Course data intelligent multi-dimensional retrieval system based on deep learning

Abstract

The invention discloses a course data intelligent multi-dimensional retrieval system based on deep learning, and relates to the technical field of course retrieval. The course data retrieval platform is in communication connection with a knowledge graph construction module, a user intention analysis module, a user portrait construction module, a personalized recommendation engine module and a retrieval display module, and all the modules are in electric signal connection; the knowledge graph construction module is used for collecting course data from multiple sources and constructing a course domain knowledge graph. According to the natural language processing technology based on deep learning, deep semantics queried by a user can be deeply analyzed, fuzzy or implicit requirements can be understood, limitation of traditional keyword matching is broken through, intent features and entity relationships can be extracted by combining a knowledge graph, specific entities and relationships queried by the user can be accurately mapped, and user experience is improved. Therefore, a more accurate retrieval result is provided, and personalized requirements of users are met.

Description

Technical Field

[0001] The present invention relates to the field of course retrieval technology, and in particular to an intelligent multi-dimensional retrieval system for course data based on deep learning. Background Art

[0002] With the deepening of educational informatization, various educational institutions have accumulated massive amounts of course data, including course information, teaching resources, student evaluations, teacher information, etc. How to efficiently manage and utilize this data has become a major challenge facing educational institutions. The modern curriculum system is becoming increasingly rich, covering course resources in different disciplines, levels, and forms. When users search for courses, they often need to filter from multiple dimensions. The use of intelligent multi-dimensional retrieval systems can integrate, analyze, and retrieve this data, providing accurate information support for educational administrators, teachers, and students, and improving the utilization efficiency of educational resources.

[0003] In the existing technology, traditional retrieval systems mainly rely on keyword matching, which makes it difficult to understand the deep semantics of user queries. Moreover, course data is often scattered across multiple systems, making it difficult to achieve cross-system data association, which in turn leads to the problem of being unable to make accurate recommendations based on the user's personalized needs. Therefore, how to use the natural language processing technology in deep learning to perform semantic analysis on user queries, and combine the entities and relationships in the knowledge graph to further parse the query intent and provide users with personalized course recommendations is the problem to be solved by the present invention. To this end, an intelligent multidimensional retrieval system for course data based on deep learning is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent multi-dimensional retrieval system for course data based on deep learning to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A deep learning-based intelligent multi-dimensional course data retrieval system includes a course data retrieval platform, wherein the course data retrieval platform is communicatively connected to a knowledge graph construction module, a user intent analysis module, a user portrait construction module, a personalized recommendation engine module, and a retrieval display module, wherein the modules are electrically connected;

[0007] The knowledge graph construction module is used to collect course data from multiple sources and construct a course domain knowledge graph;

[0008] The user intent analysis module is used to receive query information input by the user and perform semantic analysis on the user query using natural language processing technology in deep learning. It combines the course domain knowledge graph to extract intent features and entity relationships, obtain the intent deviation coefficient, analyze the degree of deviation between the user query intent and the system's understanding, and further analyze the user's true intent, thereby overcoming the limitations of traditional retrieval systems that rely on keyword matching, improving semantic understanding capabilities, and accurately identifying ambiguous or implicit requirements in user queries.

[0009] The user profile building module is used to extract user features and build a user profile by analyzing the user's search history, browsing behavior and learning progress data;

[0010] The personalized recommendation engine module is used to combine the collaborative filtering technology in deep learning and the constructed user portrait to provide users with personalized course recommendations and achieve accurate personalized recommendations;

[0011] The search display module is used to combine the results of the personalized recommendation engine to generate a search result list, sort and page the search results. Through multi-dimensional search, users can accurately filter courses according to their needs.

[0012] A further improvement of the technical solution of the present invention is that: the knowledge graph construction module includes a data integration unit and a knowledge graph unit;

[0013] The data integration unit is used to collect course data from multiple source systems, including the educational administration system, online learning platform, teaching resource library, etc., and pre-process the collected data;

[0014] The knowledge graph unit constructs a course domain knowledge graph based on the preprocessed course data, which includes entities of courses, teachers, knowledge points and learning objectives, as well as the relationships between the entities.

[0015] A further improvement of the technical solution of the present invention is that the data integration unit specifically includes:

[0016] Collect course data from the educational system, online learning platform, and teaching resource library through API interfaces, direct database connections, or web crawler technology to ensure data integrity and timeliness. Design appropriate parsing rules for the data formats of different systems to ensure that the data can be recognized by the system. At the same time, record the data source and collection timestamp.

[0017] Clean and pre-process the collected data to ensure data quality and consistency, and unify data from different sources into a standardized format;

[0018] Based on the unique identification of the course ID, multi-source data is associated and matched, and then the same course data from different sources are merged and stored in a unified data warehouse.

