A knowledge tracking method and system based on fine-grained similarity calculation

Abstract

The present invention relates to the field of intelligent education technology, and discloses a knowledge tracking method and system based on fine-grained similarity calculation, the method comprising: embedding questions and students' answers, and obtaining context-aware question embedding vectors and question-answer pair embedding vectors through self-attention aggregation; performing multiple difficulty corrections on the context-aware question embedding vectors and answer embedding vectors through a cross-attention mechanism to obtain the student's knowledge state; splicing the knowledge state and the question embedding vector, and performing performance prediction to obtain the student's performance prediction result. The present invention utilizes inherent attributes by enhancing the representation of questions, and explicitly calculates the similarity of dynamic attributes, and combines it with the attention mechanism to form a transformer model that is sensitive to response time; in addition, the present invention can also use a difficulty correction module to correct statistical attributes according to the individual perception of students.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to a knowledge tracking method and system based on fine-grained similarity calculation. Background Art

[0002] With the advent of the information age, online education has developed rapidly and gradually become one of the important learning channels for learners around the world. In recent years, the emergence of new technologies such as artificial intelligence and big data has brought unprecedented opportunities for online education. These technologies not only improve the efficiency of obtaining educational resources, but also promote the realization of personalized learning, allowing learners to choose appropriate learning content and methods according to their own needs.

[0003] The advantages of online education are obvious. First, online education resources are rich, and learners can easily access a large number of courses, videos and learning materials to meet different learning needs. Second, online learning provides a flexible and independent learning method. Learners can study according to their own schedule, breaking the time and space limitations of traditional education.

[0004] However, online learning also has some disadvantages. Due to the lack of face-to-face teacher supervision and immediate feedback, learners cannot understand their own knowledge level during the learning process and cannot choose the appropriate learning path. In addition, current online education provides students with undifferentiated learning materials and cannot provide students with personalized tutoring.

[0005] Knowledge tracking is an AI-based educational data mining technology field that extracts students’ knowledge status from their historical response data, thereby predicting the probability that students will correctly answer new questions. Based on this predictive ability, accurate knowledge level testing and personalized learning materials can be provided for each student, so that each student can have a personalized learning trajectory, making the learning process more specific and purposeful.

[0006] The knowledge tracking model based on the attention mechanism has been proven to have superior performance, but existing technologies usually only rely on the general representation of the question to calculate similarity, failing to fully utilize the fine-grained properties of the question. This prevents the attention mechanism from focusing on the answer records that truly affect the student's knowledge status, resulting in insufficient accuracy in knowledge tracking. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a knowledge tracking method and system based on fine-grained similarity calculation.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A knowledge tracking method based on fine-grained similarity calculation, comprising:

[0010] Embed the questions and students’ answers, and obtain the context-aware question embedding vector and question-answer pair embedding vector through self-attention aggregation;

[0011] The context-aware question embedding vector and the question-answer pair embedding vector are used through a cross-attention mechanism to obtain the student's initial knowledge state, and multiple difficulty corrections are performed based on subjective relative difficulty and statistical difficulty to obtain the student's knowledge state.

[0012] The student’s knowledge status and question embedding vector are concatenated, and performance prediction is performed to obtain the student’s performance prediction result.

[0013] In one embodiment, embedding the question and the student's answer specifically includes:

[0014] By fusing the embedding vector of the problem, the embedding vector of the knowledge point corresponding to the problem, the embedding vector of the statistical difficulty of the problem, and the embedding vector of the type of the problem, we can get the embedding vector of the problem with enhanced inherent attributes. ;

[0015] By fusing the embedding vector of the question, the embedding vector of the knowledge point corresponding to the question, the embedding vector of the answer corresponding to the question, the embedding vector of the statistical difficulty of the question, and the embedding vector of the question type, we can get the embedding vector of the question-answer pair with enhanced inherent attributes. .

