ICT learning resource dynamic intention recognition semantic and time enhancement recommendation method

Abstract

The invention discloses an ICT (information and communication technology) learning resource dynamic intention recognition semantic and time enhancement recommendation method, which comprises the following steps of: proposing learning requirements by a user, finding out a semantic-related resource list from ICTKB by using a semantic model, analyzing user query by using a large model, judging learning intentions (such as entry, depth, practice and the like) of the user, and recommending the learning intentions (such as entry, depth, practice and the like) of the user. And judging whether the query has timeliness sensitive words or not, calculating a timeliness score, calculating a comprehensive score in combination with semantic similarity, the identified intention (and the corresponding weight), the resource characteristics (popularity, credibility and difficulty) and the timeliness score, and reordering the recall list. And presenting the resource list which is subjected to intelligent reordering and most meets the intention and timeliness demand of the user to the user. According to the method, the dynamic learning intention can be deeply understood, high-quality learning resources advancing with times can be accurately matched, and the method can adapt to intelligent recommendation of rapid iteration characteristics of domain knowledge.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technology, and in particular to a method for dynamic intent recognition semantics and time-enhanced recommendation of ICT learning resources. Background Art

[0002] In recent years, information and communication technology (ICT) has developed and iterated at an unprecedented rate. Emerging technologies such as artificial intelligence, cloud computing, big data, and the Internet of Things have continuously emerged, driving the explosive growth of related knowledge systems. This has brought tremendous opportunities to professional learners, researchers, and practitioners in the ICT field, but it is also accompanied by severe challenges. Faced with massive, heterogeneous, and multi-source learning resources (such as online tutorials, technical blogs, research papers, open source project documentation, video lectures, etc.), learners often find it difficult to efficiently and accurately obtain timely and high-quality content that meets their current needs. Existing information retrieval and recommendation methods have exposed significant shortcomings when dealing with the characteristics unique to the ICT field, such as rapid knowledge updates, complex and diverse terminology, and dynamic and changing learning needs. Specifically, they are as follows:

[0003] 1) Information overload and difficulty in screening: There are a huge number of resources, but the quality is uneven. A large amount of low-value or even outdated information interferes with the acquisition of effective knowledge.

[0004] 2) Insufficient retrieval accuracy: Traditional methods have difficulty in deeply understanding the complex, multi-stage learning intentions behind user queries, resulting in a low degree of match between recommendation results and users' actual needs.

[0005] 3) Poor timeliness: ICT knowledge is highly time-sensitive. Existing systems cannot effectively distinguish and process the time sensitivity of different types of knowledge (such as basic theory vs. cutting-edge technology), making it difficult to recommend the latest and most relevant resources.

[0006] 4) Lack of content quality assessment: There is a lack of a reliable evaluation mechanism for the teaching value and professional depth of the learning resources themselves, which makes it easy to be misled by superficial indicators (such as click volume).

[0007] In order to solve the above problems, existing technologies mainly try to solve them through the following ways:

[0008] 1) Traditional information retrieval based on keyword matching: This is exemplified by the BM25 algorithm, which primarily relies on matching and ranking keywords in user queries with statistical features such as word frequency and inverse document frequency in document content. This approach is the foundation of search engines, is easy to implement, and can handle large amounts of data.

[0009] 2) Semantic Retrieval Based on Vector Representation: In recent years, with the development of deep learning, embedding methods based on pre-trained language models, such as Contriever, have emerged. These methods map queries and documents into a high-dimensional vector space and determine semantic relevance by calculating vector similarity, improving the ability to understand query intent, especially in general domains.

[0010] 3) Recommendation systems that incorporate user behavior or context: These include classic collaborative filtering and content-based recommendation, as well as more complex sequential recommendation models (such as TiM4Rec), multimodal recommendation models (such as HM4SR), and temporal context-aware models (such as MOJITO and TAIW). These models attempt to predict user interests based on historical user behavior, timestamps, content features, and even multimodal information, enabling personalized recommendations. Some work has also begun to focus on the interpretability of recommendations.

[0011] 4) Construction and application of domain knowledge bases: For specific professional fields (such as ICT), some work attempts to build structured or unstructured knowledge bases and use technologies such as knowledge graphs to assist in retrieval and recommendation, in order to improve professionalism and accuracy.

[0012] Although existing technologies have made some progress in the field of general retrieval and recommendation, they still have obvious limitations and shortcomings in meeting the needs of ICT learners for specialized, dynamic, and high-quality learning resources.