[0019] A further improvement of the technical solution of the present invention is that the knowledge graph unit specifically includes:

[0020] Based on the preprocessed course data, rule matching and a BiLSTM-CRF deep learning model are used to identify key entities, including course entities, teacher entities, knowledge point entities, and learning objective entities. Unique entity identifiers based on course IDs are also defined.

[0021] Through pattern matching and semantic analysis, a relationship network between entities is constructed. Entity relationships include course-teacher relationships, course-knowledge point relationships, course-learning objective relationships, course-course type relationships, and cross-system relationships. An attribute graph model is used to store relationships, forming a semantic network in the course domain and enabling deep association of cross-system data.

[0022] Based on the unique identification of the course ID, entities with the same name in multi-source data are merged, including entity alignment and relationship fusion, to resolve data conflicts and build a complete course domain knowledge graph.

[0023] A further improvement of the technical solution of the present invention is that: the user intention analysis module includes a user query input unit and a semantic analysis unit;

[0024] The user query input unit is used to receive natural language queries input by users, including queries about courses, course types, and learning needs, and convert the user's query content into a data format that can be processed by the system;

[0025] The semantic parsing unit is used to use natural language processing technology in deep learning to perform semantic analysis on natural language queries input by users, extract key information in the queries, and analyze intent features and entity relationships in combination with the knowledge graph, calculate the intent deviation coefficient, map the user queries to specific entities and relationships in the knowledge graph, clarify the user's specific needs, accurately understand the deep intention of the user query, and avoid ambiguity and misunderstanding.

[0026] A further improvement of the technical solution of the present invention is that the user query input unit specifically includes:

[0027] Receive natural language queries input by users through a front-end interface including a web form (including a text input box and a speech-to-text button) or a mobile input box (supporting keyboard input and speech recognition), and integrate a speech recognition API to convert speech into text in real time, annotating the input method. At the same time, a unique ID is generated for each query, recording the query timestamp, user ID, input method, and original query text, and storing them in a log database. The query content includes course name, course type, and learning needs, etc., and supports multiple methods such as text input and voice input to meet the needs of different users;

[0028] Preprocessing of natural language queries entered by users, including word segmentation and part-of-speech tagging, denoising and normalization, and stop word filtering. Jieba word segmentation tools are used to split queries into lexical units, annotate parts of speech, and generate a list of word segmentation results. This removes irrelevant symbols and standardized abbreviations, corrects spelling errors, loads a stop word list, removes words without actual meaning, and retains key entities.

[0029] The word segmentation results, standardized text, and filtered keyword list are associated with the original query ID and stored in the preprocessing log table.

[0030] A further improvement of the technical solution of the present invention is that the semantic parsing unit specifically includes:

[0031] A pre-trained BERT model is used to encode pre-processed user queries and generate context-sensitive word vector representations. Fine-tuning is then performed on intent classification and entity recognition tasks to extract key information from the query, including the course name, course type, and learning requirements. The query is then classified into predefined intent types. A sequence labeling model (BiLSTM-CRF) is then used to identify entities and their attributes in the query and generate a structured representation. By converting natural language queries into structured semantic representations, the core user needs are clarified, providing a foundation for subsequent knowledge graph mapping.

[0032] The extracted entities are matched with entities in the course domain knowledge graph. Unique entity IDs are determined through string similarity and context consistency verification. Based on the query intent and entity attributes, the relationship path required by the user is inferred. At the same time, the attributes and relationships of the knowledge graph are combined to expand the query scope and map the user query to specific entities and relationships in the knowledge graph, clarifying the semantic boundaries of the query and avoiding ambiguity.

[0033] Combined with the contextual information of the knowledge graph, the user intent is further refined, and the refined intent and entity relationships are mapped into the query language of the knowledge graph, thereby extracting intent features and entity relationships. Intent features include query subject, query intent type, query constraints, and query semantic context. Entity relationships include course-teacher relationships, course-knowledge point relationships, course-learning objective relationships, course-course type relationships, and cross-system relationships.

[0034] Based on the extracted intent features and entity relationships, the intent deviation coefficient is calculated to analyze the degree of deviation between the user's query intent and the system's understanding, and to check whether the query results are consistent with the user's original intent. If there is a deviation, the user is prompted to provide additional information, and an accurate knowledge graph query statement is generated to ensure that the returned results are consistent with the user's deep intention. Figure 1 Improve the query accuracy.

