Knowledge service system construction method

Abstract

The invention relates to the technical field of knowledge base construction, in particular to a knowledge service system construction method, which comprises a knowledge acquisition step for collecting knowledge data from knowledge sources of multiple fields; the preprocessing step is used for preprocessing the knowledge data; the classification step is used for classifying the obtained knowledge data; the similarity calculation step is used for calculating the similarity between different categories of knowledge bases; the first judgment step is used for judging whether the similarity is greater than a first threshold value or not; the association step is used for establishing association between the knowledge according to the similarity; the second judgment step is used for judging whether the similarity is greater than a second threshold value or not; the labeling step is used for constructing a new upper label; a label directory generation step for establishing a label directory; a query demand acquisition step for acquiring a knowledge retrieval demand of a user and generating a query request; the knowledge retrieval step is used for conducting knowledge recommendation according to the query request. The method has the effect of improving the searching efficiency.

Description

Technical Field

[0001] The present application relates to the technical field of knowledge base construction, and in particular to a method for constructing a knowledge service system. Background Art

[0002] At present, with the rapid development of information technology and the rise of the knowledge economy, knowledge services have become an important force to promote innovation and development in all walks of life. With the rapid development of technologies such as big data and artificial intelligence, the generation, dissemination and application of knowledge data have accelerated unprecedentedly, and the sources of knowledge have become more diversified, covering science, technology, economy, electricity, culture and other fields. Knowledge data is growing explosively, and the cross-integration of knowledge in multiple fields is becoming increasingly frequent.

[0003] Traditional knowledge management systems are mostly based on a single field or simple classification standards, and organize knowledge in the form of a tree structure or a flat list. Although this system facilitates the storage and search of knowledge to a certain extent, it has obvious limitations. First, the classification of knowledge is often too broad, resulting in large differences in knowledge content under the same category, causing users to have too much redundant information when searching, making it difficult to quickly find comprehensive and accurate information, resulting in poor search efficiency. Summary of the invention

[0004] In order to improve search efficiency, this application provides a method for constructing a knowledge service system.

[0005] This application provides a method for constructing a knowledge service system, which adopts the following technical solutions: A method for constructing a knowledge service system comprises the following steps: Knowledge acquisition: Collecting knowledge data from knowledge sources in multiple fields; Preprocessing: preprocess the knowledge data to eliminate noise; Classification: Classify the acquired knowledge data, generate corresponding primary labels for each category, and store the data of the classification number separately to form a knowledge base; Similarity calculation: calculate the similarity between knowledge bases of different categories; First judgment: judging whether the similarity is greater than a first threshold, if so, executing the association step; Association: Establish associations between knowledge based on similarities; Second judgment: judging whether the similarity is greater than a second threshold, if so, executing the labeling step; Labeling: Analyze their common features and construct new hyperlinks to summarize and reflect the common attributes of this knowledge; Tag catalog generation: Create a tag catalog based on knowledge base categories, primary tags, and upper-level tags; Query demand acquisition: obtain the user's knowledge retrieval needs, determine keywords or tags based on the retrieval needs, and generate query requests; Knowledge retrieval: Based on the query request, combined with the upper-level tags and knowledge associations, the results are filtered and sorted to make knowledge recommendations.

[0006] By adopting the above technical solution, primary tags are generated for each category through classification, and a knowledge base is formed, which helps users locate the required knowledge more quickly; the technical solution introduces a similarity calculation step, and by calculating the similarity between knowledge bases of different categories, potential knowledge associations can be discovered; through the first and second judgment steps, the technical solution can further analyze the common features between knowledge and construct new superordinate tags to summarize and reflect the common attributes of these knowledge. This helps users understand the knowledge content more deeply and simplifies the retrieval process. In addition, the generation of the tag directory also improves the navigability and ease of use of the knowledge system.

[0007] Combined with hyperlinks and knowledge associations, this technical solution can more accurately filter and sort search results and make knowledge recommendations, which provides a more personalized and intelligent user experience compared to the simple keyword matching or content-based search methods in the prior art. In summary, through classification, similarity calculation and association steps, this technical solution can build a more complete and coherent knowledge system, which helps users understand the knowledge content more comprehensively; through the combination of upper-level tags and knowledge associations, this technical solution can more accurately filter and sort the search results, making it easier for users to quickly find the knowledge they want to know, reducing the steps of self-screening from redundant knowledge, and improving retrieval efficiency and accuracy.

[0008] The generation of tag catalogs and intelligent knowledge recommendation functions enable users to locate required knowledge more quickly and obtain a personalized learning experience; this technical solution helps promote knowledge sharing and innovation by building a systematic knowledge service system, providing strong support for academic research and business development.

[0009] Optionally, the classification step also includes: performing multi-dimensional classification based on the subject, content, and form characteristics of the knowledge, and assigning one or more labels to the knowledge under each dimension.

[0010] By adopting the above technical solutions, multi-dimensional classification can capture the characteristics of knowledge more comprehensively, taking into account not only the subject and content of knowledge, but also its formal characteristics (such as document type, media format, etc.); this classification method is more accurate and detailed than single-dimensional classification, which helps users find the required knowledge more accurately.

[0011] Assigning one or more labels to the knowledge under each dimension not only increases the number and types of labels, but also enhances the correlation between labels. This label assignment method enables each knowledge point in the knowledge system to be connected to each other through multiple labels, forming a closer and more complex knowledge network.

[0012] Multi-dimensional classification and a rich tag system provide users with more search entries and paths. Users can choose different dimensions and tags for combined queries according to their needs and preferences, so as to obtain the required knowledge more flexibly. At the same time, this retrieval method also improves the user's retrieval efficiency, allowing users to find relevant information faster.

[0013] Multidimensional classification and labeling systems help to discover knowledge connections and intersections between different fields and topics. This connection and intersection provides the possibility for cross-integration and innovation of knowledge, and helps stimulate new ideas, viewpoints and methods.

[0014] By classifying knowledge in multiple dimensions and assigning labels, the knowledge system can be managed and organized more effectively; this management method not only improves the accessibility and comprehensibility of knowledge, but also contributes to the long-term preservation and inheritance of knowledge.

[0015] Optionally, the labeling step also includes: analyzing the commonalities and differences of similar knowledge bases, as well as their positions and functions in the knowledge system, to determine the naming and hierarchical structure of the superordinate labels.

[0016] By adopting the above technical solution and deeply analyzing the similarities and differences of similar knowledge bases, we can more accurately grasp the internal connections and essential characteristics between them. This analysis helps to determine the naming of superordinate tags and ensure that the tags can accurately reflect the common attributes of relevant knowledge. At the same time, considering the position and role of knowledge in the knowledge system, we can further ensure that the hierarchical structure and logical relationship of the superordinate tags are reasonable, thereby enhancing the representativeness of the tags.

[0017] This detailed operation in the labeling step helps to clarify the hierarchical and logical relationships between knowledge. By determining the hierarchical structure of the upper-level labels, the subordinate relationships and association paths between knowledge can be clearly displayed. This optimization of the hierarchical structure and logical relationships makes the knowledge system more organized and easy to understand, which helps users better grasp the overall structure and internal logic of the knowledge system.

[0018] The accurate and hierarchical super-tag system provides users with more search paths and choices. Users can select super-tags of different levels for combined queries as needed, thereby locating the required knowledge more quickly. At the same time, since super-tags can accurately reflect the common attributes of related knowledge, the accuracy and relevance of search results can be improved.

[0019] By analyzing the similarities and differences of similar knowledge bases, as well as their positions and roles in the knowledge system, we can discover the knowledge connections and intersections between different fields and topics. This connection and intersection makes it possible to deeply explore and integrate knowledge, and helps to form a more comprehensive and in-depth knowledge system.

[0020] The optimized upper-level tag system provides strong support for the intelligence and personalization of knowledge services; based on users' historical search records, interest preferences and other information, it can intelligently recommend upper-level tags and knowledge system paths related to user needs, thereby providing more accurate and personalized knowledge services.

[0021] Optionally, after the query requirement acquisition step, a tag delineation step is also provided; Tag delineation: Based on the query request, preliminary knowledge base categories and upper-level tags are recommended for the queryer to choose from, and the query request is enriched.

[0022] By adopting the above technical solution, the tag delineation step provides a clearer query direction for the inquirer by preliminarily recommending knowledge base categories and upper-level tags. The inquirer can select the tags and categories most relevant to his or her needs based on the recommendations, thereby quickly narrowing the query scope and improving query efficiency. At the same time, since the recommendation is based on intelligent analysis of the query request, it can also improve the accuracy of the query.