[0013] (1) Insufficient depth and precision of semantic understanding:

[0014] Traditional methods like BM25 are sensitive to lexical matching, making them difficult to handle synonyms and polysemy, and even more so, unable to understand the deeper semantics and learning objectives behind queries. They also perform poorly for the numerous specialized terms and abbreviations in the ICT field. While models like Contriever improve general semantic understanding, they fail to optimize for the specific terminology, knowledge structure, and rapidly evolving concepts of the ICT field, resulting in mediocre performance for specialized queries. They also often overlook the evolving nature of intent within the specific context of "learning." For example, users may have a chain of learning needs, from "understanding a concept" to "finding an introductory tutorial," to "practical case analysis," and finally "in-depth research." Existing models lack effective modeling of these dynamic learning paths. Most recommender systems (such as HM4SR and MOJITO) focus on user consumption behaviors (such as shopping and watching videos) rather than systematic knowledge acquisition. While they may capture short-term interest fluctuations, they struggle to identify and meet the well-defined, structured, and multi-stage "learning intent" of ICT learners.

[0015] (2) The timeliness processing mechanism is rough and unsuitable:

[0016] Many time-aware models (such as TiM4Rec) use a simple timestamp decay strategy, treating all content equally. This ignores the complexity of knowledge in the ICT field: fundamental theories (such as TCP / IP principles) may remain valid for a long time, while cutting-edge technologies (such as the application of a new framework) may quickly become outdated. Uniform decay cannot meet the recommendation needs of this structured, long-term knowledge and dynamic, cutting-edge knowledge. Even though models such as MOJITO and TAIW have deepened their temporal modeling, their design scenarios (such as news recommendation) still differ from the needs of ICT learning resource recommendation. They fail to fully support differentiated timeliness weighting based on whether the query itself is time-sensitive or the inherent temporal properties of the resource content itself (such as basic concepts vs. the latest developments).

[0017] (3) The content quality assessment system is weak and unreliable:

[0018] Existing recommendation systems widely rely on user behavior data such as click-through rate, number of collections, and number of likes as proxy indicators of content quality. However, these indicators are easily interfered with by factors such as "clickbait," marketing promotion, and group effects, and cannot truly reflect the teaching value, accuracy, and depth of the content. Collaborative filtering and content-based recommendation methods are essentially based on user preferences or content similarity. They cannot judge whether a resource is truly suitable for learning, whether the explanation is clear, whether practical guidance is provided, and other key teaching elements. Even work that focuses on explainability often stops at explaining the "reasons" for recommendations (such as user behavior correlations) rather than explaining the "quality" of the resource itself, lacking verifiable quality assurance. Summary of the Invention

[0019] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0020] To this end, the present invention proposes a dynamic intention recognition semantic and time-enhanced recommendation method for ICT learning resources, providing ICT learners with an intelligent recommendation system that can deeply understand their dynamic learning intentions, accurately match high-quality and up-to-date learning resources, and adapt to the rapid iteration characteristics of domain knowledge.

[0021] Another object of the present invention is to propose an ICT learning resource dynamic intent recognition semantic and time-enhanced recommendation system.

[0022] To achieve the above objectives, the present invention proposes a method for dynamic intent recognition, semantics, and temporal enhancement recommendation of ICT learning resources, including:

[0023] Build an ICT knowledge base containing multi-source heterogeneous ICT knowledge;

[0024] Conduct a preliminary search in the ICT knowledge base based on user demand information to generate a list of semantically related resources;

[0025] Use the big model to analyze user demand information to determine the user's learning intention, and assign an intention weight to each resource in the resource list based on the determined learning intention;

[0026] Detect timeliness-sensitive words in user demand information and calculate the timeliness score of each resource in the resource list;

[0027] Calculate the comprehensive score of each resource in the resource list based on the generated semantic similarity, intent weight, timeliness score, and resource characteristics, and re-sort the resource list according to the comprehensive score;

[0028] Based on the re-ranking results, the resource list that best meets the user's intention and timeliness requirements is recommended to the user as the search results.

[0029] The ICT learning resource dynamic intent recognition semantic and time-enhanced recommendation method according to the embodiment of the present invention may also have the following additional technical features:

[0030] In one embodiment of the present invention, building an ICT knowledge base containing multi-source heterogeneous ICT knowledge includes:

[0031] We use crawler technology to extract high-frequency domain terms from authoritative academic resources as seed keywords, establish a multi-dimensional classification system based on domain characteristics, and then use a large language model to perform two-stage semantic expansion to output a keyword set.

[0032] Based on the keyword set, browser automation technology is used to simulate complete user behavior. By executing native JavaScript expressions in the browser environment, encrypted parameters dynamically generated by the website are directly obtained. The collection strategy is customized according to the characteristics of different platforms to collect multi-platform data.

[0033] Define a unified ICT knowledge representation data structure and map multi-platform data to standardized fields. Then, deduplicate data based on unique identifiers. After deduplication, mark or degrade resources older than a preset age, and directly remove outdated content. At the same time, apply a regular expression rule library combined with text features to identify and filter low-quality content to obtain standardized data.

[0034] A dual-channel hybrid scoring model including the three dimensions of popularity, credibility, and difficulty is constructed to perform multi-dimensional scoring on each resource of standardized data to obtain a three-dimensional scoring result. Multi-dimensional labels are automatically generated through a knowledge-guided prompt project based on the six-dimensional label system, resource content, and three-dimensional scoring results.

