Intelligent q matrix generation method, system and terminal fusing topic semantic information

Abstract

The application belongs to the field of artificial intelligence and educational interactive technology, and discloses an intelligent Q matrix generation method, system and terminal fusing question semantic information, which uses BiLSTM+Transformer to obtain mathematical knowledge points in a data set, and uses CRF to correct the annotation type of each word to output the mathematical knowledge points; uses BiGRU+Attention mechanism to extract the relationship of the knowledge points, and outputs the relationship between each knowledge point pair; uses the relationship between the knowledge points to design a redundant knowledge point reduction algorithm, reduces the redundant knowledge points through the reduction algorithm, and generates a Q matrix using the knowledge point set after reduction. The application can achieve a relatively high knowledge point extraction rate by using the Transformer+BiLSTM+CRF method, and can output the relationship between the knowledge point pairs more accurately by using the BiGRU+Attention method.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and educational interaction technology, and in particular relates to a method, system and terminal for generating an intelligent Q matrix that integrates semantic information of questions. Background Technology

[0002] Currently, the Transformer+BiLSTM+CRF algorithm, where Transformer (Transformer), BiLSTM (Bidirectional Long Short-Term Memory), and CRF (Conditional Random Field) are used for text annotation, combines these three algorithms. This combined algorithm leverages the advantages of each to form a novel approach. It is primarily used in natural language processing. The algorithm consists of three layers: Transformer, BiLSTM, and CRF. In the Transformer layer, global input-output dependencies can be extracted without considering the distance between outputs. The Transformer layer outputs text features with local dependencies, which are then passed through the BiLSTM layer to obtain a direction-sensitive text feature extraction sequence. Finally, the CRF layer annotates the words.

[0003] The BiGRU+Attention algorithm combines the features of BiGRU (Bidirectional Gated Recurrent Unit) and Attention (Attention Mechanism). It consists of two layers: a BiGRU layer and an Attention layer. In the BiGRU layer, each output depends on the state of the previous and next time steps. The GRU comprises two parts: an update gate and a reset gate. The update gate determines the impact of the previous output on the current state, while the reset gate decides whether to ignore the impact. During knowledge extraction, each output is related to the GRU outputs of the previous and next time steps, and each input simultaneously provides values ​​to both GRUs in different directions. The BiGRU output then passes through the Attention layer. In the Attention layer, important information in the input sequence is emphasized and given higher weights, resulting in a sequence of important information. Finally, this sequence is used to output the relationships between relational entities.

[0004] Q-matrix. The Q-matrix generation method is an important learning resource modeling method, and the Q-matrix is ​​often used as a preprocessing part in cognitive diagnosis to model the knowledge points covered by each question.

[0005] The Redundant Knowledge Points Reduction Algorithm is designed based on the relationships between the aforementioned entities. It consists of four parts: RSS algorithm, RBA algorithm, PC algorithm, and Q-matrix generation method. After processing the knowledge point set through these four parts, an interpretable and effective Q-matrix is ​​finally generated.

[0006] Current methods for generating Q-matrices typically require domain knowledge from education experts to manually annotate the Q-matrices. This approach leads to significant annotation errors and makes it difficult to label all knowledge points.

[0007] Ignoring the connections between questions leads to redundancy in the knowledge points used to construct the Q-matrix. Similar questions often test similar knowledge points, and two questions testing the same knowledge points are often quite similar. This results in poor predictive performance of the Q-matrix.

[0008] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0009] (1) Existing Q matrix generation methods generally rely on experts’ domain knowledge for manual annotation, which leads to high annotation error and difficulty in annotating all knowledge points.

[0010] (2) Existing Q matrix generation methods ignore the connection between problems. Similar problems often examine similar knowledge points, and two problems that examine the same knowledge points are often quite similar, resulting in poor prediction performance of the constructed Q matrix.

[0011] (3) Existing Q-matrix generation methods generally generate only one Q-matrix. However, considering that the intensity of each type of question is different, such as multiple choice, fill-in-the-blank and problem-solving questions, using a single type of Q-matrix to predict different types of questions will lead to a large error. Summary of the Invention

[0012] To address the problems existing in the prior art, this invention provides an intelligent Q-matrix generation method, system, and terminal that integrates semantic information of the questions.

