Similar problem retrieval method and system relying on fusion of concept dependency relationship

Abstract

The application discloses a kind of fusion concept dependency relationship's similar problem retrieval method and system, by regarding content and concept as two important attributes of problem and introducing multiple dependency relationships between concepts, construct association graph (concept-problem-content multilayer graph), then the multiple dependency relationships between concepts and concept-problem association are comprehensively modeled to mine the implicit path between problem pairs and save their long-range similarity, improve the effect of similar problem retrieval.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and educational data mining technology, and in particular to a method and system for retrieving similar exercises that integrates concept dependency relationships. Background Technology

[0002] With the increasing availability of learning resources and the growing demand for online self-directed learning, the need to organize and utilize resource data in online education systems is growing. Faced with a vast amount of resources (such as exercises), users may struggle to select appropriate materials without suitable advice. Online education systems can recommend exercises to students based on their cognitive levels or purposefully retrieve a series of exercises to generate exam papers. Among these, finding similar exercises (FSE) is a crucial task.

[0003] Given exercises with content (text and images) and knowledge concepts (referred to as concepts), most solutions learn exercise representations by treating content and concepts as complementary materials. While some works attempt to capture long-range similarity between exercises by incorporating concept-exercise associations, they only utilize single relationships between concepts, ignoring multiple dependencies, which limits the effectiveness of similar exercise retrieval. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for retrieving similar exercises that integrates concept dependencies, which can improve the effectiveness of similar exercise retrieval.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A similar exercise retrieval method incorporating concept dependency relationships includes:

[0007] Obtain the content, concepts, and dependencies between the concepts corresponding to the exercises, and construct a multi-layer graph that includes a concept dependency graph, a concept-exercise graph, and an exercise content graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes.

[0008] The initial vector representation of each node in the multi-layer graph is obtained through initialization;

[0009] The vector representation of each node is updated through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input for the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise node. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses an attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses an attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses an attention mechanism to calculate the updated vector representation of the current exercise node.

[0010] For each exercise to be retrieved, the similarity between the final vector representation of its exercise node and the final vector representation of the exercise node of each other exercise is calculated. The other exercises are then sorted in descending order according to the similarity to generate the retrieval results.

[0011] A similar exercise retrieval system that integrates concept dependencies includes:

[0012] The data acquisition and multi-layer graph construction unit is used to acquire the content, concepts and dependencies between concepts corresponding to the exercises, and to construct a multi-layer graph that includes a concept dependency graph, a concept-exercise graph, and an exercise content graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes.

[0013] An initialization unit is used to obtain the initial vector representation of each node in a multi-layer graph through initialization.

[0014] The node vector representation update unit is used to update the vector representation of each node through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input to the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise nodes. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses an attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses an attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses an attention mechanism to calculate the updated vector representation of the current exercise node.

[0015] The similarity retrieval unit is used to calculate the similarity between the final vector representation of the question node of each question and the final vector representation of the question node of each other, and to sort the other questions in descending order according to the similarity to generate retrieval results.

[0016] A processing device includes: one or more processors; and a memory for storing one or more programs;

[0017] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0018] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0019] As can be seen from the technical solution provided by the present invention, by treating content and concept as two important attributes of exercises and introducing multiple dependencies between concepts, a relationship graph (multi-layer graph of concept-exercise-content) is constructed. Then, the concept-exercise relationship and the multiple dependencies between concepts are comprehensively modeled to mine the implicit paths between exercise pairs and preserve their long-range similarity, thereby improving the effect of similar exercise retrieval. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0021] Figure 1 A flowchart illustrating a similar exercise retrieval method incorporating concept dependencies, provided as an embodiment of the present invention;

[0022] Figure 2 A network structure diagram of a similar exercise retrieval method that integrates concept dependencies, provided in an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of a similar exercise retrieval system that integrates concept dependencies, provided in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0026] First, the following explanations are provided for the terms that may be used in this article:

[0027] The terms “including,” “comprising,” “containing,” “having,” or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, “including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.)” should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0028] The following is a detailed description of a similar exercise retrieval method and system based on fused concept dependencies provided by this invention. Contents not described in detail in the embodiments of this invention belong to prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they should be performed according to conventional conditions in the art or conditions recommended by the manufacturer.