[0035] In one embodiment of the present invention, a two-stage semantic expansion is performed using a large language model to output a keyword set, including:

[0036] Classify and verify the seed keywords to obtain classification verification results;

[0037] Based on the classification verification results and structured hint engineering, multiple rounds of knowledge expansion are performed to generate candidate keywords;

[0038] The generated candidate keywords are deduplicated and category balanced to output a keyword set.

[0039] In one embodiment of the present invention, a preliminary search is performed in the ICT knowledge base based on user demand information to generate a list of semantically related resources, including:

[0040] Based on the weighted segmentation coding strategy, the standardized data resources are decomposed into three parts: core content, label information and detailed content, and different weights are assigned to the three parts;

[0041] Each part is vectorized separately, and the vector representations of the three parts are synthesized according to the weights to obtain the final resource vector representation.

[0042] In one embodiment of the present invention, a large model is used to analyze user demand information to determine the user's learning intention, and an intention weight is assigned to each resource in the resource list according to the determined learning intention, including:

[0043] Seven ICT learning intentions were defined, and the Qwen-2.5-Max model was used as the intent identifier. Specific prompt words were used to guide the model in determining the intent category to which the user's desired information belongs.

[0044] Each intent is assigned a specific weight configuration, and a preset formula is used to calculate the total score based on the intent when the intent weight is actually calculated.

[0045] In one embodiment of the present invention, detecting timeliness-sensitive words in user demand information and calculating the timeliness score of each resource in the resource list includes:

[0046] Check whether the query text contains time-sensitive keywords and store the time-sensitive keywords as a keyword list;

[0047] Based on the keyword list, the timeliness score is calculated according to the resource release time using the time decay calculation method; it includes two different sets of decay curves for timeliness-sensitive queries and non-sensitive queries respectively.

[0048] In one embodiment of the present invention, a comprehensive score for each resource in the resource list is calculated based on the generated semantic similarity, intent weight, timeliness score, and resource characteristics, including:

[0049] Obtain multiple metrics, including similarity score, timeliness score, resource feature score, and intent type;

[0050] Obtain the corresponding weight configuration based on the intent type, and then calculate the weighted sum of each indicator to obtain the weighted sum calculation result;

[0051] Based on the weighted sum calculation results and according to whether the query is time-sensitive, the timeliness weight is adjusted, and proportional scaling is used to ensure that the sum of the weights of each component is 1, and finally a comprehensive score in the range of 0-1 is generated.

[0052] In one embodiment of the present invention, the method further includes:

[0053] Timeliness score: quantifies the timeliness of search results through weighted calculation;

[0054] Large model scoring: Two large language models, Qwen-2.5-Max and DeepSeek-v3, are used to evaluate the overall satisfaction of retrieval results based on user demand information and retrieval results;

[0055] Intent satisfaction: Use a large model to evaluate whether the retrieval results can meet the original intention of the user's demand information based on user demand information, intent, and retrieval results.

[0056] To achieve the above objectives, the present invention further proposes a dynamic intent recognition semantic and time-enhanced recommendation system for ICT learning resources, comprising:

[0057] ICT knowledge base construction module, used to build an ICT knowledge base containing multi-source heterogeneous ICT knowledge;

[0058] Semantic retrieval module, used to perform preliminary recall in the ICT knowledge base based on user demand information to generate a list of semantically related resources;

[0059] Dynamic intent recognition module, which uses a large model to analyze user demand information to determine the user's learning intention and assigns an intention weight to each resource in the resource list based on the determined learning intention;

[0060] Timeliness processing module, used to detect timeliness-sensitive words in user demand information and calculate the timeliness score of each resource in the resource list;

[0061] An enhanced re-ranking module is used to calculate the comprehensive score of each resource in the resource list based on the generated semantic similarity, intent weight, timeliness score, and resource characteristics, and re-rank the resource list according to the comprehensive score;

[0062] The result recommendation module is used to recommend the resource list that best meets the user's intention and timeliness requirements as the search results to the user based on the re-ranking results.

[0063] The dynamic intent recognition semantic and time-enhanced recommendation method and system for ICT learning resources in this embodiment of the present invention utilizes advanced semantic vector models to perform a preliminary search for highly relevant resources. A re-ranking mechanism is then designed based on the specific characteristics of learning intent and timeliness in the ICT field. This mechanism accurately identifies the learning intent behind the query, dynamically adjusts the weights of ranking factors, and distinguishes time-sensitive scenarios, applying a differentiated, segmented timeliness compensation strategy to ensure that recommendation results keep pace with technological developments.