[0013] This invention is implemented as follows: a method for generating an intelligent Q-matrix that integrates semantic information of the question, the method comprising:

[0014] The algorithm uses a knowledge point extraction method to obtain mathematical knowledge points from the dataset, and uses CRF to correct the labeling type of each word to output the mathematical knowledge points. It uses a relation extraction mechanism to extract relations from the knowledge points and outputs the relationship between each knowledge point pair. It designs a redundant knowledge point reduction algorithm based on the relationship between knowledge points, reduces redundant knowledge points, and uses the reduced knowledge point set to generate three types of Q matrices.

[0015] Furthermore, BiLSTM+Transformer is used to obtain mathematical knowledge points from the dataset, and CRF is used to correct the labeling type of each word to output the mathematical knowledge points; BiGRU+Attention mechanism is used to extract relationships between knowledge points, outputting the relationships between each knowledge point pair; a redundant knowledge point pruning algorithm is designed based on the relationships between knowledge points, and redundant knowledge points are pruned using the pruning algorithm, and three types of Q matrices are generated using the pruned knowledge point set: Q choice Q blank and Q comprehensive .

[0016] Furthermore, the intelligent Q-matrix generation method that integrates semantic information of the questions includes the following steps:

[0017] Step 1: Use the word2vec method to represent each word that appears in the dataset, and use BiLSTM+Transformer+CRF to extract the corresponding knowledge points;

[0018] Step 2: Use the BiGRU+Attention combination mechanism to extract relations for each knowledge point pair;

[0019] Step 3: Design a redundant knowledge point reduction algorithm, and use the algorithm to eliminate redundant knowledge points to generate a Q matrix.

[0020] Furthermore, in step one, the vector representing each word is passed through a Transformer layer, which includes an encoding part and a decoding part; each word vector obtains a high-dimensional representation vector through the encoding part, and the representation vector is then passed through the decoding part to obtain character-level information; the character-level information is passed through a BiLSTM layer to output the annotation type of each word.

[0021] The annotation method used is BMES, which includes four types: B-MAT, M-MAT, E-MAT, and O. B-MAT represents the first word of a knowledge point, M-MAT represents all the words in the middle of a knowledge point, E-MAT represents the last word of a knowledge point, and O indicates that the word does not belong to a certain knowledge point phrase. All words and their corresponding annotation types are corrected by the CRF layer, and finally the knowledge points corresponding to each exercise are output.

[0022] Furthermore, in step two, the knowledge points obtained in step one are processed through a BiGRU+Attention layer. The implementation of BiGRU includes the following sub-steps:

[0023] (1) The model's input passes through a BiGRU layer; the BiGRU layer model consists of two GRUs operating in opposite directions, each GRU consisting of an update gate and a reset gate. The update gate controls the influence of the previous output on the current input; the larger the value in the update gate, the greater the influence on the current input. The reset gate determines whether to retain the previous output value. Simultaneously, when extracting knowledge point relationships, the model considers the state of the knowledge point at the previous time step and the state of the knowledge point at the next time step to extract the relationship between each knowledge point pair. Relationships in the dataset include dependency, being depended upon, belonging to, being belonged to, synonyms, antonyms, near-synonyms, attributes, being attributed to, appositives, other, and no relationship.

[0024] (2) The output of step (1) is used as the input of the Attention layer. The model is represented as a layer in the neural network, which is used to add corresponding weights to meaningful word groups so that the word groups are noticed by the neural network.

[0025] Furthermore, the redundant knowledge point reduction algorithm in step three includes: algorithm RSS for deleting knowledge points with relationships of synonyms and near-synonyms; algorithm RBA for deleting knowledge points with relationships of belonging, being belonging to, being affixed to, and being attributed; and algorithm PC for deleting knowledge points with relationships of dependence, being depended upon, and antonyms.

[0026] Furthermore, the construction of the redundant knowledge point reduction algorithm in step three includes:

[0027] (1) The RSS algorithm is used to reduce redundant knowledge points with relationships in the dataset. The RSS algorithm is based on the following rule: if knowledge point A has a similar or identical knowledge point C, and A and C appear in the same dataset, then A is a redundant knowledge point. A list is used as the data structure of the RSS algorithm to remove redundant knowledge points. In the RSS algorithm, a list is created for each word with the same meaning. The list contains synonyms and near-synonyms of all words. Redundancy is removed based on the list of synonyms and near-synonyms.