[0016] In one embodiment, the embedding vector of the problem with enhanced inherent attributes is obtained by fusing the embedding vector of the problem, the embedding vector of the knowledge point corresponding to the problem, the embedding vector of the statistical difficulty of the problem, and the embedding vector of the type of the problem. , specifically including:

[0017] Embedding the problem into a vector The embedding vector of the knowledge point corresponding to the question Add together the embedding vector of the statistical difficulty of the problem and the embedding vector of the type of the problem Add and then concatenate to get the embedding vector of the problem with inherent attribute enhancement .

[0018] In one embodiment, the embedding vector of the question-answer pair with enhanced inherent attributes is obtained by fusing the embedding vector of the question, the embedding vector of the knowledge point corresponding to the question, the embedding vector of the answer corresponding to the question, the embedding vector of the statistical difficulty of the question, and the embedding vector of the type of the question. , specifically including:

[0019] Embedding the problem into a vector , the embedding vector of the knowledge point corresponding to the question Embedding vector of the answer corresponding to the question Add together the embedding vector of the statistical difficulty of the problem and the embedding vector of the type of the problem Add and then concatenate to get the embedding of the question answer pair with enhanced intrinsic attributes .

[0020] In one embodiment, the obtaining of the context-aware question embedding vector and the question-answer pair embedding vector through self-attention aggregation specifically includes:

[0021] After embedding the questions and students’ answers, we get the question embedding vectors with enhanced inherent attributes: and intrinsic property-enhanced question answering pair embedding vectors ;

[0022] Students in Always answer questions and The similarity of the questions answered at the same time in terms of answering time , by Always answer questions and Embedding the answer time of questions answered at any time and After subtraction, it is obtained through a nonlinear transformation; the similarity Integrate it into the self-attention mechanism to obtain the attention coefficient , based on the attention coefficient, the problem embedding vectors enhanced by the inherent attributes are respectively and intrinsic property-enhanced question answering pair embedding vectors Perform self-attention aggregation to obtain context-aware question embedding vector and context-aware question answering pair embedding vectors :

[0023] ;

[0024] ;

[0025] for The value in the self-attention aggregation at the moment; in the self-attention aggregation, the query, key, and value are all obtained by linearly transforming the same input.

[0026] In one embodiment, the context-aware question embedding vector and the question-answer pair embedding vector are used to obtain the student's initial knowledge state through a cross-attention mechanism, specifically including:

[0027] Embedding context-aware questions into vectors and context-aware question answering pair embedding vectors , after cross attention, we get the student's initial knowledge state; among them, After linear transformation, we obtain the query and key required for cross attention. After linear transformation, the value required for cross attention is obtained.

[0028] In one embodiment, the difficulty correction is implemented by a difficulty correction module; the difficulty correction module combines the cross-attention mechanism and the self-attention mechanism, and can automatically adjust the difficulty of the question according to the individual learning history of the student, so as to obtain the student's knowledge state;

[0029] The difficulty correction module includes a first difficulty correction module DC1 and a second difficulty correction module DC2. The first difficulty correction module DC1 and the second difficulty correction module DC2 are used alternately to adjust the student's knowledge status using subjective relative difficulty and statistical difficulty respectively.

[0030] In one embodiment, in the first difficulty correction module DC1, the student's prior knowledge state is taken into account. Embedding vectors for problems with inherent property enhancement The difference between Minus , and apply nonlinear transformation to obtain subjective relative difficulty, and use subjective relative difficulty to adjust students' knowledge status:

[0031] ;

[0032] represents the second weight matrix, represents the knowledge state of DC1 input to the i-th layer, The embedding vector representing the problem with inherent attribute enhancement, represents the second bias term, is a vector representing subjective difficulty perception, represents the activation function;

[0033] An output gate is designed to control the update of the student's knowledge state:

[0034] ;

[0035] ;

[0036] represents the output gate of DC1, represents the sigmoid activation function, represents the third weight matrix, represents the third bias term, Represents the knowledge state of the DC1 output of the i-th layer.

[0037] In one embodiment, in the second difficulty correction module DC2, the difficulty information is adjusted to the statistical difficulty; the embedding vector of the statistical difficulty of the question A nonlinear transformation is performed to represent the statistical difficulty, and then an output gate is used to control the update of the knowledge state:

[0038] ;

[0039] ;

[0040] ;

[0041] represents the fourth weight matrix, represents the knowledge state of the DC2 output of the i-th layer, The embedding vector representing the problem with inherent attribute enhancement, represents the fourth bias term, is a vector representing the statistical difficulty, represents the activation function; represents the fifth weight matrix, represents the fifth bias term, Represents the output gate of DC2.