[0029] Example 1

[0030] This invention provides a method for retrieving similar exercises by incorporating concept dependencies, such as... Figure 1 As shown, it mainly includes the following steps:

[0031] Step 1: Obtain the content, concepts, and dependencies between the concepts corresponding to the exercises, and construct a multi-layer graph that includes a concept dependency graph, a concept-exercise association graph, and an exercise content association graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes.

[0032] Step 2: Obtain the initial vector representation of each node in the multi-layer graph through initialization;

[0033] Step 3: Update the vector representation of each node through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input to the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise node. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses the attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses the attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses the attention mechanism to calculate the updated vector representation of the current exercise node.

[0034] Step 4: For each exercise to be retrieved, calculate the similarity between the final vector representation of its exercise node and the final vector representation of the exercise node of each other exercise. Sort the other exercises in descending order according to the similarity and generate the retrieval results.

[0035] The above-described solution provided by the embodiments of the present invention fully considers the association between concepts and exercises, and the multiple dependencies between concepts. On the one hand, modeling the rich domain knowledge implied by the multiple dependencies between concepts enhances the understanding of the concepts under examination. On the other hand, combining the association between concepts and exercises, the implicit paths between exercise pairs are mined through an attention fusion layer to preserve their long-range similarity, greatly improving the effect of similar exercise retrieval.

[0036] For ease of understanding, the above-mentioned solutions provided in the embodiments of the present invention will be described in detail below.

[0037] I. Data Collection and Association Graph Construction.

[0038] In this embodiment of the invention, the content (text and images) corresponding to the exercises, concepts, and multiple dependencies between concepts are obtained. Concepts and content are treated as attributes of the exercises. These materials naturally present a multi-layered graph structure of concept-exercise-content, which can be represented as G = KG∪KEG∪ECG}, where G represents a multi-layered graph; KG represents a concept dependency graph composed of concept nodes and dependencies between concepts; KEG represents a concept-exercise association graph composed of exercise nodes, concept nodes, and their relationships; and ECC represents an exercise-content association graph composed of exercise nodes, content nodes, and their relationships.

[0039] Those skilled in the art will understand that "concept" is a specialized term in the field and generally refers to related knowledge points.

[0040] II. Initialization of node vector representation.

[0041] In this embodiment of the invention, an embedding layer is used to initialize the vector representations of concept nodes and exercise nodes: the embedding layer converts the one-hot vectors of related nodes into low-dimensional vectors with dense values, which serve as the initial vector representations of the corresponding nodes. For each exercise node e... q and concept node k f Each through the embedding matrix Emb q and Emb k To obtain the initial representation and

[0042] Those skilled in the art will understand that one-hot vectors have a high dimension, with only one dimension having a value of 1 and the other dimensions having a value of 0, and the non-zero values ​​are very sparse. Therefore, by using an embedding layer, one-hot vectors can be converted into low-dimensional vectors with a lower dimension, which are not just composed of 0 and 1, and have a denser non-zero value.

[0043] For content nodes containing both text and images, a Long Short-Term Memory (LSTM) with attention mechanism is used to associate different parts of the text with different images, resulting in a unified semantic representation, which serves as the initial vector representation for the content node; specifically:

[0044] Content nodes include text and images, containing rich semantic information. For each exercise, the image set {p1, p2, ..., p...} LI Each image in} i i = 1, 2, ..., L I L I To represent the number of images, a pre-trained image module, ResNet18, is used to extract image representations, m i =σ(Resnet(p i σ represents the activation function, and ResNet represents the pre-trained image module ResNet18. Then, LSTM with attention mechanism (AttLSTM) is used to leverage the correlation between text and images, that is, different parts of the text are associated with different images, to obtain a unified semantic representation of the content.

