Optimization Methods of Index Retrieval in Digital Libraries

Abstract

The present invention relates to a retrieval optimization method, specifically an index retrieval optimization method for a digital library, which adopts the following technical solutions: S1, dynamic hierarchical index construction technology: first, multi-dimensional content analysis is performed on digital resources in the library, including subject, keyword, author, and publication time information; documents are classified according to the information, and a content hierarchy is constructed; S2, an index optimization strategy based on usage patterns: by performing data analysis on user retrieval behavior, the user's access frequency, query type, and access path in different categories of documents are monitored, and typical user usage patterns are identified; S3, semantic association retrieval optimization; S4, index update and resource allocation optimization: when new document resources are added to the library or existing resources are modified, the index is updated in real time, and the index structure is adjusted according to content classification and user behavior.

Description

Technical Field

[0001] The present invention relates to a retrieval optimization method, in particular to an optimization method for index retrieval of a digital library. Background Art

[0002] Among current digital library index retrieval optimization methods, existing technologies, such as the "Digital Library Content Analysis System and Method," with Chinese invention patent number 2017103045690, do offer advantages in many areas, particularly in content analysis and user input processing. However, with the increasing sophistication of user needs and the dramatic growth in information volume, this approach still suffers from shortcomings and drawbacks in many areas, particularly in key areas such as large-scale data processing, personalized retrieval, and semantic analysis. These issues directly impact the effectiveness of digital library index retrieval optimization.

[0003] First, existing indexing and retrieval systems lack effective utilization of user behavior feedback, particularly regarding data such as click-through rate and dwell time. The system described in Patent No. 2017103045690 extracts user input and performs preliminary analysis through signal reception and judgment steps, but its core still relies on keyword matching and classification models. While the classification step groups different content and attempts to model based on content characteristics, it lacks in-depth consideration of the role of user behavior feedback in optimizing the retrieval system. For example, user click-through rate and dwell time are often important metrics for measuring the value and relevance of documents, but existing systems do not leverage this behavioral data to dynamically adjust indexing priorities and model parameters. This makes it difficult to personalize the system to meet user needs, resulting in insufficient accuracy and relevance in search results and a limited user experience. Second, existing technologies have serious limitations in semantic understanding and contextual analysis. The method described in Patent No. 2017103045690 primarily relies on a first matching device and a second matching device for preliminary content matching, but its matching approach still primarily relies on direct keyword matching. While the system can perform some analysis and grouping of user input through signal reception and classification steps, it lacks a deep understanding of the semantic intent behind the query. This is especially true when users enter queries with complex semantics. Existing systems are unable to perform contextual analysis, resulting in overly mechanistic matching results that fail to capture the user's true needs. This lack of semantic matching not only reduces the accuracy of retrieval results but also makes the system perform poorly for long or semantically ambiguous queries, making it prone to biased selection of irrelevant documents or results.

[0004] Third, existing systems lack support for personalized user needs and the ability to dynamically adjust in real time. Digital library users often have unique needs and interests. While the classification devices in existing patents can group and analyze content based on its characteristics, they fail to fully consider users' personalized needs and long-term behavioral habits. Over time, users' interests and behavioral patterns change, yet the system's index optimization process is static, and index results do not adjust in real time based on long-term user behavioral feedback. This design limits the system's adaptive capabilities and makes it impossible to dynamically optimize the index structure based on user behavior patterns. This results in the system being unable to achieve accurate and rapid responses when faced with a large number of users searching simultaneously.

[0005] In addition, although the signal receiving and judgment steps in the existing technology have improved the speed of information extraction to a certain extent, their performance in large-scale data processing and real-time index updating is not ideal. With the continuous growth of the amount of documents in digital libraries, it is difficult for existing systems to effectively process massive amounts of data in a short period of time. In particular, in terms of real-time indexing and rapid response, the system's processing speed is difficult to meet the user's efficient retrieval needs. For example, when the second matching device faces a large amount of user input, it is easy for the system response speed to decrease due to excessive processing time. This not only increases the user's waiting time, but may also miss some high-value real-time information. This processing bottleneck is particularly prominent when dealing with modern large-scale data retrieval. Summary of the Invention

[0006] The purpose of the present invention is to provide an optimization method for index retrieval of a digital library, thereby solving some of the drawbacks and deficiencies pointed out in the background art.

[0007] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0008] S1. Dynamic hierarchical index construction technology:

[0009] S1.1. First, conduct a multi-dimensional content analysis of the digital resources in the library, including subject, keyword, author, and publication date information; classify the documents based on this information and construct a content hierarchy;

[0010] S1.2. Based on the classification, a dynamic index is constructed for different document levels, including topics, content details, and semantic associations. This allows frequently used content to be quickly retrieved, while detailed content is retained in the deep index.

[0011] S1.3. Regularly adjust the index structure based on user usage frequency and search habits to bring frequently used information closer to the surface and optimize search speed;

[0012] S2. Index optimization strategy based on usage pattern:

[0013] S2.1. Analyze user search behavior data to monitor user access frequency, query types, and access paths in different document categories, and identify typical user usage patterns.

[0014] S2.2. Dynamically adjust the priorities of different index items based on user behavior monitoring results, and promote frequently accessed documents or related topics to the top of the search results;

[0015] S2.3. Locally cache high-frequency indexes based on usage patterns, quickly return results, and control and reduce resource usage.

[0016] S3, semantic association retrieval optimization:

[0017] S3.1. Introducing natural language processing technology to perform semantic analysis on user queries, identify keywords, understand the underlying semantic intent, and expand the scope of indexing.

[0018] S3.2. Based on semantic analysis, intelligently match user queries with the semantics of document content, optimize the index, and enable retrieval of documents with similar semantics but different keywords.

[0019] S3.3. Optimize the semantic matching model and index structure based on the click-through rate of search results and user dwell time feedback;

[0020] S4. Index update and resource allocation optimization:

[0021] S4.1. Update the index in real time when new literature resources are added to the library or existing resources are modified, and adjust the index structure based on content classification and user behavior;

[0022] S4.2. Use distributed database technology to store indexes on multiple nodes and support large-scale parallel retrieval;

[0023] S4.3. During the retrieval process, resources are dynamically allocated and each retrieval node is scheduled based on the current query volume and load conditions.

[0024] Furthermore, the method for constructing the dynamic index includes:

[0025] S1. First, we use content analysis and classification algorithms to conduct a multi-dimensional analysis of digital resources in the library, including subject matter, content details, and semantic association information, and establish a preliminary structure of a dynamic index based on the document hierarchy. For each document, its index is calculated hierarchically using the formula:

[0026]

[0027] Among them, S(x) represents the dynamic index score of the document, which is the core variable used to measure the importance of the document during the index optimization process; f1(x) is the subject relevance function of the document, which reflects the degree of match between the document content and a specific topic; f2(x) is the content detail function of the document, which represents the richness of the description of the document at the detail level; α is the adjustment coefficient of the content details, which is used to adjust the weight of the details in the overall index; β(t) is the time dynamic adjustment coefficient related to user interaction, which dynamically adjusts the index structure over time; T is the time period for index weight adjustment, which is used to represent the time window for calculating the index during the optimization process;

[0028] S2. Secondly, by monitoring user behavior, including query frequency, access time, and user path, the frequency perception model is used to dynamically adjust the index position of the document. The index weight adjustment is calculated using the formula:

[0029]

[0030] Among them, W(u) is the result of index weight adjustment caused by user behavior, which represents the index priority after the user visits a specific document; p i is the access frequency of the i-th document, that is, the number of times the document is queried by users; λ is the adjustment parameter that controls the frequency sensitivity of user behavior and determines the intensity of the impact of access frequency on the document index weight; t represents the current query time, which adjusts the perception of document access frequency over time;

[0031] S3. Finally, we use semantic association analysis technology to perform deep semantic matching on the documents. We use a semantic vector model based on deep learning to dynamically optimize the semantic level of different document contents. The semantic level is determined by the formula:

[0032]

[0033] Among them, L(s) represents the semantic hierarchy score of the document, which is the result of hierarchical division of the document through semantic analysis; g1(s) is the basic semantic matching function, which represents the basic semantic relevance of the document; g2(s) is the semantic extension function, which is responsible for matching the document in the semantic space; γ is the semantic extension coefficient, which controls the impact of semantic extension on the matching results; n is the detail complexity of the document, which is used to measure the complexity of the detailed description of the document at the semantic level.

