CN-122023079-A - Course knowledge point learning path recommendation system based on multi-mode data

CN122023079ACN 122023079 ACN122023079 ACN 122023079ACN-122023079-A

Abstract

The invention discloses a course knowledge point learning path recommending system based on multi-mode data, which comprises a basic processing layer and an information alignment reasoning layer. And the basic processing layer is used for extracting knowledge topics, knowledge points and integrating the mastering conditional probabilities among knowledge pairs from the original multi-mode curriculum materials. And the information alignment reasoning layer is used for processing the mastering condition probabilities among the knowledge topics, the knowledge points and the integrated knowledge pairs, carrying out structural representation on the knowledge points in a triplet form and aligning the knowledge points in the related knowledge topics, and finally predicting the first repair relationship among the knowledge points through the judgment model to form a learning path so as to finish the recommendation of the learning path of the knowledge points of the course. The invention reduces the traditional modes of manual design rule paradigm and manual labeling of structured teaching resources, and realizes the automatic extraction of knowledge points and relations from unstructured or semi-structured teaching documents by integrating advanced natural language processing and machine learning technologies.

Inventors

HOU PINGZHI
WANG WEI
LIU RUIHANG
XU XIAOBIN
FENG JING
LI DIE
ZHANG ZHENJIE

Assignees

杭州电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (8)