[0023] The tag circling step allows the inquirer to select tags and categories according to their needs. This interactive query method enhances the user's sense of participation and satisfaction; users can make personalized choices based on their interests and preferences, thereby obtaining knowledge recommendations that better meet their needs.

[0024] Through the tagging step, the inquirer can further enrich and improve their query requests; they can add or modify query keywords based on the recommended tags and categories to more accurately express their needs; this optimized query request can enhance the matching ability of the retrieval system and improve the quality and relevance of the retrieval results.

[0025] The tagging step not only helps the inquirer to quickly find the required knowledge, but also guides them to discover more knowledge in related fields; by recommending relevant knowledge base categories and superordinate tags, the inquirer can expand his knowledge horizon and discover more potentially valuable knowledge resources.

[0026] By intelligently analyzing query requests and recommending relevant knowledge base categories and upper-level tags, the system can provide users with more intelligent and personalized knowledge services; this intelligent service not only improves query efficiency, but can also be continuously optimized and improved based on user feedback to enhance overall service quality.

[0027] Optionally, the knowledge retrieval step also includes: dynamically adjusting the weights of upper-level tags and knowledge associations according to the user's retrieval history, preferences and behavior patterns, and optimizing the knowledge recommendation ranking.

[0028] By adopting the above technical solution, by analyzing the user's search history, preferences and behavior patterns, the system can more accurately understand the user's real needs; on this basis, dynamically adjust the weights of upper-level tags and knowledge associations to make the recommended knowledge more in line with the user's personalized needs; this personalized recommendation method can significantly improve user satisfaction and loyalty.

[0029] The system dynamically adjusts the weight of knowledge associations based on user preferences and behavior patterns, so that knowledge most relevant to user interests is displayed first; this helps reduce invalid clicks by users and improves retrieval efficiency; users can find the knowledge they need in a shorter time, thereby improving the overall user experience.

[0030] By dynamically adjusting the weights of upper-level tags, the system can more accurately evaluate the relevance and importance of knowledge; on this basis, the knowledge is sorted so that the most relevant and important knowledge to user needs is placed at the front; this helps to improve the accuracy and quality of knowledge recommendations, allowing users to more easily obtain valuable information.

[0031] Personalized recommendations and optimized knowledge sorting can enhance user stickiness and make users more willing to stay and explore in the system; this helps promote the sharing and dissemination of knowledge and form a more active and rich knowledge community.

[0032] Through continuous learning and optimization, the system can more accurately grasp user preferences and behavior patterns, thereby providing more intelligent knowledge services; this improvement in the level of intelligence helps to enhance the system's competitiveness and market position.

[0033] As user needs and interests change, the system can dynamically adjust the weights of upper-level tags and knowledge associations to ensure that the recommended knowledge is always consistent with user needs; this helps promote the updating and maintenance of knowledge and maintain the timeliness and accuracy of the knowledge system.

[0034] Optionally, a deviation obtaining step is further provided after the query requirement obtaining step; Deviation acquisition: including demand similarity calculation step, third judgment step, difference statistics step, difference acquisition step, difference judgment step, correction step, fourth judgment step, supplementary calculation step, supplementary judgment step and supplementary labeling step; Demand similarity calculation: Calculate multiple query requests from the same query requester and calculate the similarity between the previous and next query requests; Third judgment: judging whether the similarity is greater than a preset third threshold, if so, executing the difference statistics step, otherwise, executing the knowledge retrieval step; Difference statistics: counting the number of similar differences and judging whether it is greater than the fourth threshold, if so, executing the difference judgment step; Difference acquisition: extract the words that have been modified multiple times and record them as difference words, obtain the common points between multiple difference words and record them as content words, calculate the correlation between content words and superordinate labels, and then perform the difference judgment step; Difference judgment: judging whether the correlation degree is less than the preset fifth threshold value, if so, executing the correction step, otherwise, executing the fourth judgment step; Correction: Intervene in the labeling step and modify and update the upper label; Fourth judgment: judging whether the number of times of difference similarity is greater than the sixth threshold value, if so, executing the supplementary calculation step; Supplementary calculation: Calculate the similarity between the hyperlinked tags and record it as synonymous similarity; Supplementary judgment: judge whether the synonymous similarity is greater than the seventh threshold, if so, perform the supplementary labeling step; Supplementary tags: Set multiple upper-level tags in a pop-up window for the query requester to make additional selections and correct the query request.

[0035] By adopting the above technical solution, the demand similarity calculation step can identify and compare multiple query requests of the same query requester, and by calculating the similarity of the previous and subsequent query requests, it is helpful to discover slight changes or continuous focus points in the query requests; when the similarity is higher than the preset threshold, the difference statistics step can further analyze the differences between the query requests, especially when the number of difference similarities exceeds the threshold, it indicates that the query requester may have encountered some query obstacles or the demand is not fully met; The difference acquisition step can intelligently identify the true intention of the query requester by extracting words that have been modified multiple times (difference words) and finding the common points between these words (content words); the difference judgment step can determine whether the current label accurately reflects the needs of the query requester by calculating the correlation between the content words and the upper-level labels; if the correlation is lower than the threshold, the labeling step is intervened through the correction step to ensure the accuracy and effectiveness of the label; when the number of difference similarities exceeds another threshold, the supplementary calculation step and the supplementary judgment step can identify the similarity between the upper-level labels, and then find possible synonymous labels or label redundancy; The supplementary label step not only enhances the interactivity of the query by providing a pop-up window with multiple upper-level labels for the query requester to choose again, but also provides an opportunity to correct the query request, thereby improving the satisfaction with the query results; this technical solution can flexibly respond to changes in different types of query needs by introducing multiple judgment steps and threshold settings, avoiding the limitations that may be brought about by a single query path; by dynamically adjusting the upper-level labels and label correlations during the query process, the query process is optimized and the query efficiency and accuracy are improved.

[0036] The correction step and the label supplement step not only help solve the problem of the current query request, but also promote the updating and maintenance of the knowledge system; by continuously adjusting and optimizing the upper-level labels, it can ensure that the knowledge system keeps pace with the needs of the query requester.

[0037] Optionally, a user participation step, a manual correction step, an associated label matching step, a feedback statistics step and an update step are further provided between the difference determination step and the correction step; User participation: recommending the content words to the query requester, judging whether the content words meet the query requester's needs, if yes, executing the correction step and the associated tag matching step; otherwise, executing the manual correction step; Manual correction: The query requester corrects the content words and then performs the associated tag matching step; Related tag matching: Match relevant hypernyms through lexical words and push them to the query requester. According to the query requester's feedback, perform feedback statistics. Feedback statistics: obtain the number of times a certain content word is corrected in relation to its hypernym, and determine whether it is greater than the eighth threshold. If so, perform the update step; Update: Update the association between the hypernym and the content word for matching subsequent query requests.

[0038] By adopting the above technical solution, the user participation step significantly enhances user participation and interactivity by directly recommending lexical words to the query requester and allowing him to judge whether these words meet the needs, which helps to ensure that the query results are closer to the user's true intentions.

[0039] When the user confirms that the lexical content meets the needs, the related superordinate tags are further pushed through the associated tag matching step, which not only improves the accuracy of the query but also enhances the user's satisfaction; the manual correction step allows the query requester to correct the lexical content, which provides the user with an opportunity to directly intervene in the query process, thereby helping to correct possible misunderstandings or deviations and improve the accuracy of the query results; the associated tag matching step further refines the query results by matching the lexical content with the superordinate tags, making the results more accurate and targeted.

[0040] The feedback statistics step collects the number of times users correct the association between content words and superordinate labels, and can identify which label associations may have problems or need improvement; the update step updates the association between superordinate labels and content words based on the results of feedback statistics, which helps to optimize the knowledge system and make it more in line with users' query needs and habits.

[0041] By introducing user feedback and manual correction mechanisms, it helps to improve the intelligence and adaptability of the system, enabling it to better respond to future query needs; steps such as user participation, manual correction, and feedback statistics enhance the interaction between users and the system, allowing users to feel the system's attention and respect for their needs; this helps to establish a trusting relationship between users and the system and improve user loyalty and satisfaction with the system.

[0042] Optionally, after the association tag matching step, a demand determination step and a single association step are further provided; Demand judgment: judge whether the generated tags meet the needs of the query requester. If so, execute the feedback statistics step; otherwise, execute the single association step; Single association: The query requester enters a custom tag, matches it from the upper-level tags, and establishes a single association for the query requester's query request, and then executes the knowledge retrieval step.