[0034] Furthermore, the method for identifying typical user usage patterns includes:

[0035] S1. First, we use log tracking and cookie technology to record users' search behavior data in real time, including the document category, query type, access path, and dwell time of each visit. We then use the following formula to perform multi-dimensional modeling and optimization of user behavior data:

[0036]

[0037] Among them, B(x) represents the comprehensive score of user behavior; f1(x) is the access frequency function, which represents the frequency of user access to a certain document; f2(x) is the residence time function, which reflects the length of time the user stays in the document; ω1 and ω2 are the weight factors of access frequency and residence time respectively; ψ(t) is the behavior dynamic adjustment function, which affects the index priority over time; T is the behavior time window for real-time monitoring;

[0038] S2. Based on the collected behavioral data, cluster analysis algorithms such as K-Means or DBSCAN are used to classify users, identify typical user usage patterns, including users who frequently search for documents in a specific category or often follow similar paths, and mathematically model these patterns.

[0039] Then, based on the user's historical behavior data, a deep learning model based on the recurrent neural network (RNN) is constructed to predict the user's future search needs through the following formula:

[0040]

[0041] Among them, P(u) is the prediction of the user's future behavior; u i Indicates the i-th retrieval behavior, f3(u i ) is the user behavior sequence feature function, which describes the user's retrieval behavior pattern in the past; λ is the adjustment parameter that controls the prediction sensitivity, t i Indicates the time when the behavior occurs.

[0042] Furthermore, based on the prediction results of P(u), the index of relevant documents is loaded in advance, and the adaptive preloading mechanism is used to optimize the retrieval efficiency; then, according to the actual feedback of user behavior, the display order of the index is adjusted, and the dynamic sorting of the index is optimized using the formula:

[0043]

[0044] Among them, S(y) represents the priority ranking score of the document in the index; r1(y i ) and r2(y i ) are the user's query frequency and path click rate function for the document, γ1 and γ2 are adjustment coefficients, which control the influence of the two on the priority.

[0045] Furthermore, the framework method used to identify keywords and understand the semantic intent behind them is:

[0046] S1. First, use the deep learning natural language processing model to perform deep semantic analysis on the user query and calculate the context semantic vector of the user query using the following formula:

[0047]

[0048] Among them, V(q) represents the context semantic vector of the query statement, which is used to quantify the semantic content of the user query; f1(w i ) is the keyword semantic weight function, which represents the i-th keyword w in the query i The weight in the semantic space reflects the relevance between the keyword and the query topic; f2(c i ) is the context-dependent function, indicating that the keyword w i Its context c i The semantic relationship between them; α and β are adjustment coefficients used to balance the weights of keywords and context; γ(t) is a time dynamic adjustment function that represents the semantic characteristics of the query that change over time; T is the time window of semantic parsing, which represents the duration of semantic parsing;

[0049] S2. Then, using semantic intent recognition technology, we classify user queries into different types through an intent classification model, including information queries, navigation queries, and transaction queries. We then use the following formula to construct a multi-level index for the intents:

[0050]

[0051] Among them, I(s) represents the hierarchical score of query intent, which is an indicator used to quantify user query intent; g i (x) is a semantic level function, which represents the score of document x in different intent levels and reflects the semantic matching degree between the document and the user query; λ is a parameter that controls the time sensitivity of the model and determines the speed at which the query intent changes over time; t i is the timestamp of the query;

[0052] S3. Finally, through the dynamic semantic expansion mechanism, the knowledge graph or deep semantic model is used to expand keywords according to the semantic intent of the user's current query statement. The calculation formula is:

[0053]

[0054] Among them, R(q) is the recommended search result set after expansion, which represents the final search result after semantic expansion; h1(y i ) is a direct keyword matching result function, which represents the document y directly related to the query q i score; h2(y i ) is a semantic expansion result function, which represents the documents y that are semantically related to the query q but not directly matched by keywords.i score; μ is the semantic expansion coefficient, which controls the influence of the expansion result on the final recommendation; n is the total number of retrieved documents.

[0055] Furthermore, the optimization of the semantic matching model and index structure includes the following steps:

[0056] S1. First, collect the user's click behavior and dwell time in the search results in real time, associate the data with the semantic relevance of the document, and calculate the impact of user behavior feedback on the semantic priority of the document based on the formula:

[0057]

[0058] Among them, P(u) represents the semantic priority score of document u, C u (t) is the change of the click rate of the document over time t, S u (t) is the user's residence time function of the document, α and β are adjustment coefficients that control the relative importance of click-through rate and residence time, λ is the time decay factor, and T is the collection period of feedback data;

[0059] S2. The behavioral feedback is then input into the semantic matching model to dynamically adjust the position of the document in the semantic space. A deep learning algorithm is used to analyze long-term behavior and optimize the parameters of the semantic model. Based on user dwell time and click-through rate, a reinforcement learning algorithm is used to dynamically adjust the index structure and update the index priority of the document using the formula:

[0060]

[0061] Among them, I(v) is the priority score of document v in the index, S i (v) is the residence time of the document during the i-th retrieval, C i (v) is the i-th click rate, μ is the coefficient of the control feedback on the priority adjustment, t i timestamp for the action;

[0062] S3. Finally, we optimize the semantic expansion model based on behavioral feedback. We dynamically optimize and expand the keyword pool based on user clicks and dwell time, and use the formula to calculate personalized semantic expansion recommendations:

[0063]

[0064] Among them, E(q) is the semantic expansion score of query q, f1(w i ) is the keyword w i The expansion weight, f2(c i ) is the context semantic expansion function, k1 and k2 are expansion parameters that control the relative influence of keywords and context, and t is the time when the query occurs.

[0065] The beneficial effects of the present invention are as follows: by dynamically adjusting the index priority of documents and using an adaptive preloading mechanism, the system can preload the index of relevant documents based on the user's query history, behavioral feedback, and semantic matching, thereby significantly improving the retrieval speed and reducing user waiting time.

[0066] By combining behavioral data such as user click-through rates and dwell time, the system optimizes the semantic priority of documents in real time, prioritizing the presentation of documents that best match user needs. Through a semantic expansion mechanism, the system can also find documents that are semantically related but not completely matched, further enriching search results and ensuring a more precise search experience for users.

[0067] Based on deep learning and reinforcement learning algorithms, the system dynamically adjusts the index structure based on users' long-term behavior patterns, ensuring personalized document recommendations. Furthermore, the semantic expansion mechanism optimizes the keyword pool by analyzing users' clicks and dwell time, providing personalized semantic expansion recommendations to meet the search needs of different users. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of the optimization method for index retrieval of a digital library of the present invention.

[0069] Figure 2 Flowchart of the method used to construct a dynamic index in the present invention.

[0070] Figure 3 A flow chart of a method for identifying typical user usage patterns is provided for the present invention.

[0071] Figure 4 Flowchart of the framework method used by the present invention to identify keywords and understand the semantic intent behind them.

[0072] Figure 5 This is a flow chart of the method for optimizing the semantic matching model and index structure of the present invention. DETAILED DESCRIPTION

[0073] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0074] Combined with attachment Figure 1 , improve retrieval efficiency and user experience through dynamic hierarchical index construction technology.

[0075] First, in step S1.1, the system performs a multi-dimensional content analysis of the library's digital resources, analyzing key information such as the document's subject, keywords, author, and publication date. This information is used to categorize the documents, thereby constructing a content hierarchy that allows subsequent retrieval operations to more efficiently organize and manage document data.

[0076] Based on this classification, in step S1.2, a dynamic index structure is constructed for different levels of the document, such as topic, content details, and semantic associations. This dynamic index structure is divided into two parts: shallow and deep. Common, high-frequency content is placed in the shallow index for rapid retrieval, while more detailed content is retained in the deep index, awaiting more complex or specific queries. This index structure ensures that commonly used information can be retrieved quickly without affecting overall efficiency due to the system processing detailed information.

[0077] To further optimize the user's search experience, in step S1.3, the system regularly and dynamically adjusts the index structure based on user usage frequency and search habits. By analyzing user usage data, the system can place frequently accessed documents or commonly used information closer to the surface of the index, making these high-frequency contents more quickly accessible. At the same time, the system continuously monitors user behavior to ensure that the index structure is always in optimal condition.

[0078] Improve retrieval efficiency and user experience through index optimization strategies based on usage patterns.