[0028] (2) Redundant knowledge points are removed using the RBA algorithm. The RBA algorithm focuses on resolving redundant knowledge points whose relationships are "belong to", "belong to", "attribute", and "attribute to". The RBA algorithm is based on the following rules:

[0029] • If knowledge point A belongs to knowledge point B, then A is a subset of B; if A and B both appear in previous knowledge points, then A is a redundant knowledge point.

[0030] • If A is an attribute of B, then A is also considered a subset of B; when A and B appear in the dataset at the same time, A is a redundant knowledge point.

[0031] The RBA algorithm consists of two parts: a parent node list and a node set. A parent node is defined as follows: if A belongs to B, then B is A's parent node and A is B's child node; if A is an attribute of B, then B is also A's parent node and A is also B's child node. A node set is constructed for each parent node: the first element is the parent node, and the other elements are its child nodes, with all child nodes considered redundant. A parent node list is then created for all parent nodes, containing all their parent nodes.

[0032] (3) Use the PC algorithm to remove redundant knowledge points; the PC algorithm is used to remove redundant knowledge point relationships that are dependencies, dependents, and antonyms, and the rules of the PC algorithm are as follows:

[0033] • If A depends on B, A and B appear as a joint knowledge point in the knowledge dataset: (A, B); if B depends on A, the result is the same as if A depends on B.

[0034] If the meanings of A and B are opposite, A and B will share a parent node, denoted as AB, and the parent node will replace A and B in the dataset.

[0035] The PC algorithm consists of a forest of binary trees; when A depends on B, a tree is created with A and B as child nodes and (A,B) as the parent node. (A,B) If B depends on A, then the tree is similar to a tree. (A,B)When considering multiple dependencies, knowledge point C depends on one of the knowledge point sets (A, B); a binary tree is constructed using the joint knowledge point set (A, B) and C. (A,B,C) tree (A,B,C) Replace tree (A,B) .

[0036] (4) Generate a Q matrix using the reduced set of knowledge points; take the number of knowledge points as the dimension of the Q matrix, with each row representing a middle school math problem; mark the corresponding position in the Q matrix with 1 according to the knowledge points involved in the math problem, and mark 0 for knowledge points not involved; the generated Q matrix is: Q matrix of multiple choice questions - Q choice Fill in the blanks: Q matrix - Q blank The Q-matrix of the comprehensive problem - Q comprehensive .

[0037] Another objective of this invention is to provide an intelligent Q-matrix generation system that integrates question semantic information using the aforementioned intelligent Q-matrix generation method. The intelligent Q-matrix generation system integrating question semantic information includes:

[0038] The knowledge point extraction module is used to represent each word that appears in the dataset using the word2vec method, and to extract the corresponding knowledge points using BiLSTM+Transformer+CRF.

[0039] The relation extraction module is used to extract relations for each knowledge point pair using a combination mechanism of BiGRU and Attention.

[0040] The Q-matrix generation module is used to design a redundant knowledge point reduction algorithm, and to use the redundant knowledge point reduction algorithm to eliminate redundant knowledge points and generate a Q-matrix.

[0041] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the step of generating an intelligent Q matrix that integrates semantic information of the questions.

[0042] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the intelligent Q-matrix generation method that integrates semantic information of the questions.

[0043] Another objective of this invention is to provide an information data processing terminal, which is used to implement the intelligent Q-matrix generation system that integrates semantic information of the questions.

[0044] Based on the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solution to be protected by this invention from the following aspects:

[0045] First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0046] The intelligent Q-matrix generation method and system integrating semantic information of questions provided by this invention are based on the middle school mathematics section of the development dataset from East China Normal University. Developed using Python, the dataset contains three types of mathematical questions: multiple choice, fill-in-the-blank, and comprehensive questions. The Q-matrix generation method provided by this invention mainly consists of three parts: the first part uses BiLSTM+Transformer to obtain mathematical knowledge points from the dataset and uses CRF to correct the labeling type of each word to output the mathematical knowledge points; the second part uses BiGRU+Attention mechanism to extract relationships between these knowledge points and output the relationships between each knowledge point pair; the final part uses the relationships between these knowledge points to design a redundant knowledge point reduction algorithm, which reduces redundant knowledge points, and then uses the reduced knowledge point set to generate three types of Q-matrices: Q... choice Q blank and Q comprehensive .

[0047] 1. The Transformer+BiLSTM+CRF method is used to extract knowledge points from the corpus of middle school mathematics knowledge points. This approach considers the influence of the previous time step on the present time step, as well as the contextual information of the knowledge point's location, thus achieving a relatively high knowledge point extraction rate.