[0042] A knowledge tracking system based on fine-grained similarity calculation, comprising:

[0043] The embedding aggregation module embeds questions and students’ answers, and obtains context-aware question embedding vectors and question-answer pair embedding vectors through self-attention aggregation;

[0044] The state acquisition module uses the cross-attention mechanism to obtain the student's initial knowledge state by embedding the context-aware question vector and the question-answer pair embedding vector. It then makes multiple difficulty corrections based on the subjective relative difficulty and statistical difficulty to obtain the student's knowledge state.

[0045] The performance prediction module concatenates the student’s knowledge status and question embedding vector, and performs performance prediction to obtain the student’s performance prediction result.

[0046] The system of the present invention corresponds to the method, and the preferred technical solutions applicable to the method are also applicable to the system.

[0047] Compared with the prior art, the beneficial technical effects of the present invention are:

[0048] The present invention first analyzes the importance of question similarity in the knowledge tracing (KT) task and divides the question attributes into three types according to their characteristics. The present invention mines question similarity in a more detailed and comprehensive way according to the characteristics of the three types. Specifically, the present invention exploits inherent attributes (such as question type) by enhancing the representation of questions. In addition, the present invention designs a dynamic similarity calculation module to explicitly calculate the similarity of dynamic attributes (response time) and combines it with the attention mechanism to form a transformer model that is sensitive to response time. Finally, the present invention uses a difficulty correction module to correct the statistical attribute (question difficulty) according to the individual perception of students. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of the overall knowledge tracking model in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0051] The present invention proposes a knowledge tracking method based on fine-grained similarity calculation, comprising:

[0052] Construction of the overall knowledge tracking model: The overall knowledge tracking model includes an input embedding module, a self-attention aggregation module, a knowledge extractor, a performance prediction module, and three fine-grained similarity calculation modules: an inherent attribute enhancement embedding module, a dynamic similarity calculation module, and a difficulty correction module. The structure of the knowledge tracking model is as follows Figure 1 shown.

[0053] In an online education environment, students' learning process is no longer limited to traditional classrooms and books, but is carried out through various online platforms and tools. These platforms and tools record every interaction between students and the system during the learning process, including the courses they browse, the discussions they participate in, the homework they submit, and the questions they answer. In order to gain a deeper understanding of students' learning behaviors and knowledge mastery, it is necessary to systematically analyze and process these interaction data. Therefore, the present invention introduces the concept of "student interaction sequence". The student interaction sequence contains all the interaction records between students and the system during the online learning process. By analyzing this interaction sequence, the present invention can more accurately grasp the students' learning trajectory and knowledge status, and provide strong support for subsequent personalized teaching.

[0054] The student's interaction sequence is recorded as , represents the total number of interaction records in the student interaction sequence, Interaction records at all times It can be represented by a tuple, .in, Indicates the problem, It is the knowledge point contained in the question. Is the student's question Among them, is a binary indicator variable, Indicates the answer is correct. Indicates an incorrect answer. Predicted student performance Is a continuous value between 0 and 1.

[0055] The main inventive concept of the present invention is as follows: first, embed the question and the student's answer, and then obtain the context-aware question embedding vector and answer embedding vector through self-attention aggregation. Then, the two embedding vectors are used to obtain the student's knowledge state at time t through the cross-attention mechanism. , the knowledge state at time t and the question embedding vector at time t After splicing, the student's performance prediction results are obtained through the performance prediction module.

[0056] In order to make better use of the properties of the problem, the properties of the problem are divided into three categories according to their characteristics:

[0057] (1) Intrinsic attributes: These attributes are inherent to the problem, including the knowledge points corresponding to the problem and the type of problem. They are objective factors that directly reflect the similarity of problems. Previous studies usually represent problems by the combination of the problem itself and the knowledge points it contains. Among them, the type of problem is an inherent attribute that is highly correlated with the similarity of problems. Using the embedding matrix , and to represent all questions, knowledge points and question types respectively, where , , are their respective dimensions, , , Respectively represent the number of questions, knowledge points and question types. For the answer, the present invention uses the embedding matrix to represent a binary answer, where is the embedding matrix Dimension.