[0064] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0066] Figure 1 is a flow chart of a method for dynamic intent recognition semantic and time-enhanced recommendation of ICT learning resources according to an embodiment of the present invention;

[0067] Figure 2 is a typical interaction flow chart used by a user according to an embodiment of the present invention;

[0068] Figure 3 is a flow chart of knowledge base construction according to an embodiment of the present invention;

[0069] Figure 4 is a schematic diagram of a data structure JSON according to an embodiment of the present invention;

[0070] Figure 5 2. It is a schematic diagram of an intention recognition prompt word template according to an embodiment of the present invention;

[0071] Figure 6 It is a diagram showing the design of the intention weight matrix and the heat map and radar chart according to an embodiment of the present invention;

[0072] Figure 7 2. It is a schematic diagram of a large model scoring prompt word template according to an embodiment of the present invention;

[0073] Figure 8 2. Schematic diagram of an intention satisfaction prompt word template according to an embodiment of the present invention;

[0074] Figure 9 is a schematic diagram of a case analysis according to an embodiment of the present invention;

[0075] Figure 10 2 is a structural diagram of a dynamic intent recognition semantic and time-enhanced recommendation system for ICT learning resources according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0077] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0078] The following describes the method and system for dynamic intent recognition semantics and time-enhanced recommendation of ICT learning resources proposed in accordance with an embodiment of the present invention with reference to the accompanying drawings.

[0079] Figure 1 FIG. 1 is a flowchart of a method for dynamic intention recognition semantic and time-enhanced recommendation of ICT learning resources according to an embodiment of the present invention. Figure 1 As shown, the method includes but is not limited to the following steps:

[0080] S1, building an ICT knowledge base containing multi-source heterogeneous ICT knowledge;

[0081] S2, based on user demand information, conducts preliminary recall in the ICT knowledge base to generate a list of semantically related resources;

[0082] S3, using the big model to analyze user demand information to determine the user's learning intention, and assign an intention weight to each resource in the resource list based on the determined learning intention;

[0083] S4, detect timeliness sensitive words in user demand information and calculate the timeliness score of each resource in the resource list;

[0084] S5, calculates the comprehensive score of each resource in the resource list based on the generated semantic similarity, intention weight, timeliness score and resource characteristics, and re-sorts the resource list according to the comprehensive score;

[0085] S6, based on the re-ranking results, recommends the resource list that best meets the user's intention and timeliness requirements to the user as the search results.

[0086] It is understandable that if Figure 2 As shown, the typical interaction process used by users:

[0087] (1) User input query: The user proposes a learning need (such as entering a question or keyword).

[0088] (2) Semantic retrieval and recall: The system receives a query and uses the semantic model to find a list of semantically relevant resources from the ICTKB.

[0089] (3) Dynamic intent recognition: The system uses a large model to analyze user queries and determine their learning intentions (such as entry, in-depth, practice, etc.).

[0090] (4) Timeliness processing: The system determines whether the query contains timeliness-sensitive words and calculates the timeliness score.

[0091] (5) Enhanced re-ranking: The system combines semantic similarity, identified intent (and its corresponding weight), resource characteristics (popularity, credibility, difficulty) and timeliness score to calculate a comprehensive score and re-rank the recall list.

[0092] (6) Result recommendation: The system presents to the user a list of resources that has been intelligently re-sorted and best meets the user's intentions and timeliness requirements.

[0093] The knowledge base construction process of the present invention is as follows Figure 3 shown.

[0094] In one embodiment of the present invention, a multi-source heterogeneous ICT knowledge base (ICTKB) is constructed.

[0095] (1) Construction of retrieval keyword set based on the collaboration of domain features and large language model

[0096] First, we use crawler technology to extract high-frequency domain terms from authoritative academic resources (such as master's and doctoral dissertations in CNKI) as seed keywords. We then establish a multidimensional classification system based on domain characteristics (such as the strong coupling between ICT and AI). Finally, we apply a large language model to perform a two-stage semantic expansion: First, we perform classification verification on the seed terms. Then, we perform multiple rounds of knowledge expansion based on a structured hinting process (requiring new and specific terms) to generate candidate keywords. Finally, we perform duplicate removal and category balancing on the generated keywords.

[0097] Compared to traditional keyword construction methods that rely on manual or simple statistics, this method combines the semantic understanding capabilities of a large model with prior domain knowledge (seed words, classification system) to systematically generate a keyword set that is broad in coverage, closely aligned with the latest technological advances, and highly specific. The intent suffix system further enhances the adaptability of keywords to the search mechanisms of specific platforms (such as Bilibili and Zhihu, which are title-sensitive and core data-driven), improving the accuracy and efficiency of subsequent data collection.

[0098] (2) Multi-platform data collection based on browser automation and native JS calls

[0099] The present invention independently designed a modular crawler framework that supports multi-platform (such as Bilibili and Zhihu) extensions, adopts browser automation technology (Playwright) as the core, simulates complete user behavior (including code scanning login, session persistence, and page interaction) to break through anti-crawling restrictions, and directly obtains encrypted parameters dynamically generated by the website by executing native JavaScript expressions in the browser environment, and customizes collection strategies for different platform characteristics (such as Bilibili video metadata, Zhihu answer depth and author authority).