[0048] 2. The BiGRU+Attention method is used to extract the relationship between each knowledge point and its corresponding counterpart. This approach can store the features between knowledge point pairs over a period of time, and rely on the features from the previous time and the features from the next time to output the relationship between knowledge point pairs more accurately.

[0049] 3. An algorithm for reducing redundant knowledge points is employed. This method is based on the relationships between knowledge point pairs. It removes redundant knowledge points, generating an improved set of knowledge points. Finally, an efficient Q-matrix is ​​generated based on this improved set of knowledge points.

[0050] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0051] This invention uses the Transformer+BiLSTM+CRF algorithm to acquire all knowledge points in a corpus of middle school mathematics knowledge points. Then, it employs the BiGRU+Attention algorithm to extract the relationships between knowledge point pairs. Finally, it uses a redundant knowledge point removal algorithm to eliminate all redundant knowledge points. The key focus of this invention is the removal of redundant knowledge points and the generation of an interpretable and efficient Q-matrix.

[0052] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0053] The technical solution of this invention fills a technological gap in the industry both domestically and internationally:

[0054] Currently, most traditional Q-matrix generation techniques in China rely on manually labeling the Q-matrix using the domain knowledge of education experts. This traditional method requires significant manpower and time to generate the Q-matrix, and its large labeling error and low knowledge point labeling rate greatly reduce the interpretability and usability of the Q-matrix. Furthermore, traditional Q-matrix generation methods do not consider the redundant knowledge points contained in secondary school mathematics problems and the connections between these knowledge points, nor do they consider the different levels of emphasis in different types of secondary school mathematics problems, resulting in less than ideal cognitive diagnostic results for Q-matrixes generated using traditional methods. Therefore, this invention aims to fill the gaps in domestic and international Q-matrix design methods by designing a Q-matrix generation technique that combines question type, text information, and a Reduction redundant knowledge points algorithm. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the intelligent Q-matrix generation method that integrates semantic information of the questions provided in this embodiment of the invention;

[0057] Figure 2 This is a schematic diagram of the intelligent Q-matrix generation method that integrates semantic information of the questions, provided in an embodiment of the present invention.

[0058] Figure 3This is a schematic diagram of the RBA algorithm structure provided in an embodiment of the present invention;

[0059] Figure 4 This is a schematic diagram of the PC algorithm structure provided in an embodiment of the present invention;

[0060] Figure 5 This is a schematic diagram of the tree creation process provided in an embodiment of the present invention;

[0061] Figure 6 The tree provided in the embodiments of the present invention (A A structural diagram of -B);

[0062] Figure 7 This is a schematic diagram of the Transformer+BiLSTM+CRF algorithm provided in an embodiment of the present invention;

[0063] Figure 8 This is a schematic diagram of the BiGRU+Attention algorithm provided in an embodiment of the present invention;

[0064] Figure 9 This is a schematic diagram of the redundant knowledge point reduction algorithm provided in an embodiment of the present invention;

[0065] Figure 10 Q is provided in the embodiments of the present invention. choice Matrix result display diagram;

[0066] Figure 11 Q is provided in the embodiments of the present invention. blank Matrix result display diagram;

[0067] Figure 12 Q is provided in the embodiments of the present invention. comprehensive Matrix results display diagram. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] To address the problems existing in the prior art, this invention provides an intelligent Q-matrix generation method, system, and terminal that integrates semantic information of the questions. The invention will be described in detail below with reference to the accompanying drawings.

[0070] I. Explanatory and Illustrative Embodiments. To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory and illustrative description of the embodiments described in the claims.

[0071] like Figure 1As shown, the intelligent Q-matrix generation method that integrates semantic information of the questions provided in this embodiment of the invention includes the following steps:

[0072] S101 uses BiLSTM+Transformer to obtain mathematical knowledge points from the dataset, and uses CRF to correct the labeling type of each word to output the mathematical knowledge points.

[0073] S102 uses the BiGRU+Attention mechanism to extract the relationship between knowledge points, outputs the relationship between each knowledge point pair, and designs a redundant knowledge point deletion algorithm based on the relationship between knowledge points.

[0074] S103, redundant knowledge points are removed using a pruning algorithm, and the resulting set of knowledge points is used to generate three types of Q matrices: Q choice Q blank and Q comprehensive .