[0079] First, in step S2.1, the system conducts a comprehensive data analysis of users' search behavior, monitoring their access frequency, query types, and access paths across different document categories. This includes recording data such as user query history, clickthroughs, and browsing paths, thereby identifying typical user usage patterns. These usage patterns not only reflect user preferences but also reveal the document types and topics that users frequently access during their searches.

[0080] In step S2.2, the system dynamically adjusts the priority of different index items based on user behavior monitoring. This process automatically elevates frequently accessed documents or related topics to the top of search results, ensuring that these documents are prioritized for similar queries. This dynamic adjustment mechanism enables the system to adapt to changing user needs, continuously optimizing the order in which search results are displayed, and improving the relevance of search results and user satisfaction.

[0081] Next, in step S2.3, the system locally caches frequently accessed index sections based on user usage patterns. This caching mechanism significantly improves search speed because when a user issues a query, the system can quickly return results directly from the cache, eliminating the need to retrieve the global index each time. Furthermore, local caching effectively controls and reduces resource usage, enabling the system to maintain low resource consumption while providing efficient search services.

[0082] Through the semantic association retrieval optimization strategy, the relevance and accuracy of the retrieval results are improved.

[0083] First, in step S3.1, the system incorporates natural language processing (NLP) technology to perform deep semantic analysis of the user's query. This analysis not only identifies the keywords entered by the user but also understands the semantic intent behind the query by analyzing the context and structure of the sentence. This allows the system to expand the scope of its index beyond the superficial keywords entered by the user. Instead, it incorporates semantic intent to expand the index more broadly, capturing the user's potential search needs.

[0084] Next, in step S3.2, the system intelligently matches the user's query with the semantics of the document content based on semantic analysis. Through a deep semantic understanding model, the system can identify portions of the document that are semantically similar to the user's query, even if the keywords in those documents don't exactly match the user's query. This semantic-level optimization allows the system to more accurately retrieve documents that meet the user's actual needs, rather than just those with the same keywords. This semantic matching can significantly improve retrieval accuracy, especially when users use complex queries.

[0085] Finally, in step S3.3, the system continuously optimizes the semantic matching model and index structure based on the click-through rate and user dwell time of the search results. Click-through rate and dwell time are important behavioral metrics for measuring document relevance. By analyzing this feedback in real time, the system dynamically adjusts the weights and parameters of the semantic matching model, prioritizing documents with long user dwell times and high click-through rates. Simultaneously, the system also optimizes the index structure based on this behavioral feedback, ensuring that future search results are more aligned with user interests and needs.

[0086] Through index update and resource allocation optimization strategies, we ensure that the system can efficiently handle the addition and modification of new documents, while improving the performance of large-scale retrieval.

[0087] In step S4.1, the system updates the index in real time when new resources are added to the library or existing resources are modified. This process is accomplished through an automated index update mechanism, ensuring that the content of each document remains up to date in the index. Furthermore, the system adjusts the index structure based on content classification and user behavior data, prioritizing commonly used and frequently accessed documents at the top of the index to improve retrieval speed and user experience.

[0088] In step S4.2, the system uses distributed database technology to distribute and store index data across multiple nodes. This architectural design supports large-scale parallel search operations. When users initiate a large number of search requests simultaneously, the system can fully utilize the advantages of a distributed database to schedule multiple nodes to work together, ensuring load balancing of search tasks and avoiding performance bottlenecks on a single node.

[0089] Finally, in step S4.3, the system dynamically allocates resources and intelligently schedules search nodes based on the current query volume and load. When the system detects that certain nodes are overloaded, it automatically assigns some query tasks to other, less-loaded nodes, thereby optimizing overall system performance. This dynamic resource allocation mechanism not only ensures system efficiency under high concurrency conditions but also optimizes system response time, ensuring users receive search results quickly.

[0090] Example 1:

[0091] Combined with attachment Figure 2 Flowchart. A large digital library holds over five million academic papers and documents, covering a wide range of disciplines, including physics, chemistry, biology, and computer science. To enhance the user search experience, the library decided to adopt a dynamic index optimization method. This method optimizes the index structure based on different user queries, enabling rapid retrieval of the most relevant documents.

[0092] Imagine a user searches for the topic "Applications of Artificial Intelligence in Healthcare" in a library. The library's indexing system analyzes and dynamically sorts all relevant documents, ensuring that highly relevant and detailed documents are prioritized in the search results.

[0093] First, the system uses content analysis and classification algorithms to conduct a multi-dimensional analysis of digital resources in the library. For example, for the literature related to the topic "Application of Artificial Intelligence in Healthcare", the system will analyze it from three dimensions:

[0094] 1. Topic relevance: frequency and importance of keywords such as “artificial intelligence,” “medical care,” “diagnostic system,” and “machine learning.”

[0095] 2. Content details: Whether the literature contains detailed algorithm descriptions, specific cases, experimental data, research methods, etc.

[0096] 3. Semantic association: The system will understand the context of the document through natural language processing technology and analyze the potential semantic relationships in the document to determine which documents have content closely related to the topic, even though they may not have direct keyword matches.

[0097] The analysis results will construct a preliminary hierarchical structure of the literature and perform preliminary calculations of the dynamic index based on S(x) in the formula.

[0098] Next, based on the previous analysis, the system will use the following formula to calculate the dynamic index score of the document:

[0099]

[0100] For example, if a user searches for "application of artificial intelligence in healthcare," the relevance and content analysis results for a document x1 are as follows:

[0101] 1. Topic relevance function f1(x1): This function reflects the degree of match between the document and the topics of "artificial intelligence" and "medical care." Assume that this document contains a large amount of discussion on "artificial intelligence applied to medical diagnostic systems," and its topic relevance is 0.85.

[0102] 2. Content detail function f2(x1): The document contains detailed machine learning model descriptions and clinical trial data, with high detail richness. Assume that the detail score is 0.75.

[0103] 3. Adjustment coefficient α: Since libraries prioritize documents with rich details, assume that α is 1.2, indicating the system's preference for content details.

[0104] 4. Time dynamic adjustment coefficient β(t): Assuming that a user has searched for similar topics multiple times in the past month, the system will dynamically increase the priority of documents related to the user's historical queries. Assume that the initial value of β(t) is 1.1 and gradually increases over time.

[0105] Substitute these data into the formula to calculate the index score of the document. Assuming the time window T is 30 days, then:

[0106]

[0107] The calculation shows that:

[0108]

[0109] S(x1)=1.84×30=55.2

[0110] Therefore, the dynamic index score of document x1 is 55.2.

[0111] To ensure the system can quickly respond to user queries, it regularly adjusts its index structure based on user usage frequency and search habits. For example, suppose a user recently clicked on documents related to "AI medical applications" and spent significant time on them. This behavioral data will be fed back to the system, which will further prioritize these documents. For example, if document x1 is accessed multiple times and has a high click-through rate, the system will prioritize it higher in the next query, optimizing the user's search experience.

[0112] In this example, the index score of document x1 is 55.2. Based on this score, the system places it in a higher priority position, ensuring that users can quickly find this highly relevant and detailed document when searching for "Application of artificial intelligence in medicine".

[0113] It is assumed that user behavior monitoring is used to further optimize the index position of documents.

[0114] If a user frequently searches for literature related to "Applications of Artificial Intelligence in Healthcare" in the library, the system will monitor the user's query frequency, access time, and browsing path between documents in real time. Based on this behavioral data, the system will dynamically adjust the indexing priority of the documents using a frequency-aware model.

[0115] In the previous search, document x1 had a dynamic index score of 55.2, and the system prioritized its display to users based on this score. Now, the system will further incorporate user behavior feedback to adjust the index weights of document x1 and other related documents, ensuring that documents that users access more frequently are displayed more quickly in future searches.

[0116] Suppose a user recently searched for the topic "Applications of Artificial Intelligence in Healthcare" multiple times within a month, and the system recorded their visits to documents x1, x2, and x3. Specifically, the user clicked on document x1 10 times, document x2 7 times, and document x3 5 times. Furthermore, the user spent a long time on document x, indicating the document's high relevance.

[0117] In order to dynamically adjust the priority of documents based on user behavior, the system uses a frequency-aware model to calculate the weight of documents. The formula is:

[0118]

[0119] Assume that λ is a tuning parameter that controls the frequency sensitivity of user behavior, typically ranging from 0.1 to 0.5. Assume that the system is set to λ = 0.3 to moderately respond to changes in user behavior. Also, t represents the current query time, for example, set to 30 days.