[0100] Unlike traditional crawlers that rely on HTTP requests and require reverse engineering to crack encryption algorithms, this invention leverages browser automation to effectively circumvent complex anti-crawling mechanisms (particularly those based on session and user behavior detection). By directly calling front-end JavaScript functions to obtain encrypted parameters, this significantly reduces development and maintenance costs and improves the crawler's stability and adaptability. Its modular design and platform customization ensure efficient and accurate acquisition of the required multi-dimensional data from different platforms.

[0101] (3) Systematic multi-source heterogeneous data integration and quality control

[0102] This invention defines a unified ICT knowledge representation data structure (JSON format), maps data from different sources to standardized fields, and then deduplicates the data based on unique identifiers. After integrating the crawler data, multi-level timeliness filtering can be performed, marking or scoring resources that exceed a set age limit (such as 2 years), and directly removing outdated content (such as before 2020). At the same time, a regular expression rule library is applied in combination with text features (such as length) to identify and filter low-quality content (such as "clickbait"). Figure 4 It is a schematic diagram of the data structure JSON implemented in the present invention.

[0103] This paper proposes a systematic integration and quality control process for multi-source online learning resources. Compared to simple merging or basic deduplication, this approach significantly improves the standardization, timeliness, and reliability of incoming data through unified formatting, unique ID-based deduplication, a clear time-sensitive attenuation / filtering mechanism, and rule-based identification of low-quality content. This lays the foundation for the subsequent construction of a high-quality knowledge base.

[0104] (4) Resource quality assessment based on a multi-dimensional hybrid scoring and labeling system

[0105] The present invention constructs a dual-channel hybrid scoring model of "formula calculation + large model evaluation", which includes three dimensions: popularity (weighted calculation of playback, likes, comments, etc. in combination with platform characteristics), credibility (based on author information), and difficulty (dominated by large model semantic understanding, combined with title and content analysis). In addition, a six-dimensional labeling system (domain, content type, difficulty level, target audience, timeliness, and learning path) is designed simultaneously. Utilizing a large language model, combined with resource content and three-dimensional scoring results, multi-dimensional labels are automatically generated through a knowledge-guided prompting project.

[0106] Unlike solely relying on platform metrics or purely subjective assessments, the dual-channel scoring model proposed in this paper combines objective data with the deep semantic understanding capabilities of a large model, enabling a more comprehensive and interpretable quantitative assessment of resource quality. The six-dimensional tagging system, combined with large-scale model automatic generation technology, provides rich metadata for resources, significantly enhancing the knowledge base's organization, retrieval, and personalized recommendation capabilities. Scoring results guide tag generation, demonstrating the inherent connection between assessment and organization.

[0107] In one embodiment of the present invention, a key technology of retrieval and re-ranking oriented to knowledge services is proposed.

[0108] Specifically, this paper proposes a re-ranking method based on dynamic intent recognition and segmented timeliness compensation, aiming to optimize the result list of semantic search to better meet the user's deep learning needs and the characteristics of ICT domain knowledge. The method mainly includes the following key parts:

[0109] (1) Retrieval

[0110] The core retrieval module uses BGE-Large-zh as the main semantic retrieval model, which performs well in Chinese tasks and achieves high-quality resource recall through vector similarity matching. In order to adapt to the diversity of ICT educational resources, the present invention designs a set of weighted segmented coding strategies to decompose resources into three parts: core content (title, introduction), tag information, and detailed content, and assigns different weights (determined to be 0.55, 0.30, and 0.15 after experimental tuning). The vectorization process of each part is carried out separately, and then the final representation is synthesized according to the weight. This method significantly improves the retrieval accuracy, especially when dealing with structured educational resources. The specific formula is as follows:

[0111]

[0112] Among them, f BGErepresents the BGE model encoding function, which can represent the original text into a high-dimensional vector (1024 dimensions based on the model characteristics). In addition, each part of the vectorization method uses a prefix prompt word prefix: "Generate a representation for this sentence to retrieve similar text:" to improve the quality of the vectorization result. In the formula, i represents the core content (title), tag information (tag), and detailed content (content). w i Represents the weights of different parts, and finally obtains the vector representation E corresponding to each retrieval object.

[0113] (2) Reordering

[0114] Based on real-world query characteristics, this article defines seven typical ICT learning intents in the dataset construction section. To apply these intent characteristics to reranking, the system assigns specific weights to each intent, reflecting the resource characteristics preferences of different learning objectives. The system uses the Qwen-2.5-Max large model as the intent identifier, using carefully designed specific prompts to guide the model in accurately determining the intent category to which the user's query belongs. Figure 5 It is the intention recognition prompt word template, Figure 6 It is the intention weight matrix design and heat map and radar map representation. Figure 7 It is a template for large model scoring prompts. Figure 8 It is an intention satisfaction prompt word template.

