicates that knowledge point k in the knowledge graph depends on knowledge point r, and n indicates the total number of knowledge points in the section to which the current knowledge point belongs; The mastery of the test questions is the same as the fuzzy cognitive diagnosis model, which is the fuzzy intersection of the user's mastery of the knowledge points. This method believes that the mastery of the test questions and the mastery of all the knowledge points tested by the test questions are related to the weight of the test questions on the knowledge points. Among them, η ui Indicates user u’s understanding of question i, W ik is the weight of knowledge point k in question i, α uk is the user u’s mastery of knowledge point k; The fuzzy cognitive diagnosis model uses Monte Carlo algorithm to uk 、a uk 、b uk and θ u The final model output is the mastery of knowledge points α, the discrimination of knowledge points to users a, the difficulty of knowledge points to users b, the probability of guessing the test questions g, the probability of error in the test questions s, the user's mastery of the test questions X, and the user's potential ability θ.

4. The method for recommending test questions using knowledge graphs according to claim 3, characterized in that User similarity analysis is performed based on the diagnosis results. The specific analysis is as follows: Select the target knowledge point t as the knowledge point to be recommended; 1. Map out the position of user u in the entire graph, that is, G u ∈G, where G u All the knowledge points in G belong to the points involved in the questions that user u has already done; u As the input of the LFM model, obtain the knowledge points that the user may master at the current cognitive level; 2. The user selects the target knowledge point t. Assume that the user can only select knowledge points that do not contain any child nodes. Using the LFM model to make predictions, the steps are: The result of cognitive diagnosis, user-knowledge point mastery, is preprocessed and the matrix after preprocessing is recorded as α c ; Consider the user's mastery level α of knowledge points uk When it is less than the median, it is considered that the user has not mastered it. As the value to be predicted in the recommendation system, it is recorded as 0, that is, α uk < median(k), then Perform matrix decomposition on the processed user-knowledge point mastery matrix: α c =p·q T , randomly initialize matrices p and q, and compare the diagnosis results of the user-knowledge point mastery matrix with the corresponding p u and The value obtained by dot multiplication establishes the squared error loss function, where the loss function is defined as: λ is the regularization coefficient, r ui is the actual value, is the predicted value, Use the gradient descent method to optimize the loss function, and after a certain number of iterations, obtain the matrices p and q, as well as the prediction matrix α p .

5. The method for recommending test questions using knowledge graphs according to claim 4, characterized in that The recommendation of test questions based on the current user's diagnosis and the similarity between users is specifically implemented as follows: According to the prediction matrix α p Get the knowledge point t that is most suitable for the current user u to learn; search for related questions about the target knowledge point t: The target user is defined as u. If the target user u has not mastered the dependent knowledge point k, that is, α uk <median(k), then the knowledge point k is queued because this knowledge point is the knowledge required for the target user u to learn the knowledge point t. After establishing the knowledge point k, according to the weight matrix W, find the questions most relevant to the knowledge point k, and the measurement method is as follows: i is the question with the largest proportion of this knowledge point in the examination questions, i is entered into the QR team, Represents the weight sum of all knowledge points examined by question i; in the wrong question record of user u, find question N related to the current knowledge point k and similar to the wrong question, N is added to the queue QR, hat k =QR; where hat is the set of predicted knowledge points and questions, and QR is the recommended question; User u’s recommended test questions hat ki Correct answer, find hat according to W matrix ki Focus on the knowledge points P i , and check the mastery of each knowledge point, P i is the set of knowledge points tested by question i; If all knowledge points are mastered, it is considered that the knowledge tested by the relevant questions is mastered, and the knowledge point P is obtained. i Chapter C, P i ∈C, get the next knowledge point to be learned in this chapter and make a prediction of t; If you have mastered all of this chapter, proceed to the next chapter; User u’s recommended test questions hat ki Wrong answer; find the knowledge points that knowledge point k depends on, as well as the previous knowledge points of knowledge point k in chapter C. The reason for the wrong answer may be that the dependent knowledge points are not mastered or the previous knowledge points of this knowledge point are not mastered. Traverse the mastery of each knowledge point set in turn. If it is greater than the threshold, skip it and assume it is mastered; if it is less than the threshold, it is considered not mastered, and recommend related questions for it.

Images