[0120] For document x1, its access frequency p1 is 10 times, the access frequency p2 of document x2 is 7 times, and the access frequency p3 of document x3 is 5 times. Substituting these data into the formula:

[0121]

[0122] Through calculation, we can see that since the query time t is relatively short, the exponential part e in the formula -λt It will quickly approach 0, causing the index weight to be almost equal to the access frequency p i Therefore, the index weight of document x1 is 10, that of document x2 is 7, and that of document x3 is 5.

[0123] Based on these weighted results, the system adjusts the index position of the documents. In the future, when users perform similar searches, document x1 will be displayed first due to its highest weight, and documents x2 and x3 will be displayed in descending order of weight.

[0124] Suppose a new user searches for the topic "Applications of Artificial Intelligence in Healthcare." The system automatically ranks document x1 at the top of the search results. Because previous user data indicates that the document is highly relevant and rich in content, the new user can quickly find the most useful documents, improving the overall search experience. As more and more user behavior is monitored and fed back into the system, the system continuously adjusts the index weight, maintaining dynamic updates and optimizations, ensuring that frequently accessed documents are always prioritized.

[0125] In the previous calculations, documents x1, x2, and x3 already received high index weights based on user feedback. However, in this scenario, the user wants to search for documents with greater semantic depth, such as "specific applications of artificial intelligence in medical diagnosis and treatment." This requires the system to not only match keywords but also understand the semantic intent behind the document content. Therefore, the system uses semantic association analysis technology, combining the user's query with the document content, to dynamically calculate the document's semantic hierarchy score.

[0126] The system uses the following formula to optimize the semantic level matching of documents:

[0127]

[0128] in:

[0129] 1. L(s) is the semantic level score of the document, which measures the degree of semantic matching of the document.

[0130] 2. g1(s) is the basic semantic matching function, which represents the basic semantic relevance between the document and the user's query topic. For example, the number and importance of keywords such as "artificial intelligence," "medical care," and "diagnosis" in the document.

[0131] 3. g2(s) is a semantic expansion function responsible for expanding matching semantics based on context and identifying potentially relevant document content.

[0132] 4. γ is the semantic expansion coefficient, ranging from 0.5 to 1.5, which is used to adjust the impact of semantic expansion on the final matching results. For example, when γ is large, semantic expansion has a greater influence on document matching, and the system will be more inclined to match documents with expanded semantics.

[0133] 5. n is the detail complexity of the document, which measures the complexity of the detailed description at the semantic level. A higher n value indicates that the document contains more detailed technical descriptions and experimental data.

[0134] Assume that the basic semantic matching function g1(s) for document x1 is 0.8, indicating that the document contains a large amount of basic content related to "artificial intelligence" and "medical care." The semantic expansion function g2(s) for document x1 is 0.6, indicating that the document contains a large amount of potentially relevant content and may be combined with other research. Assume that the semantic expansion coefficient γ is 1.2, indicating that the expanded semantics have a strong influence on the document matching results. The document's detail complexity n is 10, indicating that the document contains a large amount of detailed descriptions.

[0135] Substitute into the formula to calculate:

[0136]

[0137] L(x1)=1.52·0.316=0.48

[0138] Through this calculation, the semantic level score of document x1 is 0.48.

[0139] Similarly, assume that the basic semantic matching function g1(s) of document x2 is 0.6, the semantic expansion function g2(s) is 0.7, the detail complexity n is 8, and the semantic expansion coefficient γ is 1.0. Substitute the formula into the calculation:

[0140]

[0141] L(x2)=1.3·0.353=0.46

[0142] The semantic level score of document x2 is 0.46.

[0143] Assume that the basic semantic matching function g1(s) of document x3 is 0.7, the semantic expansion function g2(s) is 0.4, the detail complexity n is 12, and the semantic expansion coefficient γ is 0.9. Substitute into the formula for calculation:

[0144]

[0145] L(x3)=1.06·0.288=0.31

[0146] The semantic level score of document x3 is 0.31.

[0147] According to the calculation results, the semantic hierarchy score for document x1 is 0.48, for document x2 is 0.46, and for document x3 is 0.31. The system will dynamically adjust the order of document display based on these scores. When a user queries "Specific applications of artificial intelligence in medical diagnosis and treatment," documents x1 and x2 will be prioritized due to their high semantic match scores. Especially with the support of semantic expansion, these documents can provide richer and more relevant content.

[0148] Through semantic association analysis technology, the system can more intelligently understand the intent behind user queries and perform in-depth matching based on the basic and extended semantics of the document. Document x1, due to its rich content and high semantic match, ranks high in the search results. However, despite having strong basic semantics, document x3, due to its weaker extended semantics and greater detail complexity, ultimately scores lower and is therefore placed lower in the results.

[0149] Example 2:

[0150] Combined with attachment Figure 3 In the flowchart, a user accesses documents related to "machine learning" and "NLP" (natural language processing) multiple times over a period of time. The system uses logs to track each query. For example, when a user searches for "machine learning applied to NLP," they access documents x1, x2, and x3. The system stores this query data in the log and uses cookies to track the user's subsequent behavior.

[0151] The system records the following behavioral data:

[0152] Document x1: The user visited 8 times, with each visit lasting 15 minutes.

[0153] Document x2: The user visited the document 5 times, with each visit lasting 10 minutes.

[0154] Document x3: The user visited the document 3 times, with each visit lasting 20 minutes.

[0155] In order to analyze the user's search behavior, the system substitutes these behavioral data into the following formula for calculation:

[0156]

[0157] in:

[0158] B(x) represents the comprehensive score of user behavior, which measures the user's interest in a certain document.

[0159] f1(x) is the access frequency function, which indicates the number of times document x is accessed by users.

[0160] f2(x) is the residence time function, which reflects the time the user stays on the document.

[0161] ω1 and ω2 are weight factors of access frequency and residence time, respectively. It is assumed that the system sets the weights of access frequency and residence time to ω1 = 0.7 and ω2 = 0.9, respectively.

[0162] ψ(t) is a time-dynamic adjustment function. Assume that the initial value of ψ(t) is 1.05 and gradually increases over time, indicating that the system gives higher priority to the most recent behavior.

[0163] T is the behavior time window, which is assumed to be 30 days.

[0164] Substitute actual data for calculation:

[0165] 1. Document x1:

[0166] Visit frequency: 8 times.

[0167] Duration of each stay: 15 minutes.

[0168] Weight factors: ω1=0.7, ω2=0.9.

[0169] Dynamic adjustment coefficient: ψ(t) = 1.05.

[0170] Substituting this data into the formula:

[0171]

[0172] The calculation shows that:

[0173]

[0174] B(x1)=19.775×30=593.25

[0175] 2. Documents x2:

[0176] Visit frequency: 5 times.

[0177] Duration of each session: 10 minutes.

[0178] Weight factors: ω1=0.7, ω2=0.9.

[0179] Dynamic adjustment coefficient: ψ(t) = 1.05.

[0180] Substitute the data into the formula:

[0181]

[0182] B(x2)=12.95×30=388.5

[0183] 3. Documents x3:

[0184] Visit frequency: 3 times.

[0185] Duration of each stay: 20 minutes.

[0186] Weight factors: ω1=0.7, ω2=0.9.

[0187] Dynamic adjustment coefficient: ψ(t) = 1.05.

[0188] Substitute the data into the formula:

[0189]

[0190] B(x3)=21×30=630

[0191] According to the above calculations, document x3 has the highest comprehensive user behavior score, at 630. This indicates that although users visit x3 less frequently, the system considers it highly relevant due to its longer visit time. Document x1 scored 593.25, indicating that it is also highly attractive to users, primarily due to its high visit frequency. Document x2 scored relatively low at 388.5, reflecting slightly lower user appeal.

[0192] Through this multi-dimensional modeling and calculation, the system can dynamically adjust the priority of documents in the index. When a user searches for content related to "Applications of Machine Learning in Natural Language Processing," document x3 will be displayed first because it has the highest overall user behavior score, despite its lower access frequency. Document x1 will also be displayed with higher priority due to its higher access frequency.

[0193] In the previous scenario, the user has repeatedly searched for literature related to "Applications of Machine Learning in Natural Language Processing." The system records the user's query behavior data through log tracking and cookie technology. This data includes the document categories accessed, frequency of access, query type, and access path. Now, the system needs to further analyze this data, using a clustering analysis algorithm to identify typical user behavior patterns and then using a recurrent neural network (RNN) model to predict the user's future search needs.