[0043] By adopting the above technical solution, the demand judgment step provides an additional verification link by judging whether the generated tags meet the needs of the query requester; this helps to ensure that the tags recommended by the system are highly matched with the actual needs of the user, thereby improving the accuracy and relevance of the query results.

[0044] When the system-generated tags do not meet user needs, a single association step allows users to enter custom tags and match them from upper-level tags. This provides users with a way to directly express their needs, further enhancing the precise satisfaction of user needs.

[0045] The single association step allows users to enter custom tags by themselves, which significantly improves user participation. Users can flexibly select or create tags based on their own understanding and needs, thereby more accurately describing their query intentions. By establishing a single association, the system can respond to the specific needs of the query requester and provide a more personalized and customized query service.

[0046] The introduction of the demand judgment step and the single association step makes the query process more efficient and flexible. When users find that the labels generated by the system do not meet their needs, they can immediately correct them through the single association step without having to repeat the entire query process. This not only improves query efficiency, but also reduces user waiting time and frustration, thereby improving user experience.

[0047] By executing the knowledge retrieval step after the single association step, the system is able to perform more accurate knowledge retrieval based on user-defined tags, which helps to dig deeper information and associations and provide users with more comprehensive and in-depth query results.

[0048] Through the tag association established in the single association step, the system can also better understand and process the user's query intention, thereby providing more accurate and relevant knowledge retrieval results; the introduction of the demand judgment step and the single association step enables the system to better adapt to the query needs and habits of different users, which helps to improve the system's adaptability and scalability, enabling it to cope with more complex and diverse query scenarios; at the same time, by allowing users to enter custom tags themselves, the system can also continuously learn and accumulate new tags and related information, thereby continuously optimizing its query algorithm and knowledge system.

[0049] Optionally, after the single association step, an association statistics step is further provided; Association statistics: Count the number of single associations established with the same upper-level tag, and determine whether it is greater than the ninth threshold. If so, perform the tag conversion step; Label conversion: Convert custom labels into higher-level labels and associate them with the knowledge base.

[0050] By adopting the above technical solution, the association statistics step can identify which upper-level tags frequently appear in the user query process and establish a single association with them by counting the number of times a single association is established with the same upper-level tag; this helps the system better understand the user's query needs and habits, thereby optimizing tag management and the knowledge system.

[0051] When the number of single associations of a certain upper-level tag exceeds the set ninth threshold, the custom tag is converted into an upper-level tag through the tag conversion step and associated with the knowledge base; this can not only reduce the redundancy and confusion of tags, but also improve the accuracy and authority of tags, making the knowledge system clearer and more orderly.

[0052] The tag conversion step converts frequently used custom tags into hyperlinked tags and associates them with the knowledge base, which helps the system to more quickly identify and match relevant tags in future queries, thereby improving query efficiency. At the same time, since superordinate tags usually have broader meanings and higher generalization, using superordinate tags for queries can reduce ambiguity and misunderstanding and improve the accuracy of queries; the introduction of association statistics steps and label conversion steps enables the system to adaptively adjust and optimize according to user query behaviors and needs; this helps to improve the system's adaptability and learning capabilities, enabling it to better respond to query needs that may arise in the future.

[0053] By continuously learning and accumulating new tags and related information, the system can continuously improve its tag system and knowledge system, thereby providing more intelligent and efficient query services; the introduction of the associated statistical step and the tag conversion step enables the system to more accurately understand the user's query intention and needs, thereby providing more personalized and customized query results; this helps to improve user experience and satisfaction, and enhance user trust and loyalty to the system Optionally, after the knowledge retrieval step, a recommendation statistics step is also provided; Recommendation statistics: obtain the click-through rate of the knowledge recommended by the query requester, determine the degree of matching, and correct the ranking of subsequent similar query requirements.

[0054] By adopting the above technical solution, the recommendation statistics step can more accurately understand the user's preference and satisfaction with the recommended content by collecting the user's click-through rate data on the recommended knowledge; by analyzing the click-through rate data, the system can evaluate the matching degree and effect of different recommended content. This helps the system to continuously optimize its recommendation algorithm and improve the accuracy and relevance of recommendations. For example, the system can adjust the weight, sorting or display method of the recommended content based on the click-through rate data to better meet user needs.

[0055] The recommendation statistics step enables the system to have a deeper understanding of the user's personalized needs. By collecting and analyzing the user's click behavior, the system can identify the user's interests, preferences and habits, thereby providing more personalized recommendation services, which helps to enhance user satisfaction and loyalty.

[0056] For subsequent similar query needs, the system can correct the sorting based on the results of the recommendation statistics step, which means that the system can find the content that best meets user needs more quickly, reduce user browsing and screening time, and improve query efficiency.

[0057] The recommendation statistics step not only focuses on the satisfaction of current query needs, but also guides future knowledge updates and iterations by collecting and analyzing data. When the system finds that some recommended content has a low click-through rate, it can consider updating or replacing these contents to better adapt to changes in user needs. By introducing the recommendation statistics step, the system can continuously learn and adapt to user query behaviors and needs. This helps to improve the intelligence of the system, enabling it to understand user intent more accurately and provide more intelligent and efficient query services.

[0058] In summary, the present application includes at least one of the following beneficial technical effects: 1. Through classification, similarity calculation and association steps, this technical solution can build a more complete and coherent knowledge system, which helps users understand the knowledge content more comprehensively; through the combination of hyperlinks and knowledge associations, this technical solution can more accurately filter and sort search results, making it easier for users to quickly find the knowledge they want to know, reducing the steps of self-screening from redundant knowledge, and improving search efficiency and accuracy; 2. By introducing multiple judgment steps and threshold settings, it can flexibly respond to changes in query requirements of different types and avoid the limitations that may be brought about by a single query path; by dynamically adjusting the upper-level tags and tag associations during the query process, the query process is optimized and the query efficiency and accuracy are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the method for constructing a knowledge service system in Example 1 of the present application; Figure 2 is a position relationship flow chart of the deviation acquisition step in Example 2 of the present application; Figure 3 is a sub-step flow chart of the deviation acquisition step in Example 2 of the present application; Figure 4 It is a flowchart from the user participation step to the update step in Example 2 of the present application; Figure 5 It is a flowchart from the demand judgment step to the single association step in Example 2 of the present application. DETAILED DESCRIPTION

[0060] The following combination Figures 1 to 5 This application is described in further detail.

[0061] This embodiment discloses a method for constructing a knowledge service system.

[0062] Example 1: Reference Figure 1 ,The knowledge service system construction method includes the following steps: Knowledge acquisition: Collecting knowledge data from knowledge sources in multiple fields; Specifically, select authoritative knowledge sources covering multiple fields, such as academic paper databases, professional books, online courses, industry reports, etc.; use crawler technology, API interfaces or manual entry to collect knowledge data in batches from the above knowledge sources, including text, pictures, videos and other forms.

[0063] Preprocessing: preprocess the knowledge data to eliminate noise; Specifically, preprocessing includes data cleaning, text processing, and data format unification; data cleaning includes removing duplicate data, invalid data (such as blanks, garbled characters), and advertising information, and cleaning the original data set through data cleaning tools; text processing includes natural language processing operations such as word segmentation, stop word removal, and stem extraction, which can be processed using natural language processing libraries (such as NLTK, jieba); data format unification includes converting data in different formats into a unified storage format for subsequent processing, and converting text data into JSON or XML format.

[0064] Classification: Classify the acquired knowledge data, generate corresponding primary labels for each category, and store the classified data separately to form a knowledge base; Specifically, the acquired knowledge is initially classified into different fields, and then a machine learning algorithm is selected for model training. The model can use a naive Bayes classifier. During the training process, a dataset with known classifications (also called a training set) is needed to train the classification model. This dataset contains a large number of knowledge data samples, and each sample has been manually labeled with the correct classification label. By allowing the model to learn the relationship between the features and labels of these samples, the model can gradually learn how to classify new knowledge data.

[0065] Then, personnel with professional knowledge background will conduct manual review and check the results of automatic classification one by one to ensure that each knowledge data is correctly classified into the corresponding category; if classification errors are found, the model needs to be adjusted and optimized to improve the accuracy of classification; after the classification is completed, a descriptive primary label needs to be generated for each classification, and this label should be able to accurately reflect the subject and content of the knowledge data contained in the classification. The process of generating primary labels can be automated or manual. The methods for automatically generating primary labels usually include natural language processing technologies such as text summarization and keyword extraction. These technologies can extract key information from the knowledge data in the classification and generate a concise and clear label to describe the classification.