[0115] The result of intent recognition will affect the weight allocation of subsequent resource sorting. For example, for the "beginner learning" intent, the system will assign a similarity weight of 0.4, a heat weight of 0.3, a credibility weight of 0.2, and a difficulty weight of -0.1 (negative values ​​indicate that low-difficulty resources are recommended first); while for the "in-depth learning" intent, it will assign a similarity weight of 0.3, a heat weight of 0.1, a credibility weight of 0.3, and a difficulty weight of 0.3 (positive values ​​indicate that high-difficulty resources are recommended first). When actually calculating the intent weight, the following formula is used to calculate the total score based on the intent:

[0116]

[0117] Where q represents the query, d represents the recalled document, and w represents the weight vector in the intent-based weight matrix (w sim is the similarity dimension, w pop is the heat dimension, w cred is the credibility dimension, w diff is the difficulty dimension), calculate the intention score S of each document in the result under the current query intent .f diff It is the difficulty weight processing function. For designs where the difficulty weight can be negative, the following processing method is adopted:

[0118]

[0119] This weight function makes the calculation process refer only to the absolute value of the weight. The positive and negative values ​​only represent their actual meaning and will not affect the calculation process, thus ensuring the credibility of the weighted algorithm.

[0120] At the same time, timeliness is a key feature of ICT learning resources, especially for fields with rapid technological iteration. The timeliness query detection method checks whether the query text contains timeliness-sensitive keywords (such as latest, recent, new technology, 2025, etc.). These keywords are stored as a list and can be expanded according to the characteristics of the query field. The time decay calculation method calculates the timeliness score based on the resource release time. This system includes two different sets of decay curves, one for timeliness-sensitive queries and the other for non-sensitive queries. For timeliness-sensitive queries (such as "the latest AI technology in 2024"), the system uses a steeper decay curve, otherwise it uses a gentler decay method.

[0121]

[0122] Where Δt is the number of days since the resource was released, and y is the number of years T is the final calculated timeliness score. The final timeliness score is limited to the range of [0.1, 1.0] to avoid completely excluding valuable classic resources due to excessive time.

[0123] The final resource ranking score is obtained by fusing the intent weighted score with the timeliness compensation. The calculation process receives the similarity score, timeliness score, resource feature score (popularity, credibility, difficulty) and intent type as input, and outputs a comprehensive floating-point score. First, the corresponding weight configuration is obtained based on the intent type, and then the weighted sum of each indicator is calculated. The function then adjusts the timeliness weight based on whether the query is time-sensitive (0.2 for sensitive queries and 0.05 for non-sensitive queries), and ensures that the sum of the weights of each component is 1 through proportional scaling, and finally generates a comprehensive score in the range of 0-1.

[0124]

[0125] Among them, w i and w t The dynamic weights of intent and timeliness are adjusted according to the query characteristics. intent The weight of the timeliness score T(d) is used to obtain the comprehensive score S for re-ranking. full .

[0126] Furthermore, based on the above method, the present invention designed an evaluation system, constructing 700 real user queries in the ICT field, covering a variety of learning intent scenarios. Each query was annotated with metadata such as keywords, intent type, and sentence type. To evaluate the effectiveness of different retrieval strategies, six experimental modes were used, including baseline comparison experiments and ablation experiments. This evaluation system supports flexible pluggability of modules, flexible switching of experimental modes, and switching of the top-n search phase length.

[0127] The evaluation uses the following three core indicators: (1) Timeliness score: Quantify the timeliness of the search results through weighted calculation. The formula below shows the detailed calculation method, based on the timeliness score of each data in the recalled resources and their relative order. (2) Large model score: Use two large language models, Qwen-2.5-Max and DeepSeek-v3, to evaluate the overall satisfaction of the search results based on (question, search results). During the evaluation process, the system designed a structured prompt template to guide the large model to quantify the scores from three dimensions: keyword matching, content relevance, and timeliness. The average of the two model scores is taken to reduce the bias of the single model evaluation. (3) Intent satisfaction: Use the large model to evaluate whether the search results can meet the original intention of the user's query based on the three contents of (question, intent, search results). This indicator pays special attention to the degree to which the search results meet the user's learning intention, and different evaluation criteria are designed for each intent type.

[0128]

[0129] Where i represents the order of the result in the query result, k represents the top-k mode of the current query, and r i represents the i-th recall result, the timeliness function is the timeliness score calculation function defined previously, R represents the entire content of this answer, and after calculation, the comprehensive timeliness score T of this answer is obtained. score .

[0130] Figure 9 This is a case study of the present invention.