1. A course knowledge point learning path recommendation system based on multi-mode data is characterized by comprising a basic processing layer and an information alignment reasoning layer; the basic processing layer is used for extracting knowledge topics, knowledge points and grasping condition probabilities among integrated knowledge pairs from the original multi-mode curriculum materials; The information alignment reasoning layer is used for processing the mastering condition probabilities among the knowledge topics, the knowledge points and the integrated knowledge pairs, carrying out structural representation on the knowledge points in a triplet form and aligning the knowledge points in the related knowledge topics, and finally predicting the first repair relationship among the knowledge points through the judgment model to form a learning path so as to finish the recommendation of the learning path of the knowledge points of the course.
2. The multi-modal data-based course knowledge point learning path recommendation system as claimed in claim 1, wherein the input multi-modal course material of the base processing layer is composed of four parts of a teaching material document, a question list, a course requirement document, and knowledge point mastery distribution data.
3. The multi-modal data-based course knowledge point learning path recommendation system of claim 2, wherein in the base processing layer, the following processes are performed for course requirement documents, problem lists, and knowledge point mastering distributions: The curriculum requirement document refers to document materials with class properties, subject knowledge and knowledge topics are contained in the document materials, key topics in the curriculum requirement document are extracted to be used as objects for aligning subsequent triplet knowledge points, a problem list contains problems of importance of the teaching materials, a large language model is assisted to grasp important points of the teaching materials in a pure text mode, knowledge point grasp distribution is the grasp condition of students on related knowledge points, and a certain knowledge point is grasped and is recorded as 1 or 0.
4. The multi-modal data-based course knowledge point learning path recommendation system of claim 2, wherein the processing of the teaching material documents in the base processing layer comprises a material blocking module, a knowledge point extraction module, a triplet extraction module and a verification merging module; The material blocking module is used for blocking the original teaching material text, firstly summarizing each section by using a large language model to obtain section summarization, and blocking each section independently by taking sentences as granularity; The knowledge point extraction module is used for screening and discarding irrelevant sentences according to chapter summarization and problem list, extracting and storing relevant knowledge sentences in text blocks, utilizing a large language model to carry out atomic resolution on problems of the problem list to obtain a group of atomic problem list, utilizing an embedding model to embed and fuse, obtaining relevant key vectors by adding each problem of the embedding chapter summarization and the atomic problem list, and setting a similarity threshold value Multiplying each sentence by the key vector point and then by a similarity threshold value Comparing, retaining similarity greater than threshold The result of (2) realizing screening; the triplet extraction module is used for converting sentences screened by the knowledge point extraction module into a triplet structure form to be stored, and the process is used for carrying out entity extraction and relation extraction on texts through large language model structuring; The verification module is used for carrying out legal verification, regularized extraction is carried out on the triple structure data in the legal verification stage, if the data does not accord with the regularized rule, the entity or the relation is firstly discarded, each field of the triple structure data is further checked, the checking comprises format checking, non-empty checking, legal checking and validity checking, and if the verification checking is wrong, all relevant information of the triple is directly ignored.
5. The multi-modal data-based course knowledge point learning path recommendation system of claim 2, wherein the information inference alignment layer comprises a knowledge point topic classification module and a first repair relationship judgment module.
6. The multi-modal data-based course knowledge point learning path recommendation system of claim 5, wherein the knowledge point topic classification module is configured to generalize the triples output by the verification module to knowledge points according to course requirements to form a hierarchical structure, and the specific process is as follows: deploying a pre-trained bert model, and representing the subject or knowledge point represented by the triplet output by the verification module by using a sequence of a token symbol, a combination of a < cls > prefix symbol and a < sep > suffix symbol; the subject and knowledge points are respectively used as the input of bert encoder, the first vector after encoding is taken, namely As subject words Or knowledge points Traversing and integrating all topics and knowledge points to obtain a topic vector set And knowledge point vector set ; Subject term vector Adding random noise to obtain m data enhanced samples Will be as a sample And enhancing the sample Combining to obtain a new vector group sequence with index k Traversing all indexes to obtain k groups of new vector group sequences, and integrating the k groups of new vector group sequences according to the index i as a sequence to obtain a sequence In addition, in the sequence Find a corresponding distance sequence based on (a) Wherein the elements are Is that Elements of the sequence And element(s) And to Each element scaling of (2) And the above mentioned middle part , And is the sign of the inversion of the sign, In order to take the remainder of the symbol, Representing sequences The largest element in (a); Above-mentioned Sequence as input to coding classification model, vector After the transformation layer, the coding vector is obtained The final classification result is obtained through a classification head, wherein the classification head parameter and the encoder layer parameter are independently trained, the classification task head parameter is reversely updated by a cross entropy loss function, and the encoder layer parameter is reversely updated by a loss function defined by the following formula, wherein As index function, when indexing The value is 1, when the index The duration is 0: By topic vector sets Training a coding classification model is completed; Vector knowledge points And as the input of the trained coding classification model, obtaining a theme classification result through prediction.
7. The multi-modal data-based course knowledge point learning path recommendation system as set forth in claim 6 wherein said first-repair relationship determination module uses elements The ith element of the subsequence representing knowledge point A and knowledge point B respectively, combines knowledge point A and knowledge point B, has a sequence Bringing bert the encoder and grabbing its corresponding query sequence from the last layer of transducer of the encoder Key sequence : For knowledge point pair A, B, intercepting the subsequence corresponding to the index 、 And 、 And summing the averages to represent knowledge point A And Vector, likewise, obtain knowledge point B And For use in (1) Respectively represent the A sequence corresponding to the A sequence Index start and end in sequence, using Respectively represent the sequences B corresponding to the sequences B In addition, according to the knowledge point subject classification module, A, B knowledge points are independently used as input to obtain the coding expression Click operation is carried out on the query vector and the key vector of the two knowledge points A, B to obtain semantic related features And And obtaining by using A, B classification codes Wherein Fusion of semantic relevance for normalized click operations 、 And classifying the encoded values 、 With semantic classification features And , , , The self-defined adjustment coefficient is adopted; Based on semantic classification features And And constructing a feature vector, and predicting and outputting a first repair relation prediction.
8. The multi-modal data-based course knowledge point learning path recommendation system of claim 7, wherein the semantic classification feature based And The feature vector is constructed, and the prediction output repair relation prediction is specifically as follows: the knowledge point first repair relation judgment is based on pairwise judgment to realize global knowledge point first repair relation judgment, wherein P (A) represents the grasping probability of the knowledge point A, and P # ) The probability of grasping the knowledge point A is represented, the probability of grasping the knowledge point B under the condition of grasping the knowledge point A is represented by P (B|A), and the probability is counted by grasping distribution data, and the rest is the same; constructing feature vectors by using knowledge points A and B , Selecting the characteristic vector representing the knowledge point A as the first repair of the knowledge point B Each row and each column element forms the first repair feature vector The first part of the feature vector is respectively mastered As under the condition of mastering Conditional probability, i.e. conditional probability Second part, grasp respectively As a condition of not being mastered Conditional probability, i.e. conditional probability The third part is not mastered As under the condition of mastering Conditional probability, i.e. conditional probability The fourth part, not mastered Is not known under the condition Conditional probability, i.e. conditional probability Fifth part, semantically classifying feature elements The first four parts have 4 elements, which are ordered according to the logic of the priority of the first repair knowledge points, namely the description sequence of the conditional probability values, and the five parts are integrated to finally obtain the first repair feature vector expression Traversing all the knowledge points, and obtaining n by n knowledge points (N-1) first-repair feature vectors; the first repair relation prediction module comprises a multi-layer perceptron MLP and a fraction mapping S part, wherein the MLP maps the first repair feature vector into a scalar quantity Design score mapping Wherein The function is composed of a hyperbolic tangent function tanh (z) and a custom threshold Collectively defined, the expression is as follows: 。