[0066] For example, in the power industry, the initial areas can be divided into: Power load forecasting: This area focuses on the changing trends and forecasts of power load. Data may include historical load data, weather forecasts, holiday information, economic indicators, etc., which are used to train models to predict future power demand.

[0067] Power market transactions: This area involves transaction data and price information in the power market. The data may include power plant quotations, power transaction volume, power prices, etc., which are used to analyze market dynamics and formulate trading strategies.

[0068] Grid operation and maintenance: This area focuses on the stable operation and fault maintenance of the grid. Data may include grid topology, equipment status monitoring data, fault records, etc., which are used to monitor the grid status, predict faults, and take corresponding maintenance measures.

[0069] Power energy management: This field involves the production, distribution and use of energy. Data may include power plant capacity, energy transmission losses, user electricity behavior, etc., which are used to optimize energy allocation and improve energy efficiency.

[0070] Power safety and emergency response: This area focuses on the safety and emergency response capabilities of the power system. Data may include power grid fault simulation data, emergency response plans, accident records, etc., which are used to assess the safety risks of the power grid and formulate emergency response strategies.

[0071] New Power Technology and Applications: This area focuses on new technologies and innovative applications in the power sector. Data may include smart grid technology, electric vehicle charging station data, distributed energy system data, etc., to explore the application potential and effects of new technologies in power systems.

[0072] Similarity calculation: calculate the similarity between knowledge bases of different categories; First judgment: judging whether the similarity is greater than a first threshold, if so, executing the association step; Association: Establish associations between knowledge based on similarities; Specifically, similarity calculation is a quantitative indicator to measure the similarity between two or more knowledge items. Similarity calculation methods may include cosine similarity, Jaccard similarity, and semantic similarity; Cosine similarity: Knowledge items can be represented as vectors, where each dimension of the vector represents a feature (such as a keyword, topic, etc.). Cosine similarity = (A·B) / (|A|*|B|), where A and B represent the vector representations of two knowledge items, · represents the dot product of the vectors, and |A| and |B| represent the modulus of vectors A and B, respectively. It is suitable for measuring the similarity between knowledge items based on feature vectors. Jaccard similarity: Knowledge items can be represented as sets. The elements in the set represent the characteristics of the knowledge items (such as keywords, concepts, etc.). Jaccard similarity = |A∩B| / |A∪B|, where A and B represent the set representations of two knowledge items respectively, ∩ represents the intersection, and ∪ represents the union. It is suitable for measuring the similarity between knowledge items based on set representation.

[0073] Semantic similarity: Semantic similarity measures the similarity between two knowledge items by calculating the distance between them in the semantic space. This usually requires the use of natural language processing technology and semantic models (such as word embedding, knowledge graphs, etc.) to achieve this. Pre-trained semantic models can be used to calculate the semantic distance between knowledge items, such as cosine distance, Euclidean distance, etc., and then the distance is converted into a similarity score, which is usually implemented using the inverse of the distance or a negative exponential function. It is suitable for measuring the similarity between knowledge items based on semantic understanding.

[0074] Then, the similarity calculation result is compared and screened with the set first threshold to establish the association relationship between the knowledge items. The greater the similarity, the tighter the association. After exceeding the set first threshold, the association between the two is established. This can be achieved by creating an association table, a relationship diagram in the database, or using graph database and other technologies.

[0075] Validate and optimize established associations, which can be achieved through manual review, user feedback, or statistical-based methods.

[0076] The first threshold may be determined based on past project experience or expert opinions, and a fixed similarity threshold may be directly set.

[0077] Or by analyzing a large amount of data, calculating the similarity distribution, and then setting the first threshold according to the distribution characteristics, for example, a percentile in the similarity distribution can be selected as the threshold, such as the 90% or 95% percentile.

[0078] Alternatively, the first threshold may be dynamically adjusted according to the characteristics of the current data set. A machine learning algorithm, such as a support vector machine (SVM) or a decision tree, may be used to learn and predict the optimal threshold according to the data characteristics.

[0079] Second judgment: judging whether the similarity is greater than a second threshold, if so, executing the labeling step; Labeling: Analyze their common features and construct new hyperlinks to summarize and reflect the common attributes of this knowledge; Tag catalog generation: Create a tag catalog based on knowledge base categories, primary tags, and upper-level tags; Query demand acquisition: obtain the user's knowledge retrieval needs, determine keywords or tags based on the retrieval needs, and generate query requests; Knowledge retrieval: Based on the query request, combined with the upper-level tags and knowledge associations, the results are filtered and sorted to make knowledge recommendations.

[0080] Specifically, a second threshold is set to screen out highly similar knowledge items, the common features of these knowledge items are analyzed manually or automatically, and upper-level labels are constructed based on the common features and assigned to relevant knowledge items. The hierarchical structure and logical relationship of the tag directory are designed, and various tags and knowledge items are organized according to the designed structure. A hierarchical tag directory is constructed based on knowledge base categories, primary tags and upper-level tags, and the tag directory is displayed in the form of a graphical interface or web page to facilitate user browsing and query.

[0081] Provide a user input interface for query requesters, receive users' knowledge retrieval needs, generate query requests based on the keywords entered by users or the tags selected, filter and sort the search results based on the upper-level tags and knowledge associations, and generate a recommendation list.

[0082] Among them, the second threshold can be determined by statistical analysis, experimental method and expert consultation method, among which the statistical analysis method: analyze the similarity distribution between knowledge items, observe the concentrated area and discrete degree of similarity, and set the second threshold according to the statistical characteristics of the similarity distribution (such as mean, standard deviation, percentile, etc.); for example, a higher percentile (such as 95% or 99%) can be selected as the second threshold to screen out highly similar knowledge items.

[0083] Experimental method: Test the screening effect under different thresholds through experiments. Manual evaluation or machine learning algorithm can be used to evaluate the accuracy and relevance of the screening results. The second threshold can be adjusted according to the experimental results until a satisfactory screening effect is achieved.

[0084] Expert consultation method: Invite domain experts or practitioners with rich experience to participate in the setting of the second threshold, use the experts' professional knowledge and experience to determine which knowledge items are highly similar, and set the second threshold accordingly.

[0085] For example, when conducting power load forecasting, for the specific field of "power load forecasting", we can collect relevant data from the following knowledge sources: Academic paper database: Search databases such as IEEE Xplore and CNKI to find academic papers on power load forecasting and learn about the latest forecasting models, algorithms, and case studies.

[0086] Professional books: Consult classic works in the fields of power engineering, power system analysis, etc. to understand the basic principles, methods and historical development of power load forecasting.

[0087] Online courses: Learn professional courses related to power load forecasting on platforms such as China University MOOC and NetEase Cloud Classroom, and master skills such as forecasting model construction, parameter optimization, and result evaluation.

[0088] Industry reports: Read the annual reports of power companies such as State Grid and China Southern Power Grid, as well as market analysis reports released by power industry associations and research institutions to understand the current status and trends of power load forecasting in the power industry.

[0089] The collected power load forecast data may contain noise and outliers, and needs to be preprocessed to improve data quality. The specific steps include: Data cleaning: Remove duplicate data, invalid data (such as blanks, garbled characters) and outliers (such as extreme load values) to ensure the accuracy and validity of the data.

[0090] Text processing (if the data contains text information): perform natural language processing operations such as word segmentation and stop word removal on the text data for subsequent analysis.

[0091] Data format unification: Convert data in different formats into a unified storage format, such as JSON or CSV, to facilitate subsequent processing and storage.

[0092] Data normalization: Normalize the load data to eliminate the impact of different dimensions on the data and improve the prediction performance of the model.

[0093] The preprocessed power load forecast data is classified according to different characteristics, such as time scale (daily load, weekly load, monthly load, etc.), geographical scope (city, region, province, etc.), load type (residential load, industrial load, commercial load, etc.). For each category, a corresponding primary label is generated, such as "daily load forecast-urban residents", "monthly load forecast-industrial area", etc.

[0094] Calculate the similarity between different types of power load forecast data and establish associations between knowledge based on the similarity. The specific steps include: Selection of similarity calculation method: Dividing power load subsequences: The two power load curves S and S' to be compared are divided into power load subsequences S(i) and S'(i) of equal length and interval according to the same preset rule, where i=1,2,...,n, and n is the number of power load subsequences.

[0095] The divided power load subsequences S(i) and S'(i) have the same sampling interval, that is, the number of elements m(i) and m'(i) of the power load subsequences S(i) and S'(i) are the same.

[0096] Set the weight of the power load subsequence: The weight w(i) of each power load subsequence S(i) is set according to the formula w(i)=P(i) / P, where P(i) is the sum of the load values ​​of each element in the power load subsequence S(i), and P is the total load value of the power load curve S.