[0035] A further improvement of the technical solution of the present invention is that the calculation process of the intention deviation coefficient is:

[0036] Extract the actual value of the intent feature and the actual value of the entity relationship from the user query, and obtain the expected value of the corresponding intent feature and the expected value of the entity relationship from the knowledge graph. Then determine the weight of each intent feature and entity relationship. The weight is assigned according to the importance of the intent feature or entity relationship in the query. At the same time, set a baseline deviation value according to system requirements to normalize the final intent deviation coefficient. The baseline deviation value is a constant greater than 0.

[0037] For each intent feature, calculate the square of the deviation between its actual value and the expected value and divide it by the corresponding weight. Then, sum the weighted squares of the deviations of all intent features to obtain the intent feature deviation function.

[0038] For each entity relationship, calculate the square of the deviation between its actual value and the expected value and divide it by the corresponding weight. Then, sum the weighted squares of the deviations of all entity relationships to obtain the entity relationship deviation function.

[0039] Adding the intent feature deviation function and the entity relationship deviation function yields the sum of squared deviations of the intent feature and entity relationship. Taking the square root converts the sum of squared deviations into a single value to measure the overall deviation. This is then divided by the baseline deviation value to normalize the deviation coefficient to the range of [0, 1] to obtain the intent deviation coefficient.

[0040] A further improvement of the technical solution of the present invention is that the user portrait construction module specifically includes:

[0041] Collect user data from multiple channels, including search history, browsing behavior, and learning progress. For search history, collect user historical query records, including query time, keywords, and clicked course IDs. For browsing behavior, record the course pages visited by users, length of stay, and page jump paths. For learning progress, synchronize user course learning data;

[0042] Extract user features from the collected user data, including search preferences, browsing depth, and learning progress. For search preferences, count high-frequency search keywords, calculate keyword TF-IDF values, and identify users' core areas of interest. For browsing depth, calculate the average duration of stay on the course page and bounce rate to assess users' interest in the content. For learning progress, extract learning completion rate, exercise accuracy rate, and learning frequency, and then generate user tags that include knowledge domain, learning stage, and behavioral preferences. Convert user behavior into structured tags to form an interpretable user feature representation.

[0043] Integrate the long-term behavioral characteristics of users in the past three months to generate static tags, and combine them with the user's real-time behavior on the day to generate short-term interest tags. By setting the weight of static tags to 70% and the weight of short-term interest tags to 30%, weighted fusion is performed to generate the final user portrait.

[0044] A further improvement of the technical solution of the present invention is that the personalized recommendation engine module specifically includes:

[0045] Extract static tags and short-term interest tags from user profiles, convert them into vector representations, and then combine them with user behavior features to form user feature vectors. Course attributes are also extracted and converted into vector representations, which are then combined with course feature vectors. A dual-tower model is used to perform nonlinear mapping on user and course features, generating low-dimensional dense vectors. This maximizes user-course matching and converts user profiles and course attributes into a computable vector space, providing high-quality feature input for collaborative filtering.

[0046] Based on user collaborative filtering, user similarity is calculated and courses favored by similar users are recommended. Based on the cosine similarity of course vectors, courses similar to the user's previous courses are recommended. The interactive signal (user-course click matrix) of collaborative filtering and deep features (the dot product of user and course vectors) are integrated to analyze the probability of users clicking on courses. The loss function is cross-entropy loss, and the optimization goal is to maximize the click-through rate of recommended courses. The top 50 course recommendations from collaborative filtering and the top 30 course recommendations from deep learning are combined through weighted sorting to generate a candidate recommendation list.

[0047] Combined with the user's current behavior, the update of short-term interest tags is triggered, the weight of the courses corresponding to the short-term interest tags is increased by 30%, and they are inserted at the front of the candidate recommendation list. The user's clicks on the recommended courses, learning time and other behaviors are recorded.

[0048] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art:

[0049] 1. The present invention provides an intelligent multi-dimensional retrieval system for course data based on deep learning. The natural language processing technology based on deep learning can deeply analyze the deep semantics of user queries, understand vague or implicit needs, break through the limitations of traditional keyword matching, and combine with knowledge graphs to extract intent features and entity relationships, accurately map user queries to specific entities and relationships, thereby providing more accurate retrieval results and meeting user personalized needs.

[0050] 2. The present invention provides an intelligent multi-dimensional retrieval system for course data based on deep learning. Through the knowledge graph construction module, it can collect course data from multiple sources, construct a course domain knowledge graph, and realize cross-system data association. The knowledge graph represents the course knowledge system in the form of nodes and edges, supports semantic association and reasoning, can quickly identify key entities and relationships in user queries, provide more comprehensive and accurate retrieval results, and effectively integrate course data scattered in multiple systems.