[0058] (2) Dynamic attributes: This attribute is related to students. The difference between dynamic attributes and inherent attributes is that even the same question may have different similarities when presented to different students because of their different knowledge status. Students who have mastered the required knowledge and those who have not mastered it will have very different views on the problem. The present invention selects answering time as a representative attribute to capture the dynamic factors related to students: students with different knowledge status often spend different times when answering the same question. In order to effectively capture the embedding of answering time, the present invention applies the K-means clustering method to all training data and classifies it into 20 clusters. This clustering method helps manage the continuity of answering time by grouping similar answering times, thereby achieving a more structured representation in the model. In order to avoid the problem of data centralization, the present invention performs a logarithmic transformation on the answering time to achieve a more uniform cluster distribution. The present invention uses the matrix To represent the set of answer time embeddings, where is a matrix The dimension of Indicates the number of clusters for answering time. , whose answer time is embedded in It is indexed from the matrix by its answer time type Retrieved from .

[0059] (3) Statistical properties: These properties are related to the data set, including the difficulty of the questions, which is determined based on the average correct rate of the questions. The difficulty of a question can be expressed as the objective statistical average correct rate of a certain knowledge point multiplied by a constant :in It's a problem A collection of response records. Indicates that the response record is correct, and Indicates that the response record is in error. is a predefined constant indicating the difficulty level. The present invention uses the embedding matrix to represent the set of problem difficulty embeddings, where is the embedding matrix Dimension.

[0060] The specific steps are as follows:

[0061] Step 1: Obtain the embedding vector of the problem with inherent attribute enhancement through the inherent attribute enhancement embedding module and the embedding vector of the question-answer pair . In early studies, the embedding of the knowledge points contained in the question is usually used to represent the embedding of the question. This method is too simple and ignores a lot of important information. Recent studies have adopted various methods to integrate additional information from the question. In fact, concatenation and element-by-element summation have been shown to effectively integrate the information of the question. The present invention uses a combination of these two methods to obtain the embedding of the question-answer pair with enhanced inherent properties. The specific calculation formula is as follows:

[0062] ;

[0063] ;

[0064] in, It is a splicing operation. is an element-wise sum operation. , , , and The embedding vectors represent the original question, knowledge point, answer, statistical difficulty of the question, and question type respectively.

[0065] Step 2: Use the dynamic similarity calculation module to obtain the attention mechanism for perceiving the similarity of answering time. After obtaining the embedding vector with enhanced inherent attributes, the present invention further explores the similarity of the dynamic attributes of the questions.

[0066] first, and All are self-attention aggregated to generate context-aware question embedding vectors and context-aware question answering pair embedding vectors In self-attention, the input is generated by three linear transformations: (Query), (key) and (value), then through get and :

[0067] ;

[0068] ;

[0069] in, It is the FS attention mechanism proposed in this paper. , and yes The query, key and value obtained after linear transformation. , and yes The query, key and value obtained after linear transformation.

[0070] Unlike inherent attributes, answering time can take into account individual factors of students and thus measure the similarity of questions. For example, if a student has encountered some questions recently, his answering time on these questions is usually shorter and the accuracy is higher. However, it is difficult to explain the similarity of these problems through inherent attributes. Answering time can identify this potential similarity. Previous studies have not explicitly explored the similarity of answering time for questions. In the present invention, a dynamic similarity computing (DSC) module with integrated attention mechanism is designed to address this limitation. Inspired by AKT's monotonic attention mechanism, which takes into account the time decay effect, the present invention adopts the following method to calculate the attention score :

[0071] ;

[0072] in, denote the query matrix, key matrix and value matrix respectively, is the parameter that controls the time decay rate, is the time distance between two steps, is the scaling factor, Represents the dimension of the key vector. This method improves the performance of the model by conforming to the forgetting curve theory.