[0075] As a preferred embodiment, such as Figure 2 As shown, the intelligent Q-matrix generation method that integrates semantic information of the questions provided in this embodiment of the invention specifically includes the following steps:

[0076] Step 1: Use the word2vec method to represent each word that appears in the dataset, and use BiLSTM+Transformer+CRF to extract the corresponding knowledge points.

[0077] Step 2: Use the BiGRU+Attention combination mechanism to extract relations for each knowledge point pair.

[0078] Step 3: A redundant knowledge point reduction algorithm was designed to reduce redundant knowledge points in the dataset. This algorithm consists of four parts: the Reduction Similar and SameContent Algorithm (RSS) for knowledge points with synonyms and near-synonyms; the Reduction Belong-Attribute Algorithm (RBA) for knowledge points with belonging, being belonged to, attributes, and being attributed to; and the Parent-Child Algorithm (PC) for knowledge points with dependency, being depended upon, and antonyms. These algorithms are then used to eliminate redundant knowledge points, thereby generating an efficient Q-matrix.

[0079] In a preferred embodiment, step 1 of this invention involves passing the vector representing each word through a Transformer layer, which includes an encoding part and a decoding part. Each word vector, through the encoding part, obtains a high-dimensional representation vector, which is then passed through the decoding part to obtain character-level information. This character-level information is then passed through a BiLSTM layer to output the annotation type for each word. The annotation type method used here is BMES, which has four types: B-MAT, M-MAT, E-MAT, and O. B-MAT represents the first word of a knowledge point, M-MAT represents all words in the middle of a knowledge point, E-MAT represents the last word of a knowledge point, and O indicates that the word does not belong to a certain knowledge point phrase. Then, all words and their corresponding annotation types are corrected by a CRF layer, and finally, the knowledge point corresponding to each exercise is output.

[0080] In a preferred embodiment, step 2 of this invention involves passing the knowledge points obtained in step 1 through a BiGRU+Attention layer. The specific implementation of BiGRU in step 2 includes the following sub-steps:

[0081] Step 2.1: The input to this model first passes through a BiGRU layer. This layer consists of two GRUs operating in opposite directions. Each GRU comprises an update gate and a reset gate. The update gate controls the impact of the previous output on the current input. A larger value in the update gate has a greater impact on the current input, while the reset gate determines whether to retain the previous output value. Simultaneously, when extracting knowledge point relationships, the model considers the state of the knowledge point at the previous time step and the state of the knowledge point at the next time step to extract the relationship between each knowledge point pair. In this dataset, relationships are categorized into 12 types, as shown in Table 1.

[0082] Table 1 shows the relation classifications in the dataset.

[0083]

[0084]

[0085] Step 2.2: Then, use the output of Step 2.1 as the input to the Attention layer. This layer is typically represented as a layer in a neural network, used to add appropriate weights to meaningful phrases, making these phrases more "attentional" to the neural network.

[0086] In a preferred embodiment, step 3 of this invention utilizes the relationship obtained in step 2 to design a redundant knowledge point reduction algorithm. This algorithm consists of four steps:

[0087] Step 3.1: Use the RSS algorithm to remove redundant knowledge points with similar or identical relationships in the dataset. This algorithm is based on the following rule: if knowledge point A has a similar or identical knowledge point C, and A and C both appear in the same dataset, then A is a redundant knowledge point. To remove redundant knowledge points, lists are used as the data structure for this algorithm. In this algorithm, a list is created for each word with the same meaning, containing all synonyms and near-synonyms of that word. Redundancy is then removed based on this set of synonyms and near-synonyms.

[0088] Step 3.2: Use the RBA algorithm to remove redundant knowledge points. This algorithm focuses on resolving redundant knowledge points with relationships of "belonging," "belonging to," "attribute," and "attributed to." The algorithm is based on the following rules:

[0089] If knowledge point A belongs to knowledge point B, then A is a subset of B. Therefore, if A and B both appear in previous knowledge points, A is a redundant knowledge point.

[0090] If A is an attribute of B, then A can also be considered a subset of B. Therefore, when A and B both appear in the dataset, A is a redundant knowledge point.

[0091] The algorithm consists of two parts: a list of parent nodes and a set of nodes. A parent node is defined as follows: if A belongs to B, then B is A's parent node, and A is B's child node. If A is an attribute of B, then B is also A's parent node, and A is also B's child node. Simultaneously, a set of nodes is constructed for each parent node: the first element is the parent node, and the other elements are its child nodes, with all child nodes considered redundant. A list of parent nodes is created for all parent nodes, containing all their parent nodes, see [link to relevant documentation]. Figure 3 .