[0051] 3. The present invention provides an intelligent multi-dimensional retrieval system for course data based on deep learning. Combined with a user portrait construction module and a personalized recommendation engine module, it can analyze the user's search history, browsing behavior and learning progress data, extract user features and construct user portraits. Based on collaborative filtering technology and user portraits, the system provides users with personalized course recommendations, achieves accurate personalized services, and improves user experience and course conversion rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0053] Figure 1 This is a schematic diagram of the functional modules of the system of the present invention;

[0054] Figure 2 Schematic diagram of the workflow of the semantic parsing unit of the present invention. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0056] Example 1, as Figure 1 、 Figure 2 As shown, the present invention provides an intelligent multi-dimensional retrieval system for course data based on deep learning, including a course data retrieval platform, which is communicatively connected to a knowledge graph construction module, a user intent analysis module, a user portrait construction module, a personalized recommendation engine module, and a retrieval display module, wherein the modules are connected by electrical signals;

[0057] The knowledge graph construction module is used to collect course data from multiple sources and construct a course domain knowledge graph. The knowledge graph represents the course knowledge system in the form of nodes and edges, and supports semantic association and reasoning. Through the knowledge graph, the system can quickly identify key entities and relationships in user queries, and perform semantic association and reasoning based on the graph to provide more comprehensive and accurate search results. The knowledge graph construction module includes a data integration unit and a knowledge graph unit;

[0058] Among them, the data integration unit is used to collect course data from multiple source systems, including the academic affairs system, online learning platform, teaching resource library, etc., and pre-process the collected data to ensure the quality and consistency of the data. Through API interface, direct database connection or web crawler technology, course data is collected from the academic affairs system, online learning platform and teaching resource library to ensure the integrity and timeliness of the data. According to the data format of different systems, appropriate parsing rules are designed to ensure that the data can be recognized by the system. At the same time, the data source and collection timestamp are recorded. Among them, the course outline, teacher information and teaching schedule are obtained from the academic affairs system. The online learning platform captures course videos, courseware, and user learning behavior data, extracts teaching materials, lesson plans, and academic papers from the teaching resource library, and cleans and preprocesses the collected data to ensure data quality and consistency. The cleaning process includes removing duplicate records, correcting erroneous data, and filling missing values. Duplicate course descriptions are detected and deleted using a hash algorithm, and malformed date fields are corrected using regular expressions. Data from different sources is standardized and correlated based on the unique identifier of the course ID. Data from multiple sources is then merged and stored in a unified data warehouse.

[0059] The knowledge graph unit constructs a course domain knowledge graph based on the preprocessed course data, which includes entities such as courses, teachers, knowledge points and learning objectives, as well as the relationships between entities, realizes cross-system data association, and integrates course data scattered in multiple systems. Based on the preprocessed course data, the key entities are identified through rule matching and BiLSTM-CRF deep learning model, including course entities, teacher entities, knowledge point entities and learning objective entities, and a unique entity identifier based on the course ID is defined. For course entities, attributes such as course name, course code, and credits are extracted, and the required / elective course types are marked. For teacher entities, the teacher's name, title, and department are identified, and the list of courses they teach is associated. For knowledge point entities, core knowledge points are extracted from the course syllabus and textbooks, and the subject areas to which they belong are marked. For learning objective entities, the ability requirements in the course description are parsed and converted into structured objectives. Through pattern matching and semantic analysis, a relationship network between entities is constructed. Entity relationships include course-teacher relationships, course-knowledge point relationships, course-learning objective relationships, and course- Course type relationships and cross-system relationships are stored using an attribute graph model to form a semantic network in the course domain, achieving deep association of cross-system data. For the course-teacher relationship, teachers are associated with the courses they teach and their roles are marked. For the course-knowledge point relationship, the course syllabus is parsed, an inclusion relationship between courses and knowledge points is established, and the weights of knowledge points are marked. For the course-learning objective relationship, the objectives in the course description are decomposed into a mapping of "course→learning objective". For cross-system relationships, the course schedule of the academic affairs system and the user learning records of the online learning platform are integrated to build a "course-learning behavior" relationship. Based on the unique identification of the course ID, entities with the same name in multi-source data are merged, including entity alignment and relationship fusion, to resolve data conflicts and build a complete course domain knowledge graph. For entity alignment, entities with similar names in different systems are matched using the cosine similarity algorithm to verify consistency. For relationship fusion, conflicting relationships are merged, and manual review is used to determine the final relationship. The fused course domain knowledge graph is then stored in the graph database to achieve global integration of cross-system course data.