[0097] Calculating the distance between the power load subsequences: using, for example, the Euclidean distance to calculate the distance between the corresponding power load subsequences.

[0098] According to the distance between the power load subsequences and the weight of the power load subsequences, after obtaining the distance between each subsequence, these distances are multiplied by the corresponding weights to obtain the weighted distance; then, all weighted distances are summed or averaged to obtain a total similarity score to calculate the similarity of the two power load curves to be compared.

[0099] Similarity calculation and comparison: Calculate the similarity between different load forecast data and compare it with the set first threshold. If the similarity is greater than the first threshold, establish an association between the two. For example, there may be an association between "daily load forecast - urban residents" and "daily load forecast - urban business".

[0100] Verification and optimization of association relationships: Verify the accuracy and effectiveness of association relationships through manual review or statistical methods, and perform optimization.

[0101] Filter out highly similar power load forecast data, analyze their common features, and construct new upper-level labels. The specific steps include: Second threshold setting: Set an appropriate second threshold based on statistical analysis, experimental method, expert consultation and other methods.

[0102] Superordinate label construction: Filter out data with a similarity greater than the second threshold, analyze their common characteristics (such as time scale, geographical scope, load type, etc.), and construct new superior labels, such as "urban resident load forecast", "industrial area load forecast", etc.

[0103] Tag directory generation: Based on the knowledge base categories, primary tags and upper-level tags, a hierarchical tag directory is established to facilitate user browsing and query.

[0104] According to the user's knowledge retrieval needs, combined with the upper tags and knowledge associations, the power load forecast data is screened and sorted to make knowledge recommendations. The specific steps include: User input interface design: Design a user input interface to receive users' knowledge retrieval needs about power load forecasting, such as entering the keyword "daily load forecast" or selecting the label "urban residential load forecast".

[0105] Query request generation: Generate query requests based on the keywords entered by the user or the tags selected.

[0106] Knowledge retrieval and screening: Combine the upper-level tags and knowledge associations to retrieve and screen the power load forecast data, such as screening out all data on "daily load forecast-urban residents".

[0107] Knowledge recommendation: Generate a recommendation list to show users relevant power load forecast data, upper-level tags, and related knowledge to help users quickly find the knowledge they need.

[0108] The embodiment of the present application refines the upper-level tags and adds filter items to narrow the search scope. The query requester can quickly search according to the upper-level tags, reduce the interference of redundant information, and quickly obtain relevant knowledge.

[0109] Example 2: Reference Figure 2 The difference between this embodiment and embodiment 1 is that the classification step also includes: multi-dimensional classification based on the subject, content, and form characteristics of the knowledge, and assigning one or more labels to the knowledge under each dimension.

[0110] Specifically, for each knowledge item, it is necessary to extract relevant information about its subject, content, and form characteristics. This can be achieved through natural language processing (NLP) technology, image recognition technology, and metadata extraction methods. For example, for text-based knowledge items, NLP technology can be used to extract keywords, topic sentences, paragraph structures, and other information; for image or video-based knowledge items, image recognition technology can be used to extract objects, scenes, colors, and other information; based on the extracted feature information, a classification system is constructed, which includes determining the classification criteria, classification hierarchy, and classification relationship under each dimension. For example, under the subject dimension, a subject field classification tree can be constructed to classify knowledge items hierarchically according to subject fields; under the content dimension, knowledge items can be classified according to classification criteria such as knowledge points, cases, and theories.

[0111] According to the classification system, assign one or more classification labels to each knowledge item, compare the characteristic information of the knowledge item with the classification criteria in the classification system, and find the most suitable classification label. At the same time, you can also consider using multi-label classification technology to assign multiple relevant labels to the knowledge item to more comprehensively describe its attributes and characteristics.

[0112] Before assigning labels, it is necessary to design a reasonable label set and establish labels based on the subject, content, and form characteristics of the knowledge items. At the same time, labels should maintain a certain degree of independence and mutual exclusivity to avoid redundancy and repetition.

[0113] Label allocation can be achieved by combining automatic allocation and manual review. Automatic allocation can be achieved through rule-based matching algorithms, machine learning algorithms, etc., and labels are automatically assigned according to the characteristic information of knowledge items; manual review is to verify and correct the automatic allocation results to ensure the accuracy and rationality of the labels.

[0114] As the knowledge base is continuously updated and expanded, labels also need to be updated and maintained accordingly, including adding new labels, deleting outdated labels, merging duplicate labels, etc. At the same time, the usage of labels also needs to be monitored and analyzed to discover and solve problems in a timely manner.

[0115] For example, data items include: power load data of a certain area on a certain day, including residential load, industrial load and commercial load.

[0116] Feature extraction: Subject features: power load data; Content features: residential load, industrial load, commercial load, load value, load change trend; Form features: table form Classification system: Topic dimension: power load data; Content dimension: classification by user type (residential, industrial, commercial); Label assignment: Topic label: power load data; Content tags: residential load, industrial load, commercial load, load value analysis, load change trend analysis; Tag update and maintenance: As the power data knowledge base is continuously updated and expanded, tags also need to be updated and maintained accordingly. This includes: adding new tags to cover new data types or analysis requirements; deleting outdated tags to avoid redundancy and confusion; merging duplicate tags to keep the tag collection concise and consistent.

[0117] In other embodiments, the labeling step also includes: analyzing the commonalities and differences of similar knowledge bases, as well as their positions and functions in the knowledge system, to determine the naming and hierarchical structure of the superordinate labels.

[0118] Specifically, similarity calculation methods (such as cosine similarity, Jaccard similarity, etc.) are used to calculate the similarity of knowledge items between different knowledge bases. Based on the similarity threshold, highly similar knowledge items are screened out to form a knowledge item set. The knowledge item set is analyzed in terms of subject, content, form, etc. to find out their similarities (such as common themes, similar descriptions, etc.) and differences (such as different emphases, differences in details, etc.), and the analysis results are recorded. Based on the data and analysis results of the existing knowledge base, a knowledge system framework is constructed. The framework should include the hierarchical structure of the knowledge base, the relationship between knowledge items, etc. Each knowledge item is positioned in the knowledge system framework, and its subject, category, and level are clarified. The location information of the knowledge item is recorded and analyzed. The role of the item in the knowledge system, such as whether it is core knowledge, whether it is related to other knowledge items, etc.; according to the results of the role analysis, adjust the knowledge system framework to ensure the accuracy and completeness of the knowledge items; according to the results of the analysis of the similarities and differences of the knowledge items, as well as their position and role in the knowledge system, determine the name of the superordinate label, and according to the knowledge system framework and the positioning information of the knowledge items, use different hierarchical structures such as tree structure and network structure to construct the hierarchical structure of the superordinate label; optimize and adjust the naming and hierarchical structure of the superordinate labels to ensure that they can accurately reflect the characteristics of the knowledge base and the structure of the knowledge system; invite domain experts or experienced practitioners to review and provide feedback to improve the accuracy and rationality of the superordinate labels.

[0119] According to the determined upper-level label naming and hierarchical structure, the knowledge items in the knowledge base are labeled. During the labeling process, attention should be paid to maintaining the consistency and accuracy of the labels. The labeling results should be verified and evaluated to ensure the accuracy and rationality of the labels; based on the verification results, necessary optimization and adjustment should be made to the upper-level labels and hierarchical structure.

[0120] Reference Figure 2 In other implementations, after the query requirement acquisition step, a label delineation step is also provided; Tag delineation: Based on the query request, preliminary knowledge base categories and upper-level tags are recommended for the queryer to choose from, and the query request is enriched.

[0121] Specifically, the system first receives the user's query request and parses the request. The parsing process includes identifying keywords, phrases, and possible contextual information in the query. Through natural language processing technology (such as word segmentation, part-of-speech tagging, named entity recognition, etc.), the system can more accurately understand the user's query intention.

[0122] According to the parsed query request, the system selects categories related to the query content from the knowledge base, and category recommendations are made based on methods such as keyword matching and semantic similarity calculation. For example, if the query request contains the keyword "smart grid", the system can recommend knowledge base categories related to artificial intelligence, such as "grid dispatching" and "distributed energy management".

[0123] While recommending knowledge base categories, the system can also recommend related superordinate tags based on the content of the query request. The recommendation of superordinate tags is based on the hierarchy and association relationship in the tag system. For example, if the query request is about "wind power generation technology", the system can recommend superordinate tags such as "renewable energy" and "power generation technology".