[0073] Because before students answer questions, students The answering time of questions answered at the same time is unknown. The present invention calculates the average answering time of different questions in the training set as the student's average answering time. Answer time embedding for questions answered at any time :

[0074] ;

[0075] where i is the index of the i-th student, is the set of questions answered by the i-th student, is the time it takes for the i-th student to answer the j-th question. Questions answered at the moment provide reasonable time embedding.

[0076] In order to measure students' Moment and The difference in answering time between the two moments, the present invention takes into account their embedding and The present invention converts this difference into a similarity score through nonlinear transformation (specifically, feed-forward neural network FFN can be used) , as follows:

[0077] ;

[0078] represents the sigmoid activation function, represents the first weight matrix, Represents the first bias term, similarity score represents the similarity between two questions based on answer time, which will be used to adjust students' Moment and The attention coefficient of the question answered at any time is calculated as follows:

[0079] ;

[0080] represents the query answered by the student at time t, express The key of time, represents time t and the temporal distance between moments;

[0081] Get a new attention mechanism:

[0082] ;

[0083] in, represents the weighted context vector, represents element-wise multiplication, is the similarity score The traditional attention mechanism linearly transforms the input embeddings and computes the similarity by the dot product of the query embedding and the key embedding. This general approach has some limitations in capturing all the characteristics of the questions. The answer time similarity is calculated outside the transformer head. This out-of-head attention mechanism gives the FSAttention mechanism the ability to adjust the question similarity based on the dynamic properties of the answer time similarity.

[0084] Subsequently, the present invention extends this attention mechanism to a multi-head attention mechanism to achieve a more comprehensive representation in the aggregation process. The multi-head attention mechanism concatenates these results and applies further linear transformations as follows:

[0085] ;

[0086] ;

[0087] Here, , and are the original query, key, and value entered, respectively. represents the i-th attention head, represents a multi-head attention operation, Represents a splicing operation, represents the weight matrix, Represents the number of attention heads. , and It is The complete Transformer model requires multi-head attention to go through the addition and normalization process.

[0088] Based on this, the FS attention layer (FSTransformerLayer) used in the entire technical solution of the invention is defined as follows:

[0089] ;

[0090] Representation layer normalization is a regularization method used in neural networks, which aims to improve the training effect and stability of the network.

[0091] Step 3: The present invention designs a difficulty correction (DC) module, which can automatically adjust the difficulty of the question according to the individual learning history of the student, so as to obtain the attention score of subjective difficulty perception. Specifically, the present invention proposes a novel knowledge extractor that combines the cross-attention mechanism and the self-attention mechanism in the difficulty correction (DC) module. In order to consider the difficulty of the problem more comprehensively, the present invention designs two types of DC modules, namely the first difficulty correction module DC1 and the second difficulty correction module DC2, which alternately use subjective relative difficulty and statistical difficulty to adjust the knowledge state.

[0092] In the first difficulty correction module DC1, the student's prior knowledge state is taken into account Embedding vectors for problems with inherent property enhancement Specifically, the present invention is from Minus , and apply nonlinear transformation to obtain subjective difficulty perception. This process can be expressed as:

[0093] ;

[0094] Here, i represents the index of the number of layers of the knowledge extractor. In addition, the present invention designs an output gate to control the update of the knowledge state, as follows:

[0095] ;

[0096] .

[0097] In the second difficulty correction module DC2, the present invention adjusts the difficulty information towards the statistical difficulty to ensure that it does not deviate too far from the objective difficulty. A nonlinear transformation is performed to represent the statistical difficulty information. Subsequently, the present invention also uses a gating mechanism To control the update of knowledge status, as follows:

[0098] ;

[0099] ;

[0100] .

[0101] By alternating between removing and retaining operations, the output knowledge state will contain the difficulty of the questions that are closely related to the learner's current knowledge level. This alignment allows subsequent transformation layers to focus on the question similarity information contained in the subjective and objective difficulty, thereby enhancing the performance of the model.