[0194] First, the system uses the KMeans or DBSCAN clustering algorithm to cluster the user's historical behavior data. Suppose that through analysis, the system identifies two typical user usage patterns:

[0195] Mode A: Frequently searches for literature related to "Basics of Machine Learning" and "Applications of NLP". The user repeatedly stops at literature x1, x2, and x3, and the path is relatively fixed.

[0196] Mode B: Cross-domain queries are frequently conducted, with query topics expanding from “deep learning” to “neural networks” and paths being more scattered.

[0197] The cluster analysis results show that the user belongs to mode A, that is, the user mainly focuses on literature on machine learning and natural language processing, which means that the user is likely to continue to query content related to these topics in future searches.

[0198] Based on the results of the cluster analysis, the system then uses a recurrent neural network (RNN) model to predict the user's future search needs. RNN is a deep learning model for processing sequential data, capable of learning and capturing the temporal dependencies in user behavior sequences. The model predicts the user's future search needs using the following formula:

[0199]

[0200] Among them, P(u) is the prediction of the user's future search needs, u i Indicates the i-th retrieval behavior, f3(u i ) is the user behavior sequence feature function, which describes the user's past search behavior pattern. λ is the adjustment parameter that controls the prediction sensitivity, ranging from 0.1 to 0.5, indicating the model's sensitivity to recent behavior. i Indicates the time when the behavior occurs, in days.

[0201] Assume that the user's search behavior in the past 30 days is:

[0202] Day 1: Search for literature related to "Basics of Machine Learning", behavior u1.

[0203] Day 7: Query "Deep Learning Applications in NLP", behavior u2.

[0204] Day 15: Search for "Application of Machine Learning in Healthcare", behavior u3.

[0205] Day 20: Query "Recent Advances in Natural Language Processing", behavior u4.

[0206] The corresponding behavior sequence characteristic function f3(u i) can be expressed as features related to document content and query path. Assume f3(u1) = 0.8, f3(u2) = 0.7, f3(u3) = 0.9, f3(u4) = 0.85, and set λ = 0.3.

[0207] We substitute this data into the formula to predict the user's future search needs:

[0208] 1. Day 1 behavior u1:

[0209]

[0210] 2. Day 7 behavior u2:

[0211]

[0212] 3. Day 15 behavior u3:

[0213]

[0214] 4. Day 20 behavior u4:

[0215]

[0216] Based on these calculations, the user's recent behavior u1 has the greatest impact on future search requests, followed by behavior u2, while more recent behaviors u3 and u4 have less impact on the prediction. Therefore, the system can conclude that the user's future search topics will continue to focus on content related to "Machine Learning Basics" and "Deep Learning Applications in NLP."

[0217] Based on the RNN model's predictions, the system prioritizes recommendations related to "Machine Learning Fundamentals" and "NLP Applications" when a user re-enters the library system, specifically those documents x1 and x2 that are most relevant to their previous search. This not only improves search accuracy but also saves users significant time by avoiding repeated searches for highly relevant content they've already searched for. The system also recommends a small number of additional documents, such as the latest literature on "Deep Learning" or "Neural Networks," to meet users' potential cross-disciplinary interests.

[0218] By identifying user behavior patterns through cluster analysis and using the RNN model to predict users' future search needs, the system can dynamically adjust indexing and recommendation strategies to provide users with more accurate literature recommendations.

[0219] In the preceding scenario, based on the calculations of the recurrent neural network (RNN) prediction model, the system predicted that the user's future search needs would focus on literature related to "Machine Learning Fundamentals" and "NLP Applications," specifically documents x1 and x2. To optimize the user's search experience, the system pre-loaded the index data for these relevant documents, ensuring that users could instantly retrieve highly relevant documents when searching. Furthermore, the system dynamically optimized the index order based on actual user feedback (such as query frequency and path click-through rate), further improving search efficiency.

[0220] Based on the previously calculated prediction result P(u), documents x1 and x2 are predicted to be the documents most likely to be retrieved by the user in the future. Therefore, the system uses an adaptive preloading mechanism to preload the indexes for these documents. For example, if document x1 has a high overall score (e.g., a score of 593.25 in the behavioral model), the system will prioritize loading its index. Document x2, with a slightly lower score (e.g., 388.5), will also be preloaded. This mechanism enables the system to quickly respond to user queries and reduce loading delays.

[0221] When the user begins to perform actual search behavior, the system will dynamically optimize the index sorting based on the user's query frequency and path click rate. The optimization process is calculated using the following formula:

[0222]

[0223] in:

[0224] S(y) represents the priority ranking score of the document in the index.

[0225] r1(y i ) is the query frequency function, indicating that the user queries document y i times.

[0226] r2(y i ) is the path click rate function, which indicates the frequency of users clicking on the document.

[0227] γ1 and γ2 are adjustment coefficients, which control the impact of query frequency and path click rate on document priority. Assume that γ1 = 0.6 and γ2 = 0.8.

[0228] Assume that in the next week, the user searches for documents x1, x2, and x3 multiple times:

[0229] Document x1: query frequency r1(y1)=10, path click rate r2(y1)=8.

[0230] Document x2: query frequency r1(y2)=7, path click rate r2(y2)=5.

[0231] Document x3: query frequency r1(y3)=4, path click rate r2(y3)=3.

[0232] Based on these data, the system calculates a priority ranking score for each document:

[0233] 1. Document x1:

[0234]

[0235] 2. Documents x2:

[0236]

[0237] 3. Documents x3:

[0238]

[0239] Based on the above calculations, the system determines that document x1 has the highest priority ranking score of 12.4; document x2 scores 8.2; and document x3 scores 4.8. Therefore, the system dynamically adjusts the index order. The next time a user searches for a related topic, document x1 will be displayed first, followed by document x2, and document x3 will be ranked later. Through this dynamic optimization process, the system ensures that documents with high query frequency and high path click-through rates receive higher priority in the index, improving the relevance and efficiency of document retrieval.

[0240] In this example, the system pre-loaded the document index that best matched the user's needs based on the prediction results. It also dynamically adjusted the document priority ranking based on the user's actual search behavior. Document x1, due to its high query frequency and click-through rate, maintained a high priority in the index, while documents x2 and x3 were given relatively low priorities.

[0241] Example 3:

[0242] Combined with attachment Figure 3 Flowchart: Suppose a user searches for "Applications of deep learning in natural language processing" in the library. The system needs to identify the two core keywords "deep learning" and "natural language processing" and understand the user's specific needs based on the query context. The system then uses a deep learning natural language processing model to parse the query and calculate the query context semantic vector based on a formula.

[0243] The core formula of semantic parsing is:

[0244]

[0245] Among them, V(q) represents the context semantic vector of the query statement, which is used to quantify the semantic content of the user query; f1(w i ) is the keyword semantic weight function, which represents the i-th keyword w in the query i The weight in the semantic space; f2(c i ) is the context-dependent function, indicating that the keyword w i Its context c i The semantic relationship between words and context; α and β are adjustment coefficients, controlling the weight of keywords and context, respectively. Assume that α = 0.7, which means the system has a higher weight for keywords, and β = 0.6, which means that context is less important to semantics.

[0246] When a user searches for "Application of deep learning in natural language processing," the system identifies keywords w1 as "deep learning" and w2 as "natural language processing." Using the natural language processing model, the system performs a semantic weighting analysis on these keywords:

[0247] 1. Keyword w1 = "deep learning": System analysis shows that its semantic weight f1(w1) = 0.85 because the word highly matches the query topic.

[0248] 2. Keyword w2 = "natural language processing": semantic weight f1(w2) = 0.9, which is also the core word in the query topic.

[0249] 3. Contextual relevance function f2(c1): Assuming that “deep learning” is related to words such as “application” and “algorithm” in the context of the query, the given contextual relevance value is f2(c1) = 0.75.

[0250] 4. Context association function f2(c2): For “natural language processing”, context words such as “technology”, “data”, etc. are identified as highly relevant words, f2(2)=0.8.

[0251] In addition, the system sets a time dynamic adjustment function γ(t) that changes over time, increasing from 1.0 to 1.05, indicating that the system gradually adjusts the priority of query semantics according to the time when the query occurs.