[0124] In order to improve the accuracy of recommendations, the system can also consider the user's historical query records, preference settings and other information. The system will display the recommended knowledge base categories and upper-level tags to the user for selection; the user can select relevant knowledge base categories and upper-level tags according to their needs to further enrich the query request.

[0125] User selections and feedback can also be used to optimize the system's recommendation algorithm and improve the accuracy of subsequent recommendations; based on the knowledge base categories and upper-level tags selected by the user, the system can further enrich the content of the query request; for example, the system can automatically add keywords, phrases, or query conditions related to the selected categories and tags to improve the accuracy and comprehensiveness of the query.

[0126] In other implementations, the knowledge retrieval step also includes: dynamically adjusting the weights of upper-level tags and knowledge associations to optimize the knowledge recommendation ranking according to the user's retrieval history, preferences, and behavior patterns.

[0127] Specifically, the system needs to collect the user's search history, including the query keywords entered by the user, the search results clicked, the browsing time, etc.; the user can set his or her own preferences in the system, such as the topics, fields, and types of knowledge of interest; the system infers the user's behavior patterns and potential needs by analyzing indicators such as the user's click behavior, browsing time, and return rate; the system calculates the correlation between each upper-level tag and the user's query keywords, historical search records, and preference settings; based on the correlation calculation results, the system dynamically adjusts the weight of each upper-level tag; the higher the correlation, the greater the weight of the tag, and thus it occupies a more important position in knowledge recommendation.

[0128] The system calculates the correlation between each knowledge item and the user's query keywords, historical search records, preference settings, and upper-level tags; similarly, based on the correlation calculation results, the system dynamically adjusts the weight of each knowledge item; the higher the correlation, the greater the weight of the knowledge item, and the more likely it is to be recommended to the user.

[0129] The system uses a sorting algorithm (such as PageRank, TF-IDF, etc.) combined with dynamically adjusted upper tags and knowledge association weights to sort knowledge items. Based on the sorting results, the system generates a personalized knowledge recommendation list for users. The knowledge items in the recommendation list are arranged from high to low according to weight, ensuring that users can first see the knowledge items that best meet their needs.

[0130] It is necessary to collect user feedback on the recommendation results, such as satisfaction, comments, etc.; based on user feedback, the system adjusts the weights of upper-level tags and knowledge associations, as well as the parameters of the sorting algorithm in real time to continuously optimize the effect of knowledge recommendation.

[0131] Reference Figure 2 and Figure 3 , in other implementations, a deviation acquisition step is further provided after the query requirement acquisition step; Deviation acquisition: including demand similarity calculation step, third judgment step, difference statistics step, difference acquisition step, difference judgment step, correction step, fourth judgment step, supplementary calculation step, supplementary judgment step and supplementary labeling step; Demand similarity calculation: Calculate multiple query requests from the same query requester and calculate the similarity between the previous and next query requests; Third judgment: judging whether the similarity is greater than a preset third threshold, if so, executing the difference statistics step, otherwise, executing the knowledge retrieval step; Difference statistics: counting the number of similar differences and judging whether it is greater than the fourth threshold, if so, executing the difference judgment step; Difference acquisition: extract the words that have been modified multiple times and record them as difference words, obtain the common points between multiple difference words and record them as content words, calculate the correlation between content words and superordinate labels, and then perform the difference judgment step; Difference judgment: judging whether the correlation degree is less than the preset fifth threshold value, if so, executing the correction step, otherwise, executing the fourth judgment step; Correction: Intervene in the labeling step and modify and update the upper label; Fourth judgment: judging whether the number of times of difference similarity is greater than the sixth threshold value, if so, executing the supplementary calculation step; Supplementary calculation: Calculate the similarity between the hyperlinked tags and record it as synonymous similarity; Supplementary judgment: judge whether the synonymous similarity is greater than the seventh threshold, if so, perform the supplementary labeling step; Supplementary tags: Set multiple upper-level tags in a pop-up window for the query requester to make additional selections and correct the query request.

[0132] Specifically, first, all query request records of the same query requester within a certain period of time are collected, and text preprocessing is performed on each query request. Stop words and punctuation marks are removed, and stem extraction or word form restoration is performed using an NLP library (such as NLTK, spaCy, Porter Stemmer); feature vectors of query requests are extracted using models such as TF-IDF, word vectors (such as Word2Vec) or BERT; cosine similarity or other similarity indicators (Euclidean distance, Manhattan distance, Jaccard similarity) between feature vectors of adjacent query requests are calculated as their similarity; a third threshold is preset, which represents the standard for sufficient similarity between query requests, and is determined based on historical data and experiments. In this embodiment, the third threshold is 0.75; if the calculated similarity is greater than or equal to the third threshold, it is considered that there is significant similarity between the query requests, and further analysis of the differences is required; otherwise, it is considered that the query requests have changed significantly, and the knowledge retrieval step is directly executed.

[0133] For similar query requests, the differences between them are recorded (such as changes in keywords, changes in query intent, etc.); the number of times differences appear in similar query requests, i.e., the number of similar differences, is counted; a fourth threshold is preset, and the fourth threshold is determined based on historical data and experiments. In this embodiment, the fourth threshold is preferably 5 times; if the number of similar differences is greater than or equal to the fourth threshold, the difference acquisition step is performed; otherwise, it may indicate that the change in the query request is not enough to attract attention, and observation can continue or other processing can be performed.

[0134] Extract the modified or replaced words from the differences as difference words; analyze the difference words to find out the commonalities or patterns between them, which may reflect the core or literal words of the query request; calculate the association between the literal words and the current upper tag (i.e., the topic or category tag to which the query request is classified), and a semantic-based similarity calculation method may be used; preset a fifth threshold value, which represents the minimum acceptable level of association between the literal words and the upper tag, wherein the fifth threshold value is determined based on historical data, experiments or user feedback, and the value of the fifth threshold value in this embodiment is 0.8; if the association degree is less than the fifth threshold value, it is considered that the current upper tag may be inaccurate and needs to be corrected; otherwise, perform the fourth judgment step.

[0135] According to the results of the difference analysis, the upper label of the query request is manually or automatically adjusted to more accurately reflect the query intent; the updated upper label is saved to the database for subsequent query use; a sixth threshold is preset, wherein the sixth threshold is determined based on historical data, experiments or user feedback, and the value of the sixth threshold in this embodiment is 7 times; it is used to determine whether the number of difference similarities has reached the level that requires more in-depth analysis. If the number of difference similarities is greater than or equal to the sixth threshold, a supplementary calculation step is performed; otherwise, it may indicate that the change in the query request is still within an acceptable range and can continue to be observed.

[0136] Calculate the similarity between the current upper tag and other related or potential upper tags, that is, synonymous similarity; a vocabulary-based similarity calculation method (such as Jaccard similarity, cosine similarity) or a semantic-based similarity calculation method (such as WordNet, semantic vector space model) can be used; preset a seventh threshold, indicating the minimum acceptable level of synonymous similarity, wherein the determination of the seventh threshold is based on historical data and experiments, and the value of the seventh threshold in this embodiment is 0.75; if the synonymous similarity is greater than or equal to the seventh threshold, it is considered that there are multiple upper tags that may be applicable to the current query request, and a supplementary tag step is required; otherwise, it may indicate that the current upper tag is accurate enough and no further supplementation is required.

[0137] Multiple possible upper-level tags are displayed to the query requester in a pop-up window for selection or confirmation; the query requester is allowed to make additional selections or modifications based on the tags in the pop-up window to further refine the query intent; based on the user's selection, the upper-level tags of the query request are updated and saved to the database.

[0138] For example, user A enters the following query request in the system: The first query is: "What are the methods for power load forecasting?"; The second query (later): "What are the power load forecasting models?"; The third query (later): "What is the algorithm principle of the power load forecasting model?"; The system first uses an NLP library (such as spaCy) to preprocess the query request and extract the feature vector. Then, it calculates the cosine similarity between adjacent query requests: Similarity between the first and second queries: 0.85; Similarity between the second and third queries: 0.78; The preset third threshold is 0.75. Since the similarity (0.85) between the first and second queries is greater than the third threshold, the system believes that there is a significant similarity between them and needs to enter the difference statistics step.

[0139] The system records the difference between the two queries: "method" and "model". The difference similarity count is 1 (because there is only one significant difference).

[0140] The system extracts the difference words "method" and "model" and calculates their relevance to the current superordinate label "power load forecast". Since "method" and "model" are both core words in the field of power load forecasting and have a high relevance to the label "power load forecasting" (assuming 0.9 and 0.95), they do not meet the correction condition (i.e., the relevance is less than the fifth threshold of 0.8).