[0102] Step 4: The prediction module is as follows:

[0103] At time t, the present invention obtains the knowledge state from the knowledge extractor Then, and question embedding Concatenate and input into a fully connected layer, use the sigmoid activation function to generate the output , which represents the probability that the student correctly answers the current question. The specific mathematical expression is as follows:

[0104] ;

[0105] The objective function is to calculate the probability of the correct answer predicted With the true label The cross entropy loss between them is determined, and the specific expression is as follows:

[0106] .

[0107] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0108] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A knowledge tracking method based on fine-grained similarity calculation, characterized in that: include: Embed the questions and students’ answers, and obtain the context-aware question embedding vector and question-answer pair embedding vector through self-attention aggregation; The context-aware question embedding vector and the question-answer pair embedding vector are used through a cross-attention mechanism to obtain the student's initial knowledge state, and multiple difficulty corrections are performed based on subjective relative difficulty and statistical difficulty to obtain the student's knowledge state. The difficulty correction is achieved through a difficulty correction module; the difficulty correction module combines the cross-attention mechanism and the self-attention mechanism, and can automatically adjust the difficulty of the question according to the individual learning history of the student, so as to obtain the student's knowledge state; the difficulty correction module includes a first difficulty correction module DC1 and a second difficulty correction module DC2, DC1 and DC2 are used alternately, and use subjective relative difficulty and statistical difficulty respectively to adjust the student's knowledge state; In DC1, the subjective relative difficulty is obtained by calculating the difference between the student’s knowledge state and the problem embedding, and the output gate is combined to control the update of the student’s knowledge state; In DC2, the embedding vector of the statistical difficulty of the problem is nonlinearly transformed to represent the statistical difficulty, and then an output gate is used to control the update of the student's knowledge state; The student’s knowledge status and question embedding vector are concatenated, and performance prediction is performed to obtain the student’s performance prediction result.

2. The knowledge tracking method based on fine-grained similarity calculation according to claim 1 is characterized in that: The embedding of questions and students' answers specifically includes: By fusing the embedding vector of the problem, the embedding vector of the knowledge point corresponding to the problem, the embedding vector of the statistical difficulty of the problem, and the embedding vector of the type of the problem, we can get the embedding vector of the problem with enhanced inherent attributes. ; By fusing the embedding vector of the question, the embedding vector of the knowledge point corresponding to the question, the embedding vector of the answer corresponding to the question, the embedding vector of the statistical difficulty of the question, and the embedding vector of the question type, we can get the embedding vector of the question-answer pair with enhanced inherent attributes. .

3. The knowledge tracking method based on fine-grained similarity calculation according to claim 2 is characterized in that: The embedding vector of the problem with enhanced inherent attributes is obtained by fusing the embedding vector of the problem, the embedding vector of the knowledge point corresponding to the problem, the embedding vector of the statistical difficulty of the problem, and the embedding vector of the type of the problem. , specifically including: Embedding the problem into a vector The embedding vector of the knowledge point corresponding to the question Add together the embedding vector of the statistical difficulty of the problem and the embedding vector of the type of the problem Add and then concatenate to get the embedding vector of the problem with inherent attribute enhancement .

4. The knowledge tracking method based on fine-grained similarity calculation according to claim 2 is characterized in that: The embedding vector of the question-answer pair with enhanced inherent attributes is obtained by fusing the embedding vector of the question, the embedding vector of the knowledge point corresponding to the question, the embedding vector of the answer corresponding to the question, the embedding vector of the statistical difficulty of the question, and the embedding vector of the type of the question. , specifically including: Embedding the problem into a vector , the embedding vector of the knowledge point corresponding to the question Embedding vector of the answer corresponding to the question Add together the embedding vector of the statistical difficulty of the problem and the embedding vector of the type of the problem Add and then concatenate to get the embedding of the question answer pair with enhanced intrinsic attributes .