[0045] The calculation method is as follows:

[0046] w t =(o t ||τ t )

[0047]

[0048]

[0049] Where || is the concatenation operator, w tThis is the input vector at the t-th time step of the AttLSTM; the text is segmented into tokens, thus converting the text into a token sequence. t Let W be the vector representation of the t-th token in the token sequence. oi It is the weight matrix, m oi It is a bias term, η j The text o represents the regularized text. t and the image represents m i The relevance weights. The above w is obtained. t The final content node initialization vector can be represented as L is the content node c u The length of the text sequence in h L-1 This represents the output of the (L-1)th time step of the LSTM with attention mechanism, combined with w t Calculate the output h of the last time step L. L .

[0050] The LSTM used in this embodiment of the invention is a commonly used sequence model, and its working principle can be referred to conventional techniques, so it will not be described in detail.

[0051] 3. Introduce an attention fusion layer to update the vector representation of nodes.

[0052] In this embodiment of the invention, an attention fusion layer is introduced to perform attention fusion on three types of nodes respectively, so as to fully utilize the multiple dependencies and the association between concepts and exercises in the concept-exercise-content multi-layer graph. A node may exist in multiple subgraphs. For example, an exercise node may appear in both the concept-exercise association graph and the exercise-content association graph, while a concept node may appear in both the concept dependency graph and the concept-exercise association graph. Therefore, node-level attention is used to compute the information aggregation of the node's neighbors in each subgraph. For information aggregation from different subgraphs, hierarchical attention is used to update each node. In order to model satisfactory long-range similarity during representation learning, multiple attention fusion layers are superimposed on the multi-layer graph to fully explore implicit paths.

[0053] Specifically, the update process of the vector representations of the three types of nodes can be described as follows: For the current concept node, combining the dependency relationships contained in the concept dependency association graph with the vector representation of the input concept node, the dependency aggregation information is calculated using a node-level attention mechanism; combining the vector representation of the corresponding exercise node in the concept exercise association graph, the influence from the exercise node is calculated; combining the dependency aggregation information and the influence from the exercise node, the updated vector representation of the current concept node is calculated using a hierarchical attention mechanism; For the current content node, combining the vector representations of related exercise nodes in the content association graph, the updated vector representation of the current content node is calculated using a hierarchical attention mechanism; For the current exercise node, combining the vector representations of related concept nodes in the concept exercise association graph, the aggregation vector is calculated using a node-level attention mechanism, and then combining the vector representations of related content nodes in the exercise content association graph, the updated vector representation of the current exercise node is calculated using a hierarchical attention mechanism.

[0054] The preferred implementation method for updating the vector representation of the three types of nodes is described below.

[0055] 1. Update method for concept node vector representation.

[0056] The concept dependency graph (KG) contains rich structural information representing knowledge, consisting of concepts and various dependencies. Furthermore, in the concept-exercise association graph (KEG), each concept is associated with multiple exercises. Therefore, for the current concept node k... f Its updates need to take into account the impact of multiple dependencies and concept-exercise associations.

[0057] Dependencies in a concept dependency graph can typically be categorized into directed relationships (e.g., predecessor relationships) and undirected relationships (e.g., similarity relationships), and are represented as unidirectional edges and bidirectional edges, respectively. Furthermore, the current concept node k... f There may be multiple neighbor concepts, and the degree of dependence on different neighbor concepts may vary.

[0058] Obtain the set of dependencies D from the concept dependency graph. For any dependency d, calculate the dependency aggregation information, represented as:

[0059]

[0060]

[0061] in, Indicates the current concept node k f The set of neighboring nodes under dependency d k represents dependency aggregation information under dependency relationship d. t Represents the set of neighboring nodes. The concept of a single neighbor node in the context of the network. This represents the neighboring concept node k input to the current layer n. t The vector representation, n = 0, 1, ..., N-1, where N is the total number of multi-layer attention fusion layers. When n = 0, ... For the concept of neighbor node k t The initial vector representation, when n = 1, ..., N-1, The neighbor concept node k calculated for the upper layer t The updated vector representation (used as input to the current layer n), This represents the trainable parameter matrix in the current layer n that acts on neighboring concept nodes. This represents the neighbor concept node k calculated in the current layer n. t The intermediate vector representation, For node-level attention weights, let k represent the neighbor concept node. t For the current concept node k f The degree of influence is calculated as follows:

[0062]

[0063] in, This represents the current concept node k computed in the current layer n. f The intermediate vector representation, This represents the current concept node k input to the current layer n. f The vector representation of , and These are the trainable parameter matrix and the bias term, respectively.