[0092] Step 3.3: Use the PC algorithm to remove redundant knowledge points. This algorithm aims to remove redundant knowledge point relationships that are dependencies, dependents, and antonyms. The rules of the PC algorithm are as follows:

[0093] If A depends on B, A and B can appear as a joint knowledge point in the knowledge dataset: (A, B). If B depends on A, the result is the same as if A depends on B.

[0094] If A and B have opposite meanings, A and B will share a parent node, denoted as AB. The parent node will then replace A and B in the dataset.

[0095] The algorithm consists of a forest of binary trees. When A depends on B, a tree is created with A and B as child nodes and (A, B) as their parent node: tree (A,B) Its structure is shown in Figure 4 If B depends on A, then the tree is similar to a tree. (A,B) Meanwhile, when considering multiple dependencies, i.e., knowledge point C depends on one of the knowledge point sets (A, B), the process of creating the tree is similar to the previous one: using the joint knowledge point set (A, B) and C to form a binary tree tree. (A,B,C) Then tree (A,B,C) Replace tree (A,B) This process is like Figure 5 As shown. Also, when A and B are antonyms, tree (A-B) See the structure Figure 6 .

[0096] Step 3.4: Generate the Q-matrix using the reduced set of knowledge points. The method for generating the Q-matrix is ​​based on the following idea: take the number of knowledge points as the dimension of the Q-matrix, with each row representing a middle school math problem. Then, according to the knowledge points involved in the math problem, mark the corresponding position in the Q-matrix with 1, and mark the unrelated knowledge points with 0. In this method, the present invention needs to generate three types of Q-matrixes: Q-matrix for multiple-choice questions, Q-matrix for... choice Regarding the Q-matrix for fill-in-the-blank questions - Q blank Regarding the Q-matrix for comprehensive problems - Q comprehensive .

[0097] The intelligent Q-matrix generation system that integrates semantic information of questions provided in this embodiment of the invention includes:

[0098] The knowledge point extraction module is used to represent each word that appears in the dataset using the word2vec method, and to extract the corresponding knowledge points using BiLSTM+Transformer+CRF.

[0099] The relation extraction module is used to extract relations for each knowledge point pair using a combination mechanism of BiGRU and Attention.

[0100] The Q-matrix generation module is used to design a redundant knowledge point reduction algorithm, and to use the redundant knowledge point reduction algorithm to eliminate redundant knowledge points and generate a Q-matrix.

[0101] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, uses a corpus of middle school mathematics knowledge points as the dataset to illustrate the intelligent Q-matrix generation method that integrates semantic information of questions provided in this embodiment of the invention.

[0102] Step 1: Input a corpus of middle school math knowledge points. First, convert all words in the corpus into corresponding vectors, and then use these word vectors as input to the Transformer layer. The Transformer layer consists of two parts: an encoding part and a decoding part. The word vectors, after encoding and decoding, become knowledge point information usable by BiLSTM. This information then passes through the BiLSTM layer, where the BiLSTM layer combines the context of the knowledge point information with forward and backward LSTM algorithms to output word labels for each word. These word labels are then corrected by a CRF layer, finally outputting each knowledge point and its corresponding word label. The process is as follows: Figure 7 As shown in Table 2, the performance of Transformer+BiLSTM+CRF surpasses that of traditional knowledge extraction methods such as TextRank and TF-IDF.

[0103] Table 2 Knowledge Point Extraction Results

[0104]

[0105] Step 2.1: Use the knowledge points obtained in Step 1 as input to the BiGRU+Attention algorithm, which consists of two layers: a BiGRU layer and an Attention layer. In the BiGRU layer, each knowledge point provides values ​​to two GRUs in different directions simultaneously, and finally outputs the corresponding value to the Attention layer based on the previous and next states of each value.

[0106] Step 2.2: Use the values ​​from Step 2.1 as input to the Attention layer. Meaningful phrases in the Attention layer represent the mathematical knowledge points contained in the middle school math problem. These phrases will be assigned relatively high weights, and combined with the features of the extracted text, the final output will be the category of the corresponding knowledge point pairs. See below for details. Figure 8 Furthermore, BiGRU+Attention outperforms traditional relation extraction models, as shown in Table 3.

[0107] Table 3. Results of the relation extraction model

[0108]

[0109]

[0110] Step 3.1: Use the output of Step 2.2 as the input of the RSS algorithm, and delete the corresponding redundant knowledge points. The results are shown in Table 4.