[0131] The embodiment of the present invention uses a dynamic intent recognition semantic and time-enhanced recommendation method for ICT learning resources to build a high-quality, multi-dimensionally annotated ICT domain-specific knowledge base (ICTKB): This knowledge base construction process uses a strategy that combines a large language model (LLM) with domain knowledge for keyword expansion and data collection planning. Through browser-based automated crawler technology, multi-source heterogeneous data is efficiently and stably acquired, and a systematic data integration, timeliness filtering, and low-quality content cleaning process is implemented. Finally, a knowledge base containing 60,187 high-quality real learning resources is constructed for retrieval and recall. A multi-dimensional automated evaluation system based on a large language model is constructed: a data set containing real user queries and multi-dimensional annotations is designed, and a large language model is used to automatically and quantitatively evaluate the timeliness, overall satisfaction, and intent satisfaction of the retrieval results, providing an objective, efficient, and reproducible evaluation method for verifying and iterating the method proposed in the present invention.

[0132] The beneficial effects of the present invention are:

[0133] The present invention builds a high-quality ICT knowledge base and integrates the dynamic intent recognition capability of a large language model with a segmented timeliness compensation strategy to intelligently re-rank traditional semantic search results, which can bring the following significant benefits: (1) Significantly improve recommendation accuracy. Through the large model, the real learning intention behind the user query (entry, in-depth, practice, etc.) is deeply understood, and resource recommendation preferences are dynamically adjusted accordingly, so that the recommendation results are highly matched with the user's actual needs, overcoming the recommendation bias problem caused by traditional methods based only on literal matching or fuzzy interest inference. (2) Effectively enhance the timeliness and applicability of results. By distinguishing the timeliness sensitivity of the query and the resource type, a differentiated timeliness processing strategy is adopted to ensure that the latest cutting-edge technology information can be pushed in a timely manner, while not excluding valuable basic knowledge and classic tutorials, meeting the characteristic requirements of rapid iteration and structuring of ICT knowledge. (3) Improve user learning experience. The resources provided by the present invention are not only semantically relevant, but also meet the user's specific learning goals and stages, and take timeliness into account, reducing the time cost and cognitive burden of users screening in massive information, and improving the efficiency and effectiveness of ICT learning.

[0134] In order to implement the above embodiment, Figure 10 As shown, this embodiment also provides an ICT learning resource dynamic intent recognition semantic and time-enhanced recommendation system 10, including:

[0135] ICT knowledge base construction module 100, used to construct an ICT knowledge base containing multi-source heterogeneous ICT knowledge;

[0136] Semantic retrieval module 200, for performing preliminary recall in the ICT knowledge base based on user demand information to generate a list of semantically related resources;

[0137] Dynamic intention recognition module 300, used to use the large model to analyze user demand information to determine the user's learning intention, and assign an intention weight to each resource in the resource list based on the determined learning intention;

[0138] Timeliness processing module 400, used to detect timeliness sensitive words in user demand information and calculate the timeliness score of each resource in the resource list;

[0139] Enhanced re-ranking module 500, for calculating a comprehensive score for each resource in the resource list based on the generated semantic similarity, intention weight, timeliness score, and resource characteristics, and re-ranking the resource list according to the comprehensive score;

[0140] The result recommendation module is used to recommend the resource list that best meets the user's intention and timeliness requirements as the search results to the user based on the re-ranking results.

[0141] The ICT learning resource dynamic intent recognition semantic and time-enhanced recommendation system of the embodiment of the present invention constructs a high-quality, multi-dimensionally annotated ICT domain-specific knowledge base (ICTKB): This knowledge base construction process adopts a strategy that combines a large language model (LLM) with domain knowledge for keyword expansion and data collection planning. Through browser-based automated crawler technology, multi-source heterogeneous data is efficiently and stably acquired, and a systematic data integration, timeliness filtering and low-quality content cleaning process is implemented. Finally, a knowledge base containing 60,187 high-quality real learning resources is constructed for retrieval and recall. A multi-dimensional automated evaluation system based on a large language model is constructed: a data set containing real user queries and multi-dimensional annotations is designed, and a large language model is used to perform automated and quantitative evaluations of the timeliness, overall satisfaction and intent satisfaction of the retrieval results, providing an objective, efficient and reproducible evaluation method for verifying and iterating the method proposed in the present invention.

[0142] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0143] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A dynamic intent recognition semantic and time-enhanced recommendation method for ICT learning resources, characterized by: include: Build an ICT knowledge base containing multi-source heterogeneous ICT knowledge; Conduct a preliminary search in the ICT knowledge base based on user demand information to generate a list of semantically related resources; Use the big model to analyze user demand information to determine the user's learning intention, and assign an intention weight to each resource in the resource list based on the determined learning intention; Detect timeliness-sensitive words in user demand information and calculate the timeliness score of each resource in the resource list; Calculate the comprehensive score of each resource in the resource list based on the generated semantic similarity, intent weight, timeliness score, and resource characteristics, and re-sort the resource list according to the comprehensive score; Based on the re-ranking results, the resource list that best meets the user's intention and timeliness requirements is recommended to the user as the search results.