[0252] Substituting these data into the formula for calculation, first calculate the semantic vectors of keywords w1 and w2:

[0253] 1. Calculation of semantic vector of keyword w1:

[0254]

[0255]

[0256] Assuming that the time window T for semantic analysis is 30 days, then:

[0257] V(w1)=1.0675×30=32.025

[0258] 2. Calculation of semantic vector of keyword w2:

[0259]

[0260] Assuming the time window is 30 days, then:

[0261] V(w2)=1.134×30=34.02

[0262] Through the above calculations, the system obtained a semantic vector of V(w1) = 32.025 for the keyword "deep learning," and a semantic vector of V(w2) = 34.02 for "natural language processing." These vectors are used to measure the semantic match between the query and relevant documents, helping the system better understand the user's query intent. The system will prioritize documents with a high relevance to these keywords and combine context for more accurate matching.

[0263] Based on the user's query semantic vector, the system can identify that the user's interest in "specific applications of deep learning in natural language processing" lies not only in the underlying principles of the technology but also in its specific implementation in practical applications. Because "natural language processing" and "deep learning" are at the core of the query and are highly correlated with the "application" in the context, the system prioritizes research findings and application cases in related fields. For example, literature related to "deep learning algorithms" or papers on "natural language processing applications" will be ranked high in the search results due to their high semantic match.

[0264] By using a deep learning-based natural language processing model, the system not only identifies the core keywords in user queries but also understands the user's underlying semantic intent through contextual association. By calculating the query's contextual semantic vector, the system can more accurately understand the user's needs and dynamically adjust the document retrieval results based on the query's semantics.

[0265] Suppose a user continues searching for "Applications of Deep Learning in Natural Language Processing" in a digital library. The system not only needs to identify the keywords in the query but also uses semantic intent recognition technology to distinguish the user's query type. This query might be an "information query," meaning the user wants to learn specific information or technical principles in a particular field, or a "transaction query," meaning the user wants to find the specific implementation or application of a particular technology. To better display documents that meet the user's needs, the system uses an intent classification model to hierarchically process the query and calculate the document's priority using a formula.

[0266] The system first uses semantic intent recognition technology to classify the user's query "Applications of deep learning in natural language processing" as an "information query," as the user likely primarily seeks information about the "basic principles and applications of deep learning in NLP." Meanwhile, if the query includes a clear goal or application scenario, it might be classified as a "transactional query," such as "How to use deep learning models to process NLP data."

[0267] To build a multi-level index based on the user's query intent, the system uses the following formula to calculate the hierarchical score of the intent:

[0268]

[0269] Among them, I(s) is the hierarchical score of the user's query intention, which quantifies the priority of the user's query intention; g i (x) is the score of document x in different intent levels, reflecting the semantic matching degree between the document and the user query. The system will i The query priority is adjusted by the time sensitivity parameter λ. Assume that λ ranges from 0.1 to 0.5, which determines how quickly the query intent changes over time.

[0270] To better understand this process, assume that the user's query behavior occurred in the past 30 days:

[0271] Day 5: Search “Basic Concepts of Deep Learning” (Information Search).

[0272] Day 10: Query “Common Algorithms in Natural Language Processing” (Information Query).

[0273] Day 20: Query “How to apply deep learning models in NLP” (Transactional Query).

[0274] For these queries, the system will calculate the scores of relevant documents at different intent levels. Suppose document x1 is about "deep learning foundations", document x2 is about "natural language processing algorithms", and document x3 is about "deep learning in NLP applications". We set:

[0275] The semantic level score of document x1 is g1(x1)=0.9.

[0276] The semantic level score of document x2 is g1(x2)=0.85.

[0277] The semantic level score of document x3 is g1(x3)=0.95.

[0278] The system calculates different query behaviors of users, assuming time sensitivity λ = 0.3:

[0279] 1. Day 5 query "Basic concepts of deep learning":

[0280]

[0281] 2. Day 10 query "Common algorithms in natural language processing":

[0282]

[0283] 3. Day 20 query: “How to apply deep learning models in NLP”:

[0284]

[0285] Based on the above calculations, document x1 most closely matches the user's "information query" and is more recent, resulting in a higher hierarchy score of approximately 0.1644. Therefore, when the user performs a related query again, this document will be prioritized. Document x2, with a hierarchy score of 0.0403, is highly relevant, but due to the time lapse, it has a lower display priority. Document x3, a matching document for the transaction query, has a hierarchy score of 0.0024, indicating it has the lowest priority and may be ranked lower.

[0286] By building a multi-level index, the system can dynamically adjust the priority of documents based on the user's query intent and query time, ensuring that the results the user receives are more in line with their current needs. For example, if a user has recently searched for information documents multiple times, the system will prioritize displaying relevant basic documents over application or transactional documents.

[0287] By using an intent classification model, the system builds a multi-level index based on the user's query type (information query, navigation query, or transaction query) and dynamically adjusts the priority display of documents. By calculating the scores of documents at different levels, the system can more accurately understand the user's intent and provide the documents that best match their needs.

[0288] If a user searches for "deep learning applications in natural language processing" in a digital library, the system should not only display directly matching documents but also, through semantic expansion, find other documents related to the topic that don't directly contain the keywords "deep learning" or "natural language processing." For example, a document might discuss "the role of neural networks in language modeling" or "the application of machine learning in text processing." These documents are semantically relevant to the query but don't use the exact keywords.

[0289] To optimize search results through semantic expansion, the system uses the following formula to calculate the expanded recommendation result set:

[0290]

[0291] in:

[0292] R(q) is the recommended retrieval result set after semantic expansion.

[0293] h1(y i ) is a direct keyword matching result function, indicating that document y i The score of a direct match with query q.

[0294] h2(y i ) is the semantic expansion result function, which means document y i The relevance score to the query q at the semantic level.

[0295] μ is the semantic expansion coefficient, ranging from 0.5 to 2.0, which controls the degree of influence of semantic expansion results on the final recommendation. Assume that μ = 1.2 here, it means that semantic expansion has a greater impact on the search results.

[0296] n is the total number of retrieved documents.

[0297] Assume that when a user searches for "Application of deep learning in natural language processing", the system identifies documents directly related to the query, including y1, y2, and y3:

[0298] The keywords of document y1 completely match the query, and the direct score is h1(y1)=0.95.

[0299] Document y2 discusses “Application of Neural Networks in Language Processing”, which is semantically relevant to the query but has different keywords. Its expansion score h2(y2)=0.8.

[0300] Document y3 mainly discusses the “role of machine learning in text processing” and has a certain semantic relevance, with an expansion score h2(y3) = 0.75.

[0301] According to the formula, the system combines the direct matching scores of these documents with the semantic expansion scores to calculate the expanded recommendation result set:

[0302] 1. The expanded recommendation score of document y1:

[0303]

[0304] Assuming the total number of documents n = 100, then:

[0305]

[0306] 2. The expanded recommendation score of document y2:

[0307]

[0308] 3. Recommendation score after expansion of document y3:

[0309]

[0310] Through calculation, the system found that although document y2 did not directly match the query keywords, its semantic expansion score was high, resulting in a final recommendation score of R(y2) = 0.0096, exceeding the recommendation score of R(y1) = 0.0095 for the directly matched document y1. Document y3's recommendation score was 0.009, which, while lower than the previous two, still had a high relevance. Therefore, when a user queries, the system will first recommend documents y2 and y1, and then recommend document y3.

[0311] Through this dynamic semantic expansion mechanism, the system not only displays documents that directly match the query keywords, but also discovers documents that are semantically related but not an exact match. This significantly enriches users' search results, ensuring they discover more valuable and relevant documents. For example, although document y2 does not directly contain the keyword "deep learning," it discusses "Applications of Neural Networks in Natural Language Processing" in detail, making this document equally important for users to understand the full scope of this field.

[0312] In this example, the system leverages a dynamic semantic expansion mechanism and knowledge graph to expand the scope of search results based on the semantic intent of the user's query. Documents are not only displayed in search results based on keyword matching but are also prioritized based on their semantic expansion scores. By adjusting the expansion coefficient μ and temporal characteristics, the system ensures that search results not only cover the user's immediate needs but also provide more potentially relevant content. This mechanism significantly enhances the breadth and depth of digital library searches, ensuring that users receive more comprehensive documentation support when conducting complex or multi-level queries.