[0064] For each concept k in the conceptual association graph KEG f Concept k f This is an important attribute of many exercises. Therefore, by combining the vector representation of the corresponding exercise node in the concept-exercise association graph, the influence from the exercise node is calculated, expressed as:

[0065]

[0066]

[0067] in, Indicates the current concept node k f The influence from the exercise nodes, Indicates the current concept node k f The corresponding set of exercise nodes, e b For a single exercise node in the set of exercise nodes, This represents the input to the problem node e in the current layer n. b The vector representation of , This represents the exercise node e calculated in the current layer n. b The intermediate vector representation, This represents the trainable parameter matrix that acts on the exercise nodes in the current layer. Indicates the current concept node k f With exercise node e b The degree of correlation is calculated as follows:

[0068]

[0069] in, This represents the current concept node k computed in the current layer n. f The intermediate vector representation, This represents the current concept node k input to the current layer n. f The vector representation of , and These are the trainable parameter matrix and the bias term, respectively.

[0070] Finally, combining dependency aggregation information with the influence from exercise nodes, a hierarchical attention mechanism is used to compute the updated vector representation of the current concept node.

[0071]

[0072]

[0073]

[0074] in, This represents the hierarchical attention weights calculated using dependency aggregation information. and These are the attention weights for calculating the hierarchical levels. The trainable parameter matrix and bias terms used at that time; This represents the hierarchical attention weights calculated using the influence from exercise nodes. and These are the attention weights for calculating the hierarchical levels. The trainable parameter matrix and bias terms used at that time.

[0075] 2. Update method for content node vector representation.

[0076] In the Exercise-Content Association Graph (ECG), for a content node containing rich semantic information, it is associated with the exercise node e. u For a one-to-one association, the formula for calculating the updated vector representation of the current content node is:

[0077]

[0078] in, This indicates the exercise node e corresponding to the current content node in the current layer n. u The vector representation of , This represents the current content node c input to the current layer n. u The vector representation of , This indicates the current content node and its corresponding exercise node e. u The attention weights can be calculated using the following formula:

[0079]

[0080] in, These represent the current content node c calculated in the current layer n. u and e u The intermediate vector representation, and These are the trainable parameter matrix and the bias term, respectively.

[0081] 3. Update method for the vector representation of exercise nodes.

[0082] In the Concept-Exercise Association Graph (KEG), the number of concepts associated with different exercises varies. Each exercise node has its own key concepts being tested. Furthermore, in the Exercise-Content Association Graph (ECG), each exercise node is also associated one-to-one with content nodes containing rich semantic information. Concepts and content together constitute a complete exercise, and their contributions are distinct.

[0083] In this embodiment of the invention, the current exercise node is denoted as e. q By combining the vector representations of related concept nodes in the concept-exercise association graph, a node-level attention mechanism is used to calculate the aggregation vector. Represented as:

[0084]

[0085]

[0086] in, Indicates the current exercise node e q The corresponding set of concept nodes, k j This represents a single concept node in the set of concept nodes. This represents the neighboring concept node k input to the current layer n. j The vector representation of , This represents the trainable parameter matrix that acts on each concept node in the current layer n. This represents the neighbor concept node k calculated in the current layer n. j The intermediate vector representation, Represents concept node k jFor the current exercise node e q The importance of.

[0087] By combining the vector representations of related content nodes in the problem content association graph, a hierarchical attention mechanism is used to calculate the updated vector representation of the current problem node.