Description

Course knowledge point learning path recommendation system based on multi-mode data Technical Field The application relates to the technical field of data processing and education knowledge engineering, in particular to a course knowledge point learning path recommendation system based on multi-mode data. Background In existing knowledge graph construction and teaching analysis systems, knowledge point capture and relationship construction typically relies on highly structured or semi-structured teaching resources. The mainstream technical scheme is to manually or semi-automatically extract entities and relations from texts through manual design rules or by using limited natural language processing tools. Finally, the knowledge point data are called in the form of service interfaces for downstream tasks such as question answering systems, recommendation systems and the like. However, knowledge maps constructed based on the above manner generally have the following drawbacks. The automation degree is low, the labor cost is high, and the existing method is seriously dependent on manual rule design and labeling of teaching materials in specific fields and specific formats. Once the text material field changes, the original processing method fails, and batch processing of the interdisciplinary and interdisciplinary documents is difficult to realize. This strong dependence results in poor generalization ability of the system and low automation level. The knowledge structure is flat, the hierarchy and the association are lacking, the knowledge point set generated by the traditional method is mostly independent triples, and an effective mechanism is lacking to identify and organize the hierarchy (such as partial overall relation) among the knowledge points. The deep semantic relations are important to clearly transfer the knowledge architecture to the user, assist the large language model to carry out knowledge reasoning and demonstration, and precisely locate the knowledge blind area of the user. Without these relationships, the value of the knowledge-graph would be compromised. The prior solution mostly only processes teaching materials, but fails to effectively fuse and utilize multi-mode teaching data (such as knowledge point mastering distribution data reflecting learning states of students, knowledge point course requirement documents for providing knowledge point topics, and problem lists for assisting in generating knowledge points aiming at specific interest problems). However, these heterogeneous multi-source data contain rich information and have important value for deep mining of supporting relationships among knowledge points. Disclosure of Invention Aiming at the problems, the invention provides a course knowledge point learning path recommending system based on multi-mode data. The system is structurally divided into a basic processing layer and an information alignment reasoning layer based on the multi-mode data knowledge point repair relation judgment system. The basic processing layer is mainly responsible for extracting knowledge topics, knowledge points and integrating knowledge pairs from the original multi-mode curriculum materials. And then, the knowledge topics, the knowledge points and the knowledge point integration pair mastering condition probabilities are processed in an information alignment processing layer, the knowledge points are structurally represented in a triplet form and aligned in the related knowledge topics, and finally, the first repair relationship among the knowledge points is predicted through a judgment model to form a learning path, so that the recommendation of the course knowledge point learning path is completed. The input multi-mode curriculum materials of the basic processing layer are composed of four parts of teaching material documents, problem lists, curriculum requirement documents and knowledge point mastering distribution data, and different processing means are needed for different materials. The course requirement document, the problem list and the knowledge point mastering distribution data do not need to be processed in a complex mode, wherein the course requirement document refers to document materials with the class property, and a large number of discipline knowledge modules and knowledge topics are contained in the course requirement document. The system directly extracts key topics in the course requirement document to be used as objects for aligning the subsequent triplet knowledge points, a problem list comprises the problem of importance of the teaching materials, and assists a large language model to grasp the key points of the teaching materials, so that the system can ignore a large number of invalid sentences in the teaching materials, the system can generate and reduce the burden on the reasoning task of the system, the system is in a pure text form, knowledge point grasp distribution is the grasp condition of each student on related kno