[0111] Table 4 Results of Redundant Knowledge Point Deletion

[0112]

[0113] Step 3.2: Use the set of knowledge points output in Step 3.1 as input to the RBA algorithm, and delete the corresponding redundant knowledge points. The results are shown in Table 5.

[0114] Table 5 Results of Redundant Knowledge Point Deletion

[0115]

[0116] Step 3.3: Use the set of knowledge points output in Step 3.2 as input to the PC algorithm, and delete the corresponding redundant knowledge points. The results are shown in Table 6.

[0117] Table 6 Results of Redundant Knowledge Point Deletion

[0118]

[0119] Step 3.4: Use the set of knowledge points output in Step 3.3 as the final knowledge point extraction result, and create three corresponding Q matrices based on this result: Q choice Q blank and Q comprehensive The results of these three matrices are referenced. Figure 10 , 11 12. For detailed procedures, please refer to... Figure 9 .

[0120] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating an intelligent Q-matrix by fusing question semantic information, characterized in that, The intelligent Q matrix generation method fusing question semantic information comprises the following steps: The knowledge point extraction method is used to obtain the mathematical knowledge points in the data set, and the CRF is used to correct the label type of each word to output the mathematical knowledge points; the relationship extraction mechanism is used to extract the relationship of the knowledge points, and the relationship between each knowledge point pair is output; the relationship between the knowledge points is used to design a redundant knowledge point reduction algorithm, the redundant knowledge points are reduced through the reduction algorithm, and the knowledge point set after reduction is used to generate three types of Q matrices; The construction of the redundant knowledge point reduction algorithm comprises: (1) redundant knowledge points with relationships in the data set are reduced by using an RSS algorithm, the RSS algorithm is based on the following rules: if a knowledge point A has a knowledge point C with similar or identical meaning, and A and C appear in the same data set at the same time, then A is a redundant knowledge point; a list is used as the data structure of the RSS algorithm, which is used to remove redundant knowledge points; in the RSS algorithm, a list is established for each word with the same meaning, and the list contains all synonyms and near synonyms of the word, and the redundant knowledge points are deleted according to the list set composed of the synonyms and near synonyms; (2) redundant knowledge points are reduced by using an RBA algorithm, the RBA algorithm focuses on solving knowledge points with redundant knowledge relationships belonging to, being belonged to, attribute and being attribute; the RBA algorithm is based on the following rules: · if a knowledge point A belongs to a knowledge point B, then A is a subset of B; if A and B appear in the previous knowledge points at the same time, then A is a redundant knowledge point; · if A is an attribute of B, then A is also considered as a subset of B; when A and B appear in the data set at the same time, A is a redundant knowledge point; The RBA algorithm consists of a parent node list and a node set; the definition of the parent node is that if A belongs to B, B is the parent node of A, and A is the child node of B; if A is an attribute of B, then B is also the parent node of A, and A is also the child node of B; a node set is constructed for each parent node: the first element is the parent node, and the other elements are the child nodes of the parent node, and all the child nodes are considered as redundant; a parent node list is established for all parent nodes, which contains all parent nodes; (3) redundant knowledge points are deleted by using a PC algorithm; the PC algorithm is used to remove redundant knowledge point relationships of dependency, being dependent and antonyms, and the rules of the PC algorithm are as follows: · if A depends on B, A and B appear as a joint knowledge point in the knowledge data set: (A, B); if B depends on A, the result is the same as A depending on B; · if A and B have opposite meanings, A and B will share a parent node, denoted as A-B, and A and B in the data set are replaced by the parent node; The PC algorithm consists of a forest of binary trees; when A depends on B, a tree is created with A and B as child nodes and (A,B) as the parent node. (A,B) If B depends on A, then the tree is similar to a tree. (A,B) When considering multiple dependencies, knowledge point C depends on one of the knowledge point sets (A, B); a binary tree is constructed using the joint knowledge point set (A, B) and C. (A,B,C) tree (A,B,C) Replace tree (A,B) ; (4) Generating Q matrix by using the set of knowledge points after pruning; taking the number of knowledge points as the dimension of Q matrix, each row represents a middle school mathematics problem; according to the knowledge points involved in the mathematics problem, the corresponding position in the Q matrix is marked with 1, and the knowledge points not involved are marked as 0; the generated Q matrix is: Q matrix of single-choice questions-Q choice , Q matrix of fill-in-the-blank questions-Q blank , Q matrix of comprehensive questions-Q comprehensive .