[0088]

[0089] in, This represents the current exercise node e input to the current layer n. q The vector representation of , This represents the current exercise node e input to the current layer n. q The vector representation of the corresponding content node; and These are all hierarchical attention weights, representing the importance of the concept-exercise association graph and the exercise content association graph, respectively.

[0090] Through exercise fusion operations, exercise representations that balance knowledge and semantic information are obtained. Furthermore, as the number of attention fusion layers increases, such as N=3, long-term similarity between exercises can be preserved by modeling implicit paths.

[0091] The above describes the update methods for the three types of nodes in the current layer n. Iterating continuously according to the method for the current layer n, the final vector representations of the three types of nodes are obtained through the last layer (the Nth layer). The final vector representation of the exercise nodes obtained at this point will be used for subsequent exercise retrieval. For concept nodes and content nodes, their final vector representations do not need to be used; that is, there is no need to enter the Nth layer for vector representation updates. However, for the sake of completeness, they have been described above.

[0092] IV. Exercise Retrieval.

[0093] The final vector representation of the exercise node can be obtained through the aforementioned method. This vector representation is learned while maintaining long-range similarity and balancing domain knowledge and content semantic information. During retrieval, search results are generated by calculating the similarity (similar question score) between the final vector representations of the exercise nodes of two exercises.

[0094] For a pair of exercises (E) a E b The similarity scores are calculated using a similarity scoring layer.

[0095]

[0096] Among them, exercises (E) a E b In the search query, one exercise is the question to be searched, and the other exercise is one of the other exercises. For Exercise E a The final vector representation of the corresponding exercise node, For Exercise E b The final vector representation of the corresponding exercise node; W z With b z These are the parameter matrix and bias term of the similarity scoring layer, respectively.

[0097] After calculating the similarity scores for all other questions in the search results, the other questions are sorted in descending order according to their similarity scores to generate the search results. That is, questions with higher similarity scores are ranked first, and those with lower scores are ranked last. Of course, the number of other questions in the search results can be set as needed, and only the corresponding number of other questions ranked first can be retained.

[0098] In all the formulas provided in the embodiments of the present invention, the parameter matrix (weight matrix) represented by the symbol W and the bias term represented by the symbol b are both trainable parameters. Different subscripts are used to distinguish different parameter matrices and bias terms. Their training methods can be implemented with reference to conventional techniques, which will not be elaborated in the present invention.

[0099] Figure 2 The network structure diagram of the method described above in this invention is shown, wherein: (a) represents the constructed multi-layer graph, and the lower part provides an example of a content node, i.e., each content node contains the text and image of the exercise; (b) represents the vector representation update process of the concept node; (c) represents the vector representation update process of the content node, wherein the CE (Content Encoder) corresponds to the part that processes the text and image into the initial vector representation; and (d) and (e) represent the vector representation update process of the exercise node.

[0100] The above-described scheme in this embodiment of the invention constructs a multi-layer graph and embeds each concept node, exercise node, and content node through an embedding layer. Then, an attention fusion layer is introduced, and the graph neural network is used to propagate on the multi-layer graph. This effectively maintains long-range similarity and rich domain knowledge to learn the exercise embedding representation. While improving the retrieval effect, it also provides a certain degree of interpretability to the retrieval results.

[0101] Example 2

[0102] This invention also provides a similar exercise retrieval system that integrates concept dependencies, which is mainly implemented based on the method provided in the foregoing embodiments, such as... Figure 3 As shown, the system mainly includes:

[0103] The data acquisition and multi-layer graph construction unit is used to acquire the content, concepts and dependencies between concepts corresponding to the exercises, and to construct a multi-layer graph that includes a concept dependency graph, a concept-exercise graph, and an exercise content graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes.

[0104] An initialization unit is used to obtain the initial vector representation of each node in a multi-layer graph through initialization.

[0105] The node vector representation update unit is used to update the vector representation of each node through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input to the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise nodes. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses an attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses an attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses an attention mechanism to calculate the updated vector representation of the current exercise node.