[0060] The user intention parsing module is used to receive query information input by the user and perform semantic analysis on the user query using natural language processing technology in deep learning. It combines the course domain knowledge graph to extract intent features and entity relationships, obtain the intent deviation coefficient, analyze the degree of deviation between the user query intent and the system understanding, and further analyze the user's true intention. This solves the limitations of traditional retrieval systems that rely on keyword matching, improves semantic understanding capabilities, and accurately identifies ambiguous or implicit requirements in user queries. The user intention parsing module includes a user query input unit and a semantic parsing unit;

[0061] Among them, the user query input unit is used to receive natural language queries input by users, including inquiries about courses, course types and learning needs, etc., convert the user's query content into a data format that can be processed by the system, receive the user's natural language query input through a front-end interface including a Web form (including a text input box and a voice-to-text button) or a mobile input box (supporting keyboard input and voice recognition), and integrate a voice recognition API to convert voice into text in real time, mark the input method, and at the same time, generate a unique ID for each query, record the query timestamp, user ID, input method and original query text, and store it in the log database. The query content includes course name, course type and learning needs, etc., and supports text Input, voice input and other methods are used to meet the needs of different users. The natural language queries entered by users are pre-processed, including word segmentation and part-of-speech tagging, denoising and standardization, and stop word filtering. Among them, the Jieba word segmentation tool is used to split the query into vocabulary units, mark the part of speech, generate a word segmentation result list, and then remove irrelevant symbols and unified abbreviations. At the same time, spelling errors are corrected, stop word lists are loaded, words without actual semantics are removed, key entities are retained, and word segmentation results, standardized text, and filtered keyword lists are associated with the original query ID and stored in the preprocessing log table. The query process is supported backtracking, and user input is converted into clean and structured text fragments to improve the accuracy of subsequent intent recognition and entity extraction;

[0062] The semantic parsing unit is used to use the natural language processing technology in deep learning to perform semantic analysis on the natural language queries input by users, extract the key information in the query, and analyze the intent features and entity relationships in combination with the knowledge graph, calculate the intent deviation coefficient, map the user query to the specific entities and relationships in the knowledge graph, clarify the user's specific needs, accurately understand the deep intention of the user query, avoid ambiguity and misunderstanding, use the pre-trained BERT model to encode the pre-processed user query, generate context-related word vector representations, and fine-tune the intent classification and entity recognition tasks to extract key information in the query, including course name, course type, and learning needs, and classify the query into predefined intent categories. The model is constructed by using a sequence tagging model (BiLSTM-CRF) to identify entities and their attributes in the query and generate a structured representation. By converting natural language queries into structured semantic representations, the core needs of users are clarified, providing a basis for subsequent knowledge graph mapping. The extracted entities are matched with entities in the course domain knowledge graph, and the unique entity ID is determined through string similarity and context consistency verification. Based on the query intent and entity attributes, the relationship path required by the user is inferred. At the same time, the query scope is expanded by combining the attributes and relationships of the knowledge graph, and the user query is mapped to the specific entities and relationships in the knowledge graph. The semantic boundaries of the query are clarified to avoid ambiguity. The context information of the knowledge graph is combined to further refine the user intent and convert the refined The intent and entity relationship of the user are mapped into the query language of the knowledge graph, and then the intent features and entity relationships are extracted. Among them, the intent feature is the feature related to the query intent extracted from the user query, which is used to measure the degree of match between the user query intent and the system prediction intent. The intent features include query subject, query intent type, query constraint and query semantic context. The query subject is the core topic of the user query, the query intent type is the specific intent type of the user query, the query constraint is the constraint specified in the user query, and the query semantic context is the semantic context of the user query, including the background information and implicit intent of the query. The entity relationship is the semantic relationship between entities extracted from the user query, which is used to measure the matching between the entities in the user query and the knowledge graph. The matching degree between entities in the spectrum, entity relationships include course-teacher relationship, course-knowledge point relationship, course-learning goal relationship, course-course type relationship and cross-system relationship. The course-teacher relationship is the relationship between the course and the teacher, the course-knowledge point relationship is the relationship between the course and the knowledge point, the course-learning goal relationship is the relationship between the course and the learning goal, and the course-course type relationship is the relationship between the course and the course type. Based on the extracted intent features and entity relationships, the intent deviation coefficient is calculated to analyze the degree of deviation between the user's query intent and the system's understanding, and check whether the query results are consistent with the user's original intent. If there is a deviation, the user is prompted to supplement information, and accurate knowledge graph query statements are generated to ensure that the returned results are consistent with the user's deep intention. Figure 1 To improve the query accuracy;

[0063] In addition, the calculation process of the intention deviation coefficient is:

[0064] The actual value of the intent feature and the actual value of the entity relationship are extracted from the user query, and the expected value of the corresponding intent feature and the expected value of the entity relationship are obtained from the knowledge graph, and then the weight of each intent feature and entity relationship is determined. The weight is assigned according to the importance of the intent feature or entity relationship in the query. At the same time, the baseline deviation value is set according to the system requirements to normalize the final intent deviation coefficient. The baseline deviation value is a constant greater than 0. For each intent feature, the square of the deviation between its actual value and the expected value is calculated and divided by the corresponding weight to measure the difference between the actual value and the expected value of each intent feature. Its contribution to the total deviation is adjusted by the weight, and then all intent features are normalized. The squared deviations of the graph features are weighted and summed to obtain the intent feature deviation function. For each entity relationship, the squared deviation between its actual value and expected value is calculated and divided by the corresponding weight to measure the difference between the actual value and the expected value of each entity relationship. Its contribution to the total deviation is adjusted by the weight. The squared deviations of all entity relationships are then weighted and summed to obtain the entity relationship deviation function. The intent feature deviation function and the entity relationship deviation function are added together to obtain the total squared deviation sum of the intent features and entity relationships. The square root is taken to convert the squared deviation sum into a single value to measure the size of the overall deviation. The deviation coefficient is then divided by the baseline deviation value and normalized to the range of [0, 1] to obtain the intent deviation coefficient. If the intent deviation coefficient is close to 0, it means that the deviation between the user query intent and the system understanding is small and the match is high. If the intent deviation coefficient is close to 1, it means that the deviation is large and the match is low.

[0065] The calculation expression of the intention deviation coefficient is:

[0066] ;

[0067] Where, is the intention deviation coefficient, which indicates the degree of deviation between the user's query intention and the system's understanding. For the The actual value of the intent feature (the intent feature extracted from the user query), For the The expected value of an intent feature (the intent feature determined based on the knowledge graph), For the The weight of an intent feature reflects its importance in the query. For the The actual value of the entity relationship (the entity relationship extracted from the user query), For the The expected value of entity relationships (entity relationships determined based on the knowledge graph), For the The weight of an entity relationship reflects its importance in the query. is the benchmark deviation value, which is used to normalize the results and ensure that the deviation coefficient is within a reasonable range. is the number of intent features, is the number of entity relationships;

[0068] User portrait construction module, used to extract user features and construct user portraits by analyzing user search history, browsing behavior, and learning progress data;

[0069] The personalized recommendation engine module is used to combine the collaborative filtering technology in deep learning with the constructed user portrait to provide users with personalized course recommendations and achieve accurate personalized recommendations;

[0070] The search display module is used to combine the results of the personalized recommendation engine to generate a search result list, sort and page the search results. Through multi-dimensional search, users can accurately filter courses according to their needs.

[0071] Example 2, as Figure 1 、 Figure 2 As shown, based on Example 1, the present invention provides a technical solution: preferably, the user portrait construction module specifically includes:

[0072] Collect user data from multiple channels, including search history, browsing behavior, and learning progress. For search history, collect user historical query records, including query time, keywords, and clicked course IDs. For browsing behavior, record the course pages visited by users, length of stay, and page jump paths. For learning progress, synchronize user course learning data and extract user features from the collected user data, including search preferences, browsing depth, and learning progress. For search preferences, count high-frequency search keywords, calculate keyword TF-IDF values, and identify users' core interest areas. For browsing depth, calculate the average length of stay on the course page and bounce rate to evaluate the user's interest in the content. For learning progress, extract learning completion rate, exercise accuracy rate, and learning frequency, and then generate user tags including knowledge areas, learning stages, and behavioral preferences. Convert user behavior into structured tags to form an interpretable user feature representation. Integrate the long-term behavioral characteristics of users from the past three months to generate static tags. Combined with the user's real-time behavior on the day, generate short-term interest tags. By setting a weight of 70% for static tags and 30% for short-term interest tags, weighted fusion is performed to generate the final user profile.

[0073] The personalized recommendation engine module specifically includes:

[0074] Extract static tags and short-term interest tags from user portraits, convert them into vector representations, splice user behavior features to form user feature vectors, extract course attributes, convert them into vector representations, and then splice course feature vectors. Use the double-tower model to perform nonlinear mapping on user and course features respectively to generate low-dimensional dense vectors, maximize user-course matching, convert user portraits and course attributes into computable vector spaces, and provide high-quality feature inputs for collaborative filtering. Based on user-based collaborative filtering, calculate user similarity and recommend courses that similar users like. Based on the cosine similarity of course vectors, recommend courses similar to the user's historical learning courses. Fuse the interactive signals of collaborative filtering (user-course click matrix) and deep features (points between user vectors and course vectors) The algorithm uses the product of the user's click probability on the course, and the loss function is the cross-entropy loss. The optimization goal is to maximize the click-through rate of the recommended courses. The collaborative filtering top 50 course recommendation results and the deep learning top 30 course recommendation results are combined through weighted sorting to generate a candidate recommendation list. Combined with the user's current behavior, the short-term interest tag is updated, and the weight of the course corresponding to the short-term interest tag is increased by 30%. The course is inserted at the front of the candidate recommendation list, and the user's click on the recommended course, learning time and other behaviors are recorded. The model parameters are updated in real time through online learning. If the user clicks on the introductory deep learning course, the association weight between the course and the user's short-term interest tag is enhanced, achieving real-time, diverse and dynamic optimization of the recommendation results, improving user satisfaction and course conversion rate.