[0141] However, considering that the user may be gradually exploring different aspects of power load forecasting, continue to observe subsequent queries.

[0142] When user A makes the third query, the system notices that the number of similar differences has increased to 2 (although it has not yet reached the sixth threshold of 7 in this example, in order to show the complete process, we assume that there are more similar queries in the future that make the number of similar differences reach 7).

[0143] At this time, the system calculates the synonymous similarity between the current upper label "power load forecast" and other potential upper labels (such as "power market transactions" and "power system operation"); assuming that the synonymous similarity with "power market transactions" is 0.65 (lower than the seventh threshold of 0.75), and the synonymous similarity with "power system operation" is 0.5 (also lower than the seventh threshold).

[0144] Since no other hyperlinked tags have a sufficiently high synonymous similarity with "power load forecasting", the system does not perform the additional tagging step for the time being. However, in order to cope with more similar queries that may appear in the future, the system decides to optimize the tagging system and consider introducing more detailed tags (such as "power load forecasting method", "power load forecasting model", etc.) when appropriate.

[0145] Pop-up window display and user selection (assuming that the supplementary tag conditions are met in subsequent queries): Assuming that in subsequent queries, the number of similar differences continues to increase, and the system finds that tags such as "power load forecasting model" and "power load forecasting algorithm" have a high degree of synonymous similarity. At this time, the system will display these possible upper-level tags to user A in the form of a pop-up window: Pop-up window title: "Please select or confirm your query subject"; tag options: "power load forecasting", "power load forecasting model", "power load forecasting algorithm"; User A can select or confirm a more specific tag according to his or her needs to further refine the query intention. The system updates the upper-level tag of the query request based on the user's selection and saves it to the database.

[0146] Reference Figure 4In other implementations, a user participation step, a manual correction step, an associated label matching step, a feedback statistics step, and an update step are provided between the difference determination step and the correction step; User participation: recommending the content words to the query requester, judging whether the content words meet the query requester's needs, if yes, executing the correction step and the associated tag matching step; otherwise, executing the manual correction step; Manual correction: The query requester corrects the content words and then performs the associated tag matching step; Related tag matching: Match relevant hypernyms through lexical words and push them to the query requester. According to the query requester's feedback, perform feedback statistics. Feedback statistics: obtain the number of times a certain content word is corrected in relation to its hypernym, and determine whether it is greater than the eighth threshold. If so, perform the update step; Update: Update the association between the hypernym and the content word for matching subsequent query requests.

[0147] Specifically, the system recommends the extracted lexical words to the query requester and asks whether these words accurately reflect his query intention; if the query requester confirms that these lexical words meet the requirements, the system executes the associated label matching step; if the query requester believes that these lexical words are inaccurate or do not fully meet the requirements, the manual correction step is entered; the query requester corrects the lexical words recommended by the system, adds or deletes words, so as to more accurately express his query intention; the system receives the corrected lexical words and re-executes the associated label matching step; the system matches the relevant superordinate labels according to the corrected lexical words using algorithms or models; the system pushes the matched superordinate labels to the query requester and asks about his satisfaction; based on the feedback from the query requester, the system executes the feedback statistics step, and the system records the number of times the correlation between each lexical word and the superordinate label is corrected.

[0148] The system determines whether the number of times the association between a certain content word and a superordinate label is revised is greater than an eighth threshold, wherein the eighth threshold is determined based on historical data and experiments, and the value of the eighth threshold in this embodiment is 3 times; if it is greater than the eighth threshold, it means that the association between the content word and the superordinate label is not stable or accurate enough and needs to be updated, then the update step is performed; if it is not greater than the eighth threshold, the next query request is processed continuously; the system updates the association between the superordinate label and the content word according to the feedback statistical results, and the updated association is used for subsequent query request matching to improve the matching accuracy and efficiency.

[0149] For example, user B enters the following query in the system: "What are the influencing factors of power load forecasting?" The system first calculates the similarity between this query and the previous query and finds that there is no significant similarity, so it directly enters the knowledge retrieval step. However, during the knowledge retrieval process, the system extracts the literal terms "power load forecasting" and "influencing factors".

[0150] It determines whether these lexical terms meet the query requirements, but due to the lack of sufficient context information, the system cannot determine whether these terms are completely accurate. Therefore, it decides to enter the user participation step; the system recommends the extracted lexical terms "power load forecast" and "influencing factors" to user B, and asks whether these terms accurately reflect his query intention. After seeing this, user B thinks that "power load forecast" is accurate, but "influencing factors" may not be specific enough. What he wants to know more is "which factors will affect the accuracy of power load forecast". Therefore, user B chooses to modify the lexical terms and changes "influencing factors" to "factors that affect the accuracy of power load forecast". The system receives the modified lexical terms and prepares to match the associated tags.

[0151] The system uses algorithms or models to match relevant hypernyms based on the revised content words "power load forecast" and "factors affecting the accuracy of power load forecast". The system finds hypernyms associated with these content words, such as "power load forecast technology", "power market analysis", etc. (Note: this is just an example, and the actual match may be more accurate).

[0152] The system pushes the matched upper-level tags to user B and asks him / her about his / her satisfaction. After seeing it, user B thinks that the tag "power load forecasting technology" is close, but he / she would like to have a tag specifically about "factors affecting power load forecasting accuracy". Therefore, user B gives feedback and suggests that the system add or optimize relevant tags.

[0153] The system records this feedback and counts the number of times the correlation between the content word "factors affecting the accuracy of power load forecasting" and the superordinate label is corrected. Since this is the first correction, the number of corrections is 1, which has not yet reached the eighth threshold (assuming 3 times).

[0154] Update step (assuming that the update condition is met in subsequent queries): In subsequent queries, if user B or other users make multiple revisions to the association between the content word "factors affecting the accuracy of power load forecasting" and the upper-level label (for example, the number of revisions reaches 3 times or more), the system will determine that the association between the content word and the upper-level label is not stable or accurate enough and needs to be updated.

[0155] The system updates the association between the hypertag and the content words based on the feedback statistics. For example, the system can add a new hypertag "Factors affecting the accuracy of power load forecasting" and associate it with the content word "Factors affecting the accuracy of power load forecasting". The updated association will be used for subsequent query request matching to improve the accuracy and efficiency of matching.

[0156] Reference Figure 5 In other implementations, after the association tag matching step, a demand determination step and a single association step are also provided; Demand judgment: judge whether the generated tags meet the needs of the query requester. If so, execute the feedback statistics step; otherwise, execute the single association step; Single association: The query requester enters a custom tag, matches it from the upper-level tags, and establishes a single association for the query requester's query request, and then executes the knowledge retrieval step.

[0157] Specifically, first, the system receives the query request from the query requester, and parses the query content through natural language processing technology to generate a series of initial labels. Then, the system uses a pre-trained label association model to match these initial labels with the upper-level labels in the knowledge base to find the most relevant upper-level label set; after the associated label matching is completed, the system enters the demand judgment step. The purpose of this step is to determine whether the generated label set accurately reflects the actual needs of the query requester. The system can make demand judgments in the following ways: Based on user feedback: The system displays the matched labels to the query requester and asks whether he is satisfied. The user can give feedback by clicking "yes" or "no"; the system uses machine learning algorithms, combined with the query requester's historical query records and the context information of the current query, to automatically determine the accuracy of the labels; the system presets a series of rules, such as label quantity thresholds, label relevance scores, etc., and determines whether the label set meets the needs based on these rules.

[0158] If the system determines that the label meets the requirements, it enters the manual input step, allowing the query requester to further refine or modify the label. If it does not meet the requirements, the system returns to the label generation step to regenerate or adjust the label.

[0159] After the demand judgment step, if the query requester chooses to enter a custom tag by himself, the system enters the single association step. This step allows the query requester to enter a specific custom tag according to his own needs. The system matches these custom tags with the upper tags in the knowledge base, finds the most suitable upper tag, and establishes a single association. Single association means that these custom tags are only used for this query request and will not be permanently stored in the system's tag association model.

[0160] After establishing a single association, the system retrieves relevant information from the knowledge base based on the associated upper tags and custom tags. The search results are sorted by relevance score and displayed to the query requester, who can view detailed information, perform further screening, or submit new query requests as needed.

[0161] For example, the query requester wants to query information about "the application of smart grid technology in power distribution". The system first parses the query content and generates initial tags such as "smart grid technology", "power distribution", and "application". Then, the system uses the pre-trained tag association model to match these initial tags with the upper-level tags in the knowledge base and find relevant upper-level tags such as "energy technology" and "power system management".