2. The method according to claim 1, characterized in that Build an ICT knowledge base containing multi-source heterogeneous ICT knowledge, including: We use crawler technology to extract high-frequency domain terms from authoritative academic resources as seed keywords, establish a multi-dimensional classification system based on domain characteristics, and then use a large language model to perform two-stage semantic expansion to output a keyword set. Based on the keyword set, browser automation technology is used to simulate complete user behavior. By executing native JavaScript expressions in the browser environment, encrypted parameters dynamically generated by the website are directly obtained. The collection strategy is customized according to the characteristics of different platforms to collect multi-platform data. Define a unified ICT knowledge representation data structure and map multi-platform data to standardized fields. Then, deduplicate data based on unique identifiers. After deduplication, mark or degrade resources older than a preset age, and directly remove outdated content. At the same time, apply a regular expression rule library combined with text features to identify and filter low-quality content to obtain standardized data. A dual-channel hybrid scoring model including the three dimensions of popularity, credibility, and difficulty is constructed to perform multi-dimensional scoring on each resource of standardized data to obtain a three-dimensional scoring result. Multi-dimensional labels are automatically generated through a knowledge-guided prompt project based on the six-dimensional label system, resource content, and three-dimensional scoring results.

3. The method according to claim 2, characterized in that Using a large language model, a two-stage semantic expansion is performed to output a keyword set, including: Classify and verify the seed keywords to obtain classification verification results; Based on the classification verification results and structured hint engineering, multiple rounds of knowledge expansion are performed to generate candidate keywords; The generated candidate keywords are deduplicated and category balanced to output a keyword set.

4. The method according to claim 3, characterized in that Based on user demand information, a preliminary search is performed in the ICT knowledge base to generate a list of semantically related resources, including: Based on the weighted segmentation coding strategy, the standardized data resources are decomposed into three parts: core content, label information and detailed content, and different weights are assigned to the three parts; Each part is vectorized separately, and the vector representations of the three parts are synthesized according to the weights to obtain the final resource vector representation.

5. The method according to claim 4, characterized in that The big model is used to analyze user demand information to determine the user's learning intention, and then assign an intention weight to each resource in the resource list based on the determined learning intention, including: Seven ICT learning intentions were defined, and the Qwen-2.5-Max model was used as the intent identifier. Specific prompt words were used to guide the model in determining the intent category to which the user's desired information belongs. Each intent is assigned a specific weight configuration, and a preset formula is used to calculate the total score based on the intent when the intent weight is actually calculated.

6. The method according to claim 5, characterized in that Detect timeliness-sensitive words in user demand information and calculate the timeliness score of each resource in the resource list, including: Check whether the query text contains time-sensitive keywords and store the time-sensitive keywords as a keyword list; Based on the keyword list, the timeliness score is calculated according to the resource release time using the time decay calculation method; it includes two different sets of decay curves for timeliness-sensitive queries and non-sensitive queries respectively.

7. The method according to claim 6, characterized in that The comprehensive score of each resource in the resource list is calculated based on the generated semantic similarity, intent weight, timeliness score, and the resource's own characteristics, including: Obtain multiple metrics, including similarity score, timeliness score, resource feature score, and intent type; Obtain the corresponding weight configuration based on the intent type, and then calculate the weighted sum of each indicator to obtain the weighted sum calculation result; Based on the weighted sum calculation results and according to whether the query is time-sensitive, the timeliness weight is adjusted, and proportional scaling is used to ensure that the sum of the weights of each component is 1, and finally a comprehensive score in the range of 0-1 is generated.

8. The method according to claim 7, characterized in that The method further comprises: Timeliness score: quantifies the timeliness of search results through weighted calculation; Large model scoring: Two large language models, Qwen-2.5-Max and DeepSeek-v3, are used to evaluate the overall satisfaction of retrieval results based on user demand information and retrieval results; Intent satisfaction: Use a large model to evaluate whether the retrieval results can meet the original intention of the user's demand information based on user demand information, intent, and retrieval results.

9. A dynamic intention recognition semantic and time-enhanced recommendation system for ICT learning resources, characterized by: include: ICT knowledge base construction module, used to build an ICT knowledge base containing multi-source heterogeneous ICT knowledge; Semantic retrieval module, used to perform preliminary recall in the ICT knowledge base based on user demand information to generate a list of semantically related resources; Dynamic intent recognition module, which uses a large model to analyze user demand information to determine the user's learning intention and assigns an intention weight to each resource in the resource list based on the determined learning intention; Timeliness processing module, used to detect timeliness-sensitive words in user demand information and calculate the timeliness score of each resource in the resource list; An enhanced re-ranking module is used to calculate the comprehensive score of each resource in the resource list based on the generated semantic similarity, intent weight, timeliness score, and resource characteristics, and re-rank the resource list according to the comprehensive score; The result recommendation module is used to recommend the resource list that best meets the user's intention and timeliness requirements as the search results to the user based on the re-ranking results.

Images