5. The knowledge tracking method based on fine-grained similarity calculation according to claim 1 is characterized in that: The context-aware question embedding vector and question-answer pair embedding vector obtained through self-attention aggregation specifically include: After embedding the questions and students’ answers, we get the question embedding vectors with enhanced inherent attributes: and intrinsic property-enhanced question answering pair embedding vectors ; Students in Always answer questions and The similarity of the questions answered at the same time in terms of answering time , by Always answer questions and Embedding the answer time of questions answered at any time and After subtraction, it is obtained through a nonlinear transformation; the similarity Integrate it into the self-attention mechanism to obtain the attention coefficient , based on the attention coefficient, the problem embedding vectors enhanced by the inherent attributes are respectively and intrinsic property-enhanced question answering pair embedding vectors Perform self-attention aggregation to obtain context-aware question embedding vector and context-aware question answering pair embedding vectors : ； ； for The value in the self-attention aggregation at the moment; in the self-attention aggregation, the query, key, and value are all obtained by linearly transforming the same input.

6. The knowledge tracking method based on fine-grained similarity calculation according to claim 5 is characterized in that: The context-aware question embedding vector and the question-answer pair embedding vector are used to obtain the student's initial knowledge state through a cross-attention mechanism, specifically including: Embedding context-aware questions into vectors and context-aware question answering pair embedding vectors , after cross attention, we get the student's initial knowledge state; among them, After linear transformation, we obtain the query and key required for cross attention. After linear transformation, the value required for cross attention is obtained.

7. The knowledge tracking method based on fine-grained similarity calculation according to claim 1 is characterized in that: In the first difficulty correction module DC1, the student's prior knowledge state is taken into account Embedding vectors for problems with inherent attribute enhancement The difference between Minus , and apply nonlinear transformation to obtain subjective relative difficulty, and use subjective relative difficulty to adjust students' knowledge status: ； represents the second weight matrix, represents the knowledge state of DC1 input to the i-th layer, The embedding vector representing the problem with inherent attribute enhancement, represents the second bias term, is a vector representing subjective difficulty perception, represents the activation function; An output gate is designed to control the update of the student's knowledge state: ； ； represents the output gate of DC1, represents the sigmoid activation function, represents the third weight matrix, represents the third bias term, Represents the knowledge state of the DC1 output of the i-th layer.

8. The knowledge tracking method based on fine-grained similarity calculation according to claim 1 is characterized in that: In the second difficulty correction module DC2, the difficulty information is adjusted to the statistical difficulty; the embedding vector of the statistical difficulty of the problem A nonlinear transformation is performed to represent the statistical difficulty, and then an output gate is used to control the update of the knowledge state: ； ； ； represents the fourth weight matrix, represents the knowledge state of the DC2 output of the i-th layer, The embedding vector representing the problem with inherent attribute enhancement, represents the fourth bias term, is a vector representing the statistical difficulty, represents the activation function; represents the fifth weight matrix, represents the fifth bias term, Represents the output gate of DC2.

9. A knowledge tracking system based on fine-grained similarity calculation, characterized in that: include: The embedding aggregation module embeds questions and students’ answers, and obtains context-aware question embedding vectors and question-answer pair embedding vectors through self-attention aggregation; The state acquisition module uses the cross-attention mechanism to obtain the student's initial knowledge state through the background-aware question embedding vector and the question-answer pair embedding vector, and performs multiple difficulty corrections based on the subjective relative difficulty and the statistical difficulty to obtain the student's knowledge state; the difficulty correction is implemented through the difficulty correction module; the difficulty correction module combines the cross-attention mechanism and the self-attention mechanism, and can automatically adjust the difficulty of the question according to the student's individual learning history, so as to obtain the student's knowledge state; the difficulty correction module includes a first difficulty correction module DC1 and a second difficulty correction module DC2, DC1 and DC2 are used alternately, and the subjective relative difficulty and the statistical difficulty are used to adjust the student's knowledge state respectively; in DC1, the subjective relative difficulty is obtained by calculating the difference between the student's knowledge state and the problem embedding, and the output gate is combined to control the update of the student's knowledge state; in DC2, the embedding vector of the statistical difficulty of the problem is nonlinearly transformed to represent the statistical difficulty, and then an output gate is used to control the update of the student's knowledge state; The performance prediction module concatenates the student’s knowledge status and question embedding vector, and performs performance prediction to obtain the student’s performance prediction result.

Images