[0162] In the demand judgment step, the system shows the query requester the matching tags and asks whether he is satisfied. The query requester finds that "smart grid technology" and "power distribution" are accurate, but the tag "application" is also too broad and not specific enough. Therefore, the query requester chooses to enter the single association step and enters the custom tags "energy saving effect" and "fault prediction" according to his needs.

[0163] The system matches these custom tags with the upper tags in the knowledge base and finds the most appropriate upper tags "energy efficiency improvement" and "preventive maintenance". Then, based on these associated upper tags and custom tags, the system retrieves relevant information from the knowledge base, such as how smart grid technology improves energy efficiency in power distribution, how to predict faults through smart grid technology, etc., and sorts them according to the relevance score and displays them to the query requester.

[0164] Reference Figure 5 In other embodiments, after the single association step, an association statistics step is also provided; Association statistics: Count the number of single associations established with the same upper-level tag, and determine whether it is greater than the ninth threshold. If so, perform the tag conversion step; Label conversion: Convert custom labels into higher-level labels and associate them with the knowledge base.

[0165] Specifically, in the single association step, the system allows users to enter custom tags according to query requirements, and matches these tags with upper-level tags in the knowledge base to establish a single association. This association only applies to the current query and will not permanently change the system's labeling system.

[0166] After the single association step, the system enters the association statistics step. The system maintains an association statistics table to record the number of associations between each upper tag and the custom tag, traverses the single association records, and counts the number of associations between each upper tag and the custom tag; a ninth threshold is set to determine whether the association between a certain upper tag and the custom tag is frequent enough for label conversion, where the ninth threshold is determined based on historical data or experiments.

[0167] When the number of associations between a certain upper-level tag and a certain custom tag exceeds the ninth threshold, the system enters the tag conversion step. The purpose of this step is to convert frequently used custom tags into upper-level tags and establish a permanent association with the knowledge base, thereby optimizing the tag system and improving retrieval efficiency.

[0168] In other implementations, after the knowledge retrieval step, a recommendation statistics step is also provided; Recommendation statistics: obtain the click-through rate of the knowledge recommended by the query requester, determine the degree of matching, and correct the ranking of subsequent similar query requirements.

[0169] Specifically, the user's click behavior on the recommended knowledge, including the number of clicks, click time, etc., is recorded, and the click rate of each recommended knowledge is calculated, that is, the ratio of the number of clicks to the number of displays, and the degree of match between the recommended knowledge and the user's query needs is determined based on the click rate. The higher the click rate, the higher the degree of match; the matching degree information is stored and used for subsequent sorting corrections of similar query needs; based on the results of recommendation statistics, the system corrects the sorting algorithm for subsequent similar query needs, and the basis for the correction is the user's previous click rate on the recommended knowledge, that is, the degree of match; when the system receives a new similar query request, it adjusts the sorting of the retrieval results based on the previous recommendation statistics; for knowledge items with a high degree of match, a higher sorting priority is given to make them easier to be found by users; for knowledge items with a low degree of match, their sorting priority is lowered, or they are excluded from the retrieval results.

[0170] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for constructing a knowledge service system, characterized in that: The following steps are involved: Knowledge acquisition: Collecting knowledge data from knowledge sources in multiple fields; Preprocessing: preprocess the knowledge data to eliminate noise; Classification: Classify the acquired knowledge data, generate corresponding primary labels for each category, and store the classified data separately to form a knowledge base; Similarity calculation: calculate the similarity between knowledge bases of different categories; First judgment: judging whether the similarity is greater than a first threshold, if so, executing the association step; Association: Establish associations between knowledge based on similarities; Second judgment: judging whether the similarity is greater than a second threshold, if so, executing the labeling step; Labeling: Analyze their common features and construct new hyperlinks to summarize and reflect the common attributes of this knowledge; Tag catalog generation: Create a tag catalog based on knowledge base categories, primary tags, and upper-level tags; Query demand acquisition: obtain the user's knowledge retrieval needs, determine keywords or tags based on the retrieval needs, and generate query requests; Knowledge retrieval: Based on the query request, combined with the upper-level tags and knowledge associations, the results are filtered and sorted to make knowledge recommendations.

2. The method for constructing a knowledge service system according to claim 1, characterized in that: Among them, classification The steps also include: multi-dimensional classification based on the subject, content, and form characteristics of the knowledge, and assigning one or more labels to the knowledge under each dimension.

3. The method for constructing a knowledge service system according to claim 1, characterized in that: The labeling step also includes: analyzing the commonalities and differences of similar knowledge bases, as well as their positions and functions in the knowledge system, to determine the naming and hierarchy of the superordinate labels.

4. The method for constructing a knowledge service system according to claim 1, characterized in that: After the query requirement acquisition step, a label delineation step is also provided; Tag delineation: Based on the query request, preliminary knowledge base categories and upper-level tags are recommended for the queryer to choose from, and the query request is enriched.

5. The method for constructing a knowledge service system according to claim 1, characterized in that: The knowledge retrieval step also includes: dynamically adjusting the weights of upper-level tags and knowledge associations according to the user's search history, preferences and behavior patterns, and optimizing the knowledge recommendation ranking.

6. The method for constructing a knowledge service system according to any one of claims 1 to 5, characterized in that: A deviation acquisition step is also provided after the query requirement acquisition step; Deviation acquisition: including demand similarity calculation step, third judgment step, difference statistics step, difference acquisition step, difference judgment step, correction step, fourth judgment step, supplementary calculation step, supplementary judgment step and supplementary labeling step; Demand similarity calculation: Calculate multiple query requests from the same query requester and calculate the similarity between the previous and next query requests; Third judgment: judging whether the similarity is greater than a preset third threshold, if so, executing the difference statistics step, otherwise, executing the knowledge retrieval step; Difference statistics: counting the number of similar differences and judging whether it is greater than the fourth threshold, if so, executing the difference judgment step; Difference acquisition: extract the words that have been modified multiple times and record them as difference words, obtain the common points between multiple difference words and record them as content words, calculate the correlation between content words and superordinate labels, and then perform the difference judgment step; Difference judgment: judging whether the correlation degree is less than the preset fifth threshold value, if so, executing the correction step, otherwise, executing the fourth judgment step; Correction: Intervene in the labeling step and modify and update the upper label; Fourth judgment: judging whether the number of times of difference similarity is greater than the sixth threshold value, if so, executing the supplementary calculation step; Supplementary calculation: Calculate the similarity between the hyperlinked tags and record it as synonymous similarity; Supplementary judgment: judge whether the synonymous similarity is greater than the seventh threshold, if so, perform the supplementary labeling step; Supplementary tags: Set multiple upper-level tags in a pop-up window for the query requester to make additional selections and correct the query request.

7. The method for constructing a knowledge service system according to claim 6, characterized in that: A user participation step, a manual correction step, an associated label matching step, a feedback statistics step and an update step are also provided between the difference judgment step and the correction step; User participation: recommending content words to the query requester, judging whether the content words meet the query requester's needs, and if so, executing the correction step and the associated tag matching step; Otherwise, the manual correction step is performed; Manual correction: The query requester corrects the content words and then performs the associated tag matching step; Related tag matching: Match relevant hypernyms through lexical words and push them to the query requester. According to the query requester's feedback, perform feedback statistics. Feedback statistics: obtain the number of times a certain content word is corrected in relation to its hypernym, and determine whether it is greater than the eighth threshold. If so, perform the update step; Update: Update the association between the hypernym and the content word for matching subsequent query requests.

8. The method for constructing a knowledge service system according to claim 7, characterized in that: After the association tag matching step, there are also a demand judgment step and a single association step; Demand judgment: judge whether the generated tags meet the needs of the query requester. If so, execute the feedback statistics step; otherwise, execute the single association step; Single association: The query requester enters a custom tag, matches it from the upper-level tags, and establishes a single association for the query requester's query request, and then executes the knowledge retrieval step.

9. The method for constructing a knowledge service system according to claim 8, characterized in that: After the single association step, an association statistics step is also provided; Association statistics: Count the number of single associations established with the same upper-level tag, and determine whether it is greater than the ninth threshold. If so, perform the tag conversion step; Label conversion: Convert custom labels into higher-level labels and associate them with the knowledge base.

10. The method for constructing a knowledge service system according to claim 1, characterized in that: After the knowledge retrieval step, there is also a recommendation statistics step; Recommendation statistics: obtain the click-through rate of the knowledge recommended by the query requester, determine the degree of matching, and correct the ranking of subsequent similar query requirements.

Images