[0106] The similarity retrieval unit is used to calculate the similarity between the final vector representation of the question node of each question and the final vector representation of the question node of each other, and to sort the other questions in descending order according to the similarity to generate retrieval results.

[0107] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0108] Example 3

[0109] The present invention also provides a processing device, such as Figure 4 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0110] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0111] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example:

[0112] Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc.

[0113] The output device can be a display terminal;

[0114] The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0115] Example 4

[0116] The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0117] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0118] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for retrieving similar exercises by incorporating concept dependency relationships, characterized in that, include: Obtain the content, concepts, and dependencies between the concepts corresponding to the exercises, and construct a multi-layer graph that includes a concept dependency graph, a concept-exercise graph, and an exercise content graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes. The initial vector representation of each node in the multi-layer graph is obtained through initialization; The vector representation of each node is updated through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input to the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise node. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses an attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses an attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses an attention mechanism to calculate the updated vector representation of the current exercise node. For each exercise to be retrieved, the similarity between the final vector representation of its exercise node and the final vector representation of the exercise node of each other exercise is calculated. The other exercises are then sorted in descending order according to the similarity to generate the retrieval results. The steps for calculating the updated vector representation of the current concept node include: Let the current concept node be denoted as Obtain the set of dependencies from the concept dependency graph. For any of these dependencies The computation depends on the aggregated information, represented as: ； ； in, Indicates the current concept node In dependency The set of neighboring nodes, Indicate dependency Dependency aggregation information below, Represents the set of neighboring nodes. The concept of a single neighbor node in the context of the network. Indicates input to the current layer Neighbor concept node The vector representation of , Indicates the current layer The trainable parameter matrix that acts on neighboring concept nodes. Indicates the current layer Neighbor concept nodes in the calculation The intermediate vector representation, The node-level attention weights are calculated as follows: ； in, , Indicates the current layer The current concept node calculated in the middle The intermediate vector representation, Indicates input to the current layer Current concept node The vector representation of , and These are the trainable parameter matrix and the bias term, respectively, and || is the concatenation operator; Combining the vector representations of the corresponding exercise nodes in the concept-exercise association graph, the influence from the exercise nodes is calculated, as follows: ； ； in, Indicates the current concept node The influence from the exercise nodes, Indicates the current concept node The corresponding set of exercise nodes, For a single exercise node in the set of exercise nodes, Indicates input to the current layer Exercise nodes The vector representation of , Indicates the current layer Exercise nodes in the calculation The intermediate vector representation, This represents the trainable parameter matrix that acts on the exercise nodes in the current layer. Indicates the current concept node Exercise nodes The degree of correlation; By combining dependency aggregation information with the influence from exercise nodes, a hierarchical attention mechanism is used to compute the updated vector representation of the current concept node. : ； ； in, This represents the hierarchical attention weights calculated using dependency aggregation information. and These are the attention weights for calculating the hierarchical levels. The trainable parameter matrix and bias terms used at that time; This represents the hierarchical attention weights calculated using the influence from exercise nodes. and These are the attention weights for calculating the hierarchical levels. The trainable parameter matrix and bias terms used at that time.

2. The method for retrieving similar exercises by incorporating concept dependencies according to claim 1, characterized in that, The multi-layer diagram is represented as follows: ； in, Represents multi-layer diagrams; This represents a concept dependency graph, which consists of concept nodes and the dependencies between concepts. This represents a concept-exercise relationship diagram consisting of exercise nodes, concept nodes, and the relationships between them. This represents an exercise-content relationship graph consisting of exercise nodes, content nodes, and the relationships between them.

3. The method for retrieving similar exercises by incorporating concept dependencies according to claim 1, characterized in that, The initial vector representation of each node in the multi-layer graph obtained through initialization includes: For concept nodes and exercise nodes, the embedding layer is used to convert the one-hot vectors of related nodes into vectors with dense values, which serve as the initial vector representations of the corresponding nodes. For content nodes containing both text and images, long short-term memory with attention mechanism is used to associate different parts of the text with different images to obtain a unified semantic representation, which is then used as the initial vector representation of the content node.

