A method for constructing a knowledge service platform monograph

Abstract

The application belongs to the technical field of knowledge product construction, and particularly relates to a knowledge service platform monograph construction method, which comprises the following steps: step one, document demand change sensing: analyzing user search logs of the knowledge service platform to sense changes in document demand of the user; step two, document demand content analysis: after judging that the user's document demand has changed, analyzing documents that have interacted with the knowledge service platform and the user within a demand change interval, mining core content of the user's document demand within the demand change interval, and providing topics and initial query words of the topics for monograph construction; step three, document query word expansion: performing semantic expansion on the initial query words of the topics to obtain an expanded query word sequence; and step four, monograph construction: performing semantic search on the knowledge service platform by using the multiple groups of expanded query word sequences of the topics to construct a monograph from obtained documents.

Description

Technical Field

[0001] This application belongs to the field of knowledge product construction technology, specifically involving a method for constructing a thematic collection of articles for a knowledge service platform. Background Technology

[0002] Against the backdrop of continuous innovation and accelerated iterative upgrades in aerospace equipment development models, researchers have placed higher demands on the speed and accuracy of knowledge services. How to accurately and promptly perceive changes in researchers' literature needs from user behavior logs of knowledge service platforms, and accordingly produce targeted knowledge products efficiently and with high quality, has become a core challenge for scientific and technological intelligence work under the new circumstances.

[0003] Currently, science and technology intelligence personnel primarily rely on attending technical meetings and engaging in offline exchanges with researchers to passively capture researchers' literature needs. They then construct search queries, conduct searches on knowledge service platforms, and subsequently manually filter the search results to organize them into specific knowledge products. This human-driven knowledge product production model has significant drawbacks:

[0004] Demand perception lag: Relying on non-real-time, intermittent communication, it is impossible to continuously and automatically monitor demand dynamics, resulting in a slow response to researchers' urgent and rapidly changing literature needs, especially those generated during the process of tackling technical challenges.

[0005] The production cycle is long and updates and maintenance are difficult: from demand collection, search-based construction, literature acquisition to the final organization of knowledge products, the entire process is highly dependent on manual operation, the construction cycle is long, and it is difficult to dynamically update and maintain in a timely manner according to changes in demand.

[0006] As knowledge service platforms gradually improve their user behavior recording capabilities, they can effectively capture changes in users' document needs using big data analytics. Based on technologies such as semantic clustering and semantic retrieval, knowledge products such as thematic collections can be quickly constructed according to users' document needs, making it possible to meet users' document needs in a timely manner. Therefore, this application is hereby submitted. Summary of the Invention

[0007] The purpose of this application is to provide a method for constructing thematic collections for a knowledge service platform, so as to support the automatic, timely and accurate construction of thematic collections and improve the response speed and quality of knowledge services.

[0008] The technical solution of this application is:

[0009] A method for constructing a thematic collection of articles for a knowledge service platform, comprising:

[0010] Step 1: Perceiving Changes in Literature Needs

[0011] Analyze user search logs on knowledge service platforms to understand changes in users' document needs;

[0012] Step Two: Literature Requirements Analysis

[0013] After determining that users' needs for literature have changed, the knowledge service platform analyzes the literature that users have interacted with during the interval of the change in needs, and mines the core content of users' needs for literature during the interval of the change in needs, so as to provide themes and initial query terms for the construction of thematic collections.

[0014] Step 3: Expanding the literature search terms:

[0015] Semantic expansion is performed on the initial query terms for each topic to obtain an expanded query term sequence;

[0016] Step Four: Constructing a Thematic Collection:

[0017] Using multiple sets of extended query terms for each topic, semantic retrieval is performed on a knowledge service platform, and the obtained documents are used to construct thematic collections.

[0018] Optionally, in the above method for constructing thematic collections for knowledge service platforms, step one includes:

[0019] S11. Collect and analyze user search logs from the knowledge service platform, and extract the search hot word sets K' and K'' from the two time periods. The search hot word set is a sorted list of search terms.

[0020] S12. The ranking Jaccard distance metric is used to measure the difference between the search hot word sets K' and K'' in the two time periods to determine whether the user's demand for literature has changed.

[0021] Optionally, in the above-mentioned method for constructing a thematic collection of knowledge service platforms, in step S11, when extracting the set of search hot words, the user search logs are preprocessed, including word segmentation, removal of stop words, word frequency statistics, and sorted according to word frequency from high to low, and the top 100 search words are extracted to form the set of search hot words.

[0022] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, in S12, if the S-JD distance S-JD(K',K'') of the search hot word set in two consecutive time periods is greater than the distance threshold δ, then it is determined that the user's demand for literature has changed.

[0023] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, step two involves analyzing the documents that users browse, download, collect, and share on the knowledge service platform during the interval between changes in demand.

[0024] Optionally, in the above method for constructing thematic collections for knowledge service platforms, step two includes:

[0025] S21. Vectorize the titles of documents in the knowledge service platform and construct a document title vector matrix;

[0026] S22. Reduce the dimensionality of the document title vector matrix;

[0027] S23. Cluster the titles of the vectorized documents that the knowledge service platform has interacted with users during the interval of demand changes to form multiple title clusters;

[0028] S24. Generate the topic of each title cluster and the initial query terms for that topic.

[0029] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, in S21, the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model is used to convert the titles of documents in the knowledge service platform into high-dimensional semantic vectors.

[0030] In S22, the UMAP algorithm is used to reduce the dimensionality of the high-dimensional semantic vector matrix of the document titles.

[0031] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, in S23, the Ward algorithm based on hierarchical clustering is used to cluster the titles of the vectorized representations of documents that the knowledge service platform and users have interacted with during the interval of demand changes.

[0032] S24 specifically includes:

[0033] Calculate the TF-IDF value of each word within the title cluster, where TF is the frequency of the word in the current title cluster, and IDF is the reciprocal of the frequency of the word in all title clusters.

[0034] Calculate the cosine distance (cosim) between each word in the title cluster and the cluster centroid.

[0035] Calculate the comprehensive evaluation value α(TF-IDF)+(1-α) cosim for each word within the title cluster, where α is the linear combination parameter, which is taken as 0.5;

[0036] The words with the highest overall evaluation value are selected as the topic of the title cluster, and the words with the highest overall evaluation value within the title cluster are selected as the initial query terms for the topic.

[0037] Optionally, in the above method for constructing thematic collections for knowledge service platforms, step