2. The intelligent Q-matrix generation method that fuses question semantics information according to claim 1, characterized in that, The intelligent Q matrix generation method fusing question semantic information comprises the following steps: Step one, using a word vector representation method to represent each word appearing in the data set, and using a knowledge point extraction method to extract the corresponding knowledge points; Step two, using a combination mechanism of a relationship extraction method to extract the relationship of each knowledge point pair; In step three, a redundant knowledge point reduction algorithm is designed, and the redundant knowledge points are eliminated by using the redundant knowledge point reduction algorithm to generate a Q matrix.

3. The intelligent Q-matrix generation method that fuses question semantics information according to claim 2, wherein, BiLSTM+Transformer is used to obtain the mathematical knowledge points in the dataset, and CRF is used to correct the label type of each word to output the mathematical knowledge points; BiGRU+Attention mechanism is used for relation extraction of knowledge points; the redundant knowledge points are deleted through the deletion algorithm, and the knowledge point set after deletion is used to generate three types of Q matrices: Q choice , Q blank and Q comprehensive .

4. The intelligent Q-matrix generation method that fuses question semantics information according to claim 2, wherein, In step one, each word vector is passed through a Transformer layer, which includes an encoding part and a decoding part. Each word vector obtains a high-dimensional representation vector through the encoding part, and the representation vector obtains character-level information through the decoding part. The character-level information is outputted after passing through a BiLSTM layer to output the label type of each word. The label type method adopted is BMES, which includes four types: B-MAT, M-MAT, E-MAT and O. B-MAT represents the first word of a knowledge point, M-MAT represents all words in the middle of a knowledge point, E-MAT represents the last word of a knowledge point, and O represents that the word does not belong to a knowledge point word group. All words and corresponding label types are corrected through a CRF layer, and finally the corresponding knowledge points of each exercise are outputted.

5. The intelligent Q-matrix generation method that fuses question semantics information according to claim 2, wherein, In step two, the knowledge points obtained in step one are passed through a BiGRU+Attention layer. The implementation of the BiGRU includes the following sub-steps: (1) The input of the model is passed through a BiGRU layer. The BiGRU layer model is composed of two GRUs in opposite directions, each of which is composed of an update gate and a reset gate. The update gate is used to control the influence of the previous output on the present input. The greater the value in the update gate, the greater the influence on the present. The reset gate is used to determine whether to retain the value of the previous output. At the same time, when extracting the relationship of knowledge points, the model considers the state of the previous knowledge point and the state of the next knowledge point to extract the relationship of each knowledge point pair. The relationships in the data set include dependency, dependency, belonging, belonging, synonym, antonym, near synonym, attribute, attribute, co-ordinate word, other and no relationship; (2) The output of step (1) is taken as the input of the Attention layer. The model is represented as a layer of neural network, which is used to add corresponding weights to meaningful word groups so that the word groups are paid attention to by the neural network.

6. The intelligent Q-matrix generation method that fuses question semantics information according to claim 2, wherein, The redundant knowledge point reduction algorithm in step three includes an algorithm RSS for deleting knowledge points with relationship of synonym and near synonym, an algorithm RBA for reducing knowledge point relationships of belonging, being belonging, attribute and being attribute, and an algorithm PC for reducing knowledge point relationships of dependency, dependency and antonym.

7. An intelligent Q matrix generation system for fusion of item semantic information, which applies the intelligent Q matrix generation method for fusion of item semantic information according to any one of claims 1 to 6, characterized by The intelligent Q matrix generation system fusing question semantic information includes: A knowledge point extraction module for representing each word appearing in the data set using a word2vec method, and extracting corresponding knowledge points using BiLSTM+Transformer+CRF; A relationship extraction module for extracting the relationship of each knowledge point pair using a combination mechanism of BiGRU+Attention; A Q matrix generation module for designing a redundant knowledge point reduction algorithm, and eliminating redundant knowledge points by using the redundant knowledge point reduction algorithm to generate a Q matrix.

8. A computer device, comprising: The computer device comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the intelligent Q matrix generation method of fusing question semantic information according to any one of claims 1-6. 9.A computer readable storage medium storing a computer program, and the computer program is executed by a processor to make the processor execute the steps of the intelligent Q matrix generation method of fusing question semantic information according to any one of claims 1-6.

Images