[0313] Example 4:

[0314] If a user searches for "deep learning applications in natural language processing" in a digital library, the system should not only display directly matching documents but also, through semantic expansion, find other documents related to the topic that don't directly contain the keywords "deep learning" or "natural language processing." For example, a document might discuss "the role of neural networks in language modeling" or "the application of machine learning in text processing." These documents are semantically relevant to the query but don't use the exact keywords.

[0315] To optimize search results through semantic expansion, the system uses the following formula to calculate the expanded recommendation result set:

[0316]

[0317] in:

[0318] R(q) is the recommended retrieval result set after semantic expansion.

[0319] h1(y i ) is a direct keyword matching result function, indicating that document y i The score of a direct match with query q.

[0320] h2(y i ) is the semantic expansion result function, which means document y i The relevance score to the query q at the semantic level.

[0321] μ is the semantic expansion coefficient, ranging from 0.5 to 2.0, which controls the degree of influence of semantic expansion results on the final recommendation. Assume that μ = 1.2 here, it means that semantic expansion has a greater impact on the search results.

[0322] n is the total number of retrieved documents.

[0323] Assume that when a user searches for "Application of deep learning in natural language processing", the system identifies documents directly related to the query, including y1, y2, and y3:

[0324] The keywords of document y1 completely match the query, and the direct score is h1(y1)=0.95.

[0325] Document y2 discusses “Application of Neural Networks in Language Processing”, which is semantically relevant to the query but has different keywords. Its expansion score h2(y2)=0.8.

[0326] Document y3 mainly discusses the “role of machine learning in text processing” and has a certain semantic relevance, with an expansion score h2(y3) = 0.75.

[0327] According to the formula, the system combines the direct matching scores of these documents with the semantic expansion scores to calculate the expanded recommendation result set:

[0328] 1. The expanded recommendation score of document y1:

[0329]

[0330] Assuming the total number of documents n = 100, then:

[0331]

[0332] 2. The expanded recommendation score of document y2:

[0333]

[0334] 3. Recommendation score after expansion of document y3:

[0335]

[0336] Through calculation, the system found that although document y2 did not directly match the query keywords, its semantic expansion score was high, resulting in a final recommendation score of R(y2) = 0.0096, exceeding the recommendation score of R(y1) = 0.0095 for the directly matched document y1. Document y3's recommendation score was 0.009, which, while lower than the previous two, still had a high relevance. Therefore, when a user queries, the system will first recommend documents y2 and y1, and then recommend document y3.

[0337] Through this dynamic semantic expansion mechanism, the system not only displays documents that directly match the query keywords, but also discovers documents that are semantically related but not an exact match. This significantly enriches users' search results, ensuring they discover more valuable and relevant documents. For example, although document y2 does not directly contain the keyword "deep learning," it discusses "Applications of Neural Networks in Natural Language Processing" in detail, making this document equally important for users to understand the full scope of this field.

[0338] In this example, the system leverages a dynamic semantic expansion mechanism and knowledge graph to expand the scope of search results based on the semantic intent of the user's query. Documents are not only displayed in the search results based on keyword matching but are also prioritized based on their semantic expansion scores. By adjusting the expansion coefficient μ and temporal characteristics, the system ensures that search results not only cover the user's immediate needs but also provide more potentially relevant content.

[0339] Example 5:

[0340] Combined with attachment Figure 5 Flowchart: Suppose a user searches for "Applications of Deep Learning in Natural Language Processing" in a digital library. The system adjusts the semantic priority and index structure of documents based on the user's click behavior and dwell time in the search results. The user clicks on three documents, v1, v2, and v3, and dwells on each for a different amount of time.

[0341] The system first collects the user's click behavior and dwell time in the search results, and uses the formula to calculate the semantic priority score of the document. The formula is as follows:

[0342]

[0343] in:

[0344] P(u) represents the semantic priority score of document u.

[0345] C u(t) is the change of the click rate of the document over time, which represents the user's click behavior at time t.

[0346] S u (t) is the residence time function of the document, reflecting the time the user stays on the document page.

[0347] α = 0.6 and β = 0.8 are coefficients that control the importance of click-through rate and dwell time, and λ = 0.05 is the time decay factor.

[0348] Assume that in the past 30 days, the user's behavior data for documents v1, v2, and v3 are as follows:

[0349] Document v1: Click-through rate Dwell time minute.

[0350] Document v2: Click-through rate Dwell time minute.

[0351] Literature v3: Click-through rate Dwell time minute.

[0352] Substitute these data into the formula to calculate the semantic priority score of the document.

[0353] 1. Semantic priority of document v1:

[0354]

[0355] Assume t = 30 days:

[0356] P(v1)≈0.6·2.77+0.8·10·e -1.5 ≈1.662+0.8·10·0.223≈1.662+1.784=3.446

[0357] 2. Semantic priority of document v2:

[0358]

[0359] The calculation shows that:

[0360] P(v2)≈0.6·2.2+0.8·5·0.223≈1.32+0.892=2.212

[0361] 3. Semantic priority of Document v3:

[0362]

[0363] The calculation shows that:

[0364] P(v3)≈0.6·2.56+0.8·7·0.223≈1.536+1.249=2.785

[0365] S2: Dynamically adjust the semantic model and index structure of documents

[0366] The system inputs these behavioral feedback into the semantic matching model, uses deep learning algorithms to analyze long-term behavior, and optimizes the position of documents in the semantic space. At the same time, the system uses reinforcement learning algorithms to dynamically adjust the indexing priority of documents based on dwell time and click-through rate. The formula is as follows:

[0367]

[0368] in:

[0369] I(v) is the index priority score of document v.

[0370] S i (v) and C i (v) are the dwell time and click rate during the i-th retrieval, respectively.

[0371] μ=0.4 is the coefficient of the influence of control feedback on priority adjustment.

[0372] Based on the previous data, the indexing priority of documents v1, v2, and v3 is calculated as follows:

[0373] 1. Indexing priority of document v1:

[0374]

[0375] 2. Indexing priority of document v2:

[0376]

[0377] 3. Indexing priority of Document v3:

[0378]

[0379] Combined with user behavioral feedback, the system will also dynamically optimize the semantic expansion model. Based on the user's click and stay time, the system dynamically optimizes and expands the keyword pool. The formula is as follows:

[0380]

[0381] in:

[0382] E(q) is the semantic expansion score of query q.

[0383] f1(w i ) is the keyword w iThe expansion weight, f2(c i ) is a contextual semantic expansion function.

[0384] k1=0.7 and k2=0.5 are expansion parameters that control the relative influence of keywords and context.

[0385] t is the time when the query occurs.

[0386] Suppose a user queries "Applications of deep learning in natural language processing" and is related to the keywords "neural network" and "text processing." After optimizing the semantic expansion model, the system identifies relevant documents y1 and y2 and calculates their expansion scores.

[0387] 1. Document y1:

[0388]

[0389] 2. Document y2:

[0390]

[0391] The system dynamically adjusts the order in which documents are displayed by optimizing their semantic priority and indexing structure through behavioral feedback. For example, documents v1 and v3 have higher indexing priorities and will be prioritized in future search results. Furthermore, through a semantic expansion mechanism, the system identifies documents y1 and y2 that may be of interest to the user but are not directly matched, and recommends these documents to the user. This process ensures a more personalized and accurate search experience.

[0392] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The optimization method of index retrieval of digital library is characterized by include: S1. Dynamic hierarchical index construction technology: S1.

1. First, conduct a multi-dimensional content analysis of the digital resources in the library, including subject, keyword, author, and publication date information; classify the documents based on this information and construct a content hierarchy; S1.

2. Based on the classification, a dynamic index is constructed for different document levels, including topics, content details, and semantic associations. This allows frequently used content to be quickly retrieved, while detailed content is retained in the deep index. S1.

3. Regularly adjust the index structure based on user usage frequency and search habits to bring frequently used information closer to the surface and optimize search speed; S2. Index optimization strategy based on usage pattern: S2.

1. Analyze user search behavior data to monitor user access frequency, query types, and access paths in different document categories, and identify typical user usage patterns. S2.

2. Dynamically adjust the priorities of different index items based on user behavior monitoring results, and promote frequently accessed documents or related topics to the top of the search results; S2.

3. Locally cache high-frequency indexes based on usage patterns, quickly return results, and control and reduce resource usage. S3, semantic association retrieval optimization: S3.

1. Introducing natural language processing technology to perform semantic analysis on user queries, identify keywords, understand the underlying semantic intent, and expand the scope of indexing. S3.