4. The method for retrieving similar exercises by incorporating concept dependencies according to claim 1, characterized in that, In each of the layers: For the current concept node, the dependency aggregation information is calculated using a node-level attention mechanism by combining the dependency relationships contained in the concept dependency graph with the vector representation of the input concept node; the influence from the exercise node is calculated by combining the vector representation of the corresponding exercise node in the concept exercise graph. By combining dependency aggregation information with the influence from exercise nodes, a hierarchical attention mechanism is used to compute the updated vector representation of the current concept node. For the current content node, the updated vector representation of the current content node is calculated by combining the vector representations of related exercise nodes in the content association graph and using a hierarchical attention mechanism. For the current exercise node, the aggregation vector is calculated using a node-level attention mechanism, combining the vector representations of related concept nodes in the concept-exercise association graph. Then, the updated vector representation of the current exercise node is calculated using a hierarchical attention mechanism, combining the vector representations of related content nodes in the exercise content association graph.

5. A method for retrieving similar exercises by incorporating concept dependencies according to claim 1 or 4, characterized in that, The formula for calculating the updated vector representation of the current content node is: ； in, Indicates input to the current layer The current content node corresponds to the exercise node. The vector representation of , Indicates input to the current layer The vector representation of the current content node. This indicates the current content node and its corresponding exercise node. Attention weights.

6. A method for retrieving similar exercises based on fused concept dependencies according to claim 1 or 4, characterized in that, The steps for calculating the updated vector representation of the current exercise node include: Record the current exercise node as By combining the vector representations of related concept nodes in the concept-exercise association graph, a node-level attention mechanism is used to calculate the aggregation vector. , is represented as: ； ； in, Indicates the current exercise node The corresponding set of concept nodes, This represents a single concept node in the set of concept nodes. Indicates input to the current layer Neighbor concept node The vector representation of , Indicates the current layer The trainable parameter matrix that acts on each concept node. Indicates the current layer Neighbor concept nodes in the calculation The intermediate vector representation, Representing concept nodes For the current exercise node The degree of importance; By combining the vector representations of related content nodes in the problem content association graph, a hierarchical attention mechanism is used to calculate the updated vector representation of the current problem node. : ； in, Indicates input to the current layer Current exercise node The vector representation of , Indicates input to the current layer Current exercise node The vector representation of the corresponding content node; and These are all hierarchical attention weights, representing the importance of the concept-exercise association graph and the exercise content association graph, respectively.

7. A similar problem retrieval system that integrates concept dependency relationships, characterized in that, Based on the method described in any one of claims 1 to 6, the system comprises: The data acquisition and multi-layer graph construction unit is used to acquire the content, concepts and dependencies between concepts corresponding to the exercises, and to construct a multi-layer graph that includes a concept dependency graph, a concept-exercise graph, and an exercise content graph. The nodes in the multi-layer graph include content nodes, concept nodes, and exercise nodes. An initialization unit is used to obtain the initial vector representation of each node in a multi-layer graph through initialization. The node vector representation update unit is used to update the vector representation of each node through a multi-layer attention fusion layer. The first layer takes the initial vector representation as input and outputs the updated vector representation as input to the next layer. The last layer outputs the final vector representation of each node and extracts the final vector representation of the exercise nodes. In each layer: the current concept node combines the vector representations of related concept nodes and exercise nodes in the concept dependency graph and the concept-exercise graph, and uses an attention mechanism to calculate the updated vector representation of the current concept node; the current content node combines the vector representations of related exercise nodes in the content graph, and uses an attention mechanism to calculate the updated vector representation of the current content node; the current exercise node combines the vector representations of related concept nodes and content nodes in the concept-exercise graph and the exercise-content graph, and uses an attention mechanism to calculate the updated vector representation of the current exercise node. The similarity retrieval unit is used to calculate the similarity between the final vector representation of the question node of each question and the final vector representation of the question node of each other, and to sort the other questions in descending order according to the similarity to generate retrieval results.

8. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images