[0075] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A deep learning-based intelligent multi-dimensional course data retrieval system, including a course data retrieval platform, characterized by: The course data retrieval platform is communicatively connected to a knowledge graph construction module, a user intent analysis module, a user portrait construction module, a personalized recommendation engine module, and a retrieval display module, wherein the modules are electrically connected; The knowledge graph construction module is used to collect course data from multiple sources and construct a course domain knowledge graph; The user intent analysis module is used to receive query information input by the user and perform semantic analysis on the user query using natural language processing technology in deep learning. It combines the course domain knowledge graph to extract intent features and entity relationships, obtain the intent deviation coefficient, analyze the degree of deviation between the user query intent and the system's understanding, and further analyze the user's true intent. The user profile building module is used to extract user features and build a user profile by analyzing the user's search history, browsing behavior and learning progress data; The personalized recommendation engine module is used to combine the collaborative filtering technology in deep learning and the constructed user portrait to provide users with personalized course recommendations; The search display module is used to combine the results of the personalized recommendation engine to generate a search result list, sort the search results and paginate them.

2. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 1, characterized in that: The knowledge graph construction module includes a data integration unit and a knowledge graph unit; The data integration unit is used to collect course data from multiple source systems and pre-process the collected data; The knowledge graph unit constructs a course domain knowledge graph based on the preprocessed course data, which includes entities of courses, teachers, knowledge points and learning objectives, as well as the relationships between the entities.

3. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 2 is characterized by: The data integration unit specifically includes: Collect course data from the educational system, online learning platform, and teaching resource library through API interfaces, direct database connections, or web crawler technology. Design appropriate parsing rules for the data formats of different systems, and record the data source and collection timestamp. Clean and preprocess the collected data and unify data from different sources into a standardized format; Based on the unique identification of the course ID, multi-source data is associated and matched, and then the same course data from different sources are merged and stored in a unified data warehouse.

4. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 2, characterized in that: The knowledge graph unit specifically includes: Based on the preprocessed course data, rule matching and a BiLSTM-CRF deep learning model are used to identify key entities, including course entities, teacher entities, knowledge point entities, and learning objective entities. Unique entity identifiers based on course IDs are also defined. Through pattern matching and semantic analysis, a relationship network between entities is constructed. Entity relationships include course-teacher relationships, course-knowledge point relationships, course-learning objective relationships, course-course type relationships, and cross-system relationships. An attribute graph model is used to store these relationships, forming a semantic network for the course domain. Based on the unique identification of the course ID, entities with the same name in multi-source data are merged, including entity alignment and relationship fusion, to build a complete course domain knowledge graph.

5. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 2, characterized in that: The user intention parsing module includes a user query input unit and a semantic parsing unit; The user query input unit is configured to receive natural language queries input by users, including queries about courses, course types, and learning needs, and convert the user's query content into a data format that can be processed by the system; The semantic parsing unit is used to use natural language processing technology in deep learning to perform semantic analysis on natural language queries input by users, extract key information in the queries, and analyze intent features and entity relationships in combination with the knowledge graph, calculate the intent deviation coefficient, map the user queries to specific entities and relationships in the knowledge graph, and clarify the user's specific needs.

6. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 5, characterized in that: The user query input unit specifically includes: Receive natural language queries entered by users through a front-end interface including a web form or mobile input box, and integrate a speech recognition API to convert speech into text in real time, annotate the input method, and generate a unique ID for each query, record the query timestamp, user ID, input method and original query text, and store them in a log database. The query content includes course name, course type and learning needs; Preprocessing of natural language queries entered by users, including word segmentation and part-of-speech tagging, denoising and normalization, and stop word filtering. Jieba word segmentation tools are used to split queries into lexical units, annotate parts of speech, and generate a list of word segmentation results. This removes irrelevant symbols and standardized abbreviations, corrects spelling errors, loads a stop word list, removes words without actual meaning, and retains key entities. The word segmentation results, standardized text, and filtered keyword list are associated with the original query ID and stored in the preprocessing log table.

7. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 5 is characterized by: The semantic parsing unit specifically includes: Use the pre-trained BERT model to encode pre-processed user queries and generate context-sensitive word vector representations. Fine-tune the training for intent classification and entity recognition tasks to extract key information from the query, including the course name, course type, and learning requirements. Then, classify the query into predefined intent types. Use a sequence labeling model to identify entities and their attributes in the query and generate a structured representation. The extracted entities are matched with entities in the course domain knowledge graph. Unique entity IDs are determined through string similarity and context consistency verification. Based on the query intent and entity attributes, the relationship path required by the user is inferred. At the same time, the query scope is expanded by combining the attributes and relationships of the knowledge graph, mapping the user query to specific entities and relationships in the knowledge graph, and clarifying the semantic boundaries of the query. Combined with the contextual information of the knowledge graph, the user intent is further refined, and the refined intent and entity relationships are mapped into the query language of the knowledge graph, thereby extracting intent features and entity relationships. Intent features include query subject, query intent type, query constraints, and query semantic context. Entity relationships include course-teacher relationships, course-knowledge point relationships, course-learning objective relationships, course-course type relationships, and cross-system relationships. Based on the extracted intent features and entity relationships, the intent deviation coefficient is calculated to analyze the degree of deviation between the user's query intent and the system's understanding. It is checked whether the query results are consistent with the user's original intent. If there is a deviation, the user is prompted to provide additional information to generate accurate knowledge graph query statements.

8. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 7, characterized in that: The calculation process of the intention deviation coefficient is: Extract the actual value of the intent feature and the actual value of the entity relationship from the user query, and obtain the expected value of the corresponding intent feature and the expected value of the entity relationship from the knowledge graph, and then determine the weight of each intent feature and entity relationship. At the same time, set the baseline deviation value according to system requirements; For each intent feature, calculate the square of the deviation between its actual value and the expected value and divide it by the corresponding weight. Then, sum the weighted squares of the deviations of all intent features to obtain the intent feature deviation function. For each entity relationship, calculate the square of the deviation between its actual value and the expected value and divide it by the corresponding weight. Then, sum the weighted squares of the deviations of all entity relationships to obtain the entity relationship deviation function. Adding the intent feature deviation function and the entity relationship deviation function together yields the sum of squared deviations of the intent feature and entity relationship. Taking the square root of the sum is then divided by the baseline deviation value to obtain the intent deviation coefficient.

9. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 5, characterized in that: The user portrait construction module specifically includes: Collect user data from multiple channels, including search history, browsing behavior, and learning progress. For search history, collect user historical query records, including query time, keywords, and clicked course IDs. For browsing behavior, record the course pages visited by users, length of stay, and page jump paths. For learning progress, synchronize user course learning data; Extract user features from the collected user data, including search preferences, browsing depth, and learning progress. For search preferences, count high-frequency search keywords, calculate keyword TF-IDF values, and identify users' core areas of interest. For browsing depth, calculate the average duration of stay on the course page and bounce rate to assess users' interest in the content. For learning progress, extract learning completion rate, exercise accuracy rate, and learning frequency, and then generate user tags that include knowledge domain, learning stage, and behavioral preferences. Convert user behavior into structured tags to form an interpretable user feature representation. Integrate the long-term behavioral characteristics of users in the past three months to generate static tags, and combine them with the user's real-time behavior on the day to generate short-term interest tags. By setting the weight of static tags to 70% and the weight of short-term interest tags to 30%, weighted fusion is performed to generate the final user portrait.

10. The intelligent multi-dimensional retrieval system for course data based on deep learning according to claim 9, characterized in that: The personalized recommendation engine module specifically includes: Extract static tags and short-term interest tags from user portraits, convert them into vector representations, and combine them with user behavior features to form user feature vectors. Course attributes are also extracted and converted into vector representations, which are then combined with course feature vectors. A dual-tower model is used to perform nonlinear mapping on user and course features, generating low-dimensional dense vectors to maximize user-course matching. Based on user collaborative filtering, we calculate user similarity and recommend courses that similar users like. We also recommend courses similar to the user's previous courses based on the cosine similarity of course vectors. We integrate collaborative filtering interaction signals and deep features to analyze the user's click probability on courses. We then combine the collaborative filtering top 50 course recommendations with the deep learning top 30 course recommendations through weighted sorting to generate a candidate recommendation list. Combined with the user's current behavior, the update of short-term interest tags is triggered, the weight of the courses corresponding to the short-term interest tags is increased by 30%, and they are inserted at the front of the candidate recommendation list. The user's clicks on the recommended courses and the length of study time are recorded.

Images