2. Based on semantic analysis, intelligently match user queries with the semantics of document content, optimize the index, and enable retrieval of documents with similar semantics but different keywords. S3.

3. Optimize the semantic matching model and index structure based on the click-through rate of search results and user dwell time feedback; S4. Index update and resource allocation optimization: S4.

1. Update the index in real time when new literature resources are added to the library or existing resources are modified, and adjust the index structure based on content classification and user behavior; S4.

2. Use distributed database technology to store indexes on multiple nodes and support large-scale parallel retrieval; S4.

3. During the search process, resources are dynamically allocated and each search node is scheduled based on the current query volume and load. The method for constructing the dynamic index includes: S1. First, we use content analysis and classification algorithms to conduct a multi-dimensional analysis of digital resources in the library, including subject matter, content details, and semantic association information, and establish a preliminary structure of a dynamic index based on the document hierarchy. For each document, its index is calculated hierarchically using the formula: S(x)=∫0 T (f1(x)+α·f2(x)·β(t))dt Among them, S(x) represents the dynamic index score of the document, which is the core variable used to measure the importance of the document during the index optimization process; f1(x) is the subject relevance function of the document, which reflects the degree of match between the document content and a specific topic; f2(x) is the content detail function of the document, which represents the richness of the description of the document at the detail level; α is the adjustment coefficient of the content details, which is used to adjust the weight of the details in the overall index; β(t) is the time dynamic adjustment coefficient related to user interaction, which dynamically adjusts the index structure over time; T is the time period for index weight adjustment, which is used to represent the time window for calculating the index during the optimization process; S2. Secondly, by monitoring user behavior, including query frequency, access time, and user path, the frequency perception model is used to dynamically adjust the index position of the document. The index weight adjustment is calculated using the formula: Among them, W(u) is the result of index weight adjustment caused by user behavior, which represents the index priority after the user visits a specific document; p i is the access frequency of the i-th document, that is, the number of times the document is queried by users; λ is the adjustment parameter that controls the frequency sensitivity of user behavior and determines the intensity of the impact of access frequency on the document index weight; t represents the current query time, which adjusts the perception of document access frequency over time; S3. Finally, we use semantic association analysis technology to perform deep semantic matching on the documents. We use a semantic vector model based on deep learning to dynamically optimize the semantic level of different document contents. The semantic level is determined by the formula: Where L(s) represents the semantic hierarchy score of the document, which is the result of the hierarchical division of the document through semantic analysis; g1(s) is the basic semantic matching function, which represents the basic semantic relevance of the document; g2(s) is the semantic expansion function, which is responsible for matching the document in the semantic space; γ is the semantic expansion coefficient, which controls the impact of semantic expansion on the matching results; n is the detail complexity of the document, which is used to measure the complexity of the detailed description of the document at the semantic level; The method for identifying typical user usage patterns includes: S1. First, we use log tracking and cookie technology to record users' search behavior data in real time, including the document category, query type, access path, and dwell time of each visit. We then use the following formula to perform multi-dimensional modeling and optimization of user behavior data: B(x)=∫0 T (ω1·f1(x)+ω2·f2(x)·ψ(t))dt Among them, B(x) represents the comprehensive score of user behavior; f1(x) is the access frequency function, which represents the frequency of user access to a certain document; f2(x) is the residence time function, which reflects the length of time the user stays in the document; ω1 and ω2 are the weight factors of access frequency and residence time respectively; ψ(t) is the behavior dynamic adjustment function, which affects the index priority over time; T is the behavior time window for real-time monitoring; S2. Based on the collected behavioral data, cluster analysis algorithms such as K-Means or DBSCAN are used to classify users, identify typical user usage patterns, including users who frequently search for documents in a specific category or often follow similar paths, and mathematically model these patterns. Then, based on the user's historical behavior data, a deep learning model based on the recurrent neural network (RNN) is constructed to predict the user's future search needs through the following formula: Among them, P(u) is the prediction of the user's future behavior; u i Indicates the i-th retrieval behavior, f3(u i ) is the user behavior sequence feature function, which describes the user's retrieval behavior pattern in the past; λ is the adjustment parameter that controls the prediction sensitivity, t i Indicates the time when the behavior occurs; Based on the prediction results of P(u), the index of relevant documents is loaded in advance, and the adaptive preloading mechanism is used to optimize the retrieval efficiency. Then, according to the actual feedback of user behavior, the display order of the index is adjusted, and the dynamic sorting of the index is optimized using the formula: Among them, S(y) represents the priority ranking score of the document in the index; r1(y i ) and r2(y i ) are the user's query frequency and path click rate function for the document, γ1 and γ2 are adjustment coefficients, which control the influence of the two on the priority.

2. The method for optimizing index retrieval of a digital library according to claim 1, characterized in that The framework used to identify keywords and understand the semantic intent behind them is as follows: S1. First, use the deep learning natural language processing model to perform deep semantic analysis on the user query and calculate the context semantic vector of the user query using the following formula: V(q)=∫0 T (α·f1(w i )+β·f2(c i )·γ(t))dt Among them, V(q) represents the context semantic vector of the query statement, which is used to quantify the semantic content of the user query; f1(w i ) is the keyword semantic weight function, which represents the i-th keyword w in the query i The weight in the semantic space reflects the relevance between the keyword and the query topic; f2(c i ) is the context-dependent function, indicating that the keyword w i Its context c i The semantic relationship between them; α and β are adjustment coefficients used to balance the weights of keywords and context; γ(t) is a time dynamic adjustment function that represents the semantic characteristics of the query that change over time; T is the time window of semantic parsing, which represents the duration of semantic parsing; S2. Then, using semantic intent recognition technology, we classify user queries into different types through an intent classification model, including information queries, navigation queries, and transaction queries. We then use the following formula to construct a multi-level index for the intents: Among them, I(s) represents the hierarchical score of query intent, which is an indicator used to quantify user query intent; g i (x) is a semantic level function, which represents the score of document x in different intent levels and reflects the semantic matching degree between the document and the user query; λ is a parameter that controls the time sensitivity of the model and determines the speed at which the query intent changes over time; t i is the timestamp of the query; S3. Finally, through the dynamic semantic expansion mechanism, the knowledge graph or deep semantic model is used to expand keywords according to the semantic intent of the user's current query statement. The calculation formula is: Among them, R(q) is the recommended search result set after expansion, which represents the final search result after semantic expansion; h1(y i ) is a direct keyword matching result function, which represents the document y directly related to the query q i score; h2(y i ) is a semantic expansion result function, which represents the documents y that are semantically related to the query q but not directly matched by keywords. i score; μ is the semantic expansion coefficient, which controls the influence of the expansion result on the final recommendation; n is the total number of retrieved documents.

3. The method for optimizing index retrieval of a digital library according to claim 1, characterized in that The optimization of the semantic matching model and index structure includes the following steps: S1. First, collect the user's click behavior and dwell time in the search results in real time, associate the data with the semantic relevance of the document, and calculate the impact of user behavior feedback on the semantic priority of the document based on the formula: P(u)=∫0 T (α·log(1+C u (t))+β·S u (t)·e -λt )dt Among them, P(u) represents the semantic priority score of document u, C u (t) is the change of the click rate of the document over time t, S u (t) is the user's residence time function of the document, α and β are adjustment coefficients that control the relative importance of click-through rate and residence time, λ is the time decay factor, and T is the collection period of feedback data; S2. The behavioral feedback is then input into the semantic matching model to dynamically adjust the position of the document in the semantic space. A deep learning algorithm is used to analyze long-term behavior and optimize the parameters of the semantic model. Based on user dwell time and click-through rate, a reinforcement learning algorithm is used to dynamically adjust the index structure and update the index priority of the document using the formula: Among them, I(v) is the priority score of document v in the index, S i (v) is the residence time of the document during the i-th retrieval, C i (v) is the i-th click rate, μ is the coefficient of the control feedback on the priority adjustment, t i timestamp for the action; S3. Finally, we optimize the semantic expansion model based on behavioral feedback. We dynamically optimize and expand the keyword pool based on user clicks and dwell time, and use the formula to calculate personalized semantic expansion recommendations: Among them, E(q) is the semantic expansion score of query q, f1(w i ) is the keyword w i The expansion weight, f2(c i ) is the context semantic expansion function, k1 and k2 are expansion parameters that control the relative influence of keywords and context, and t is the time when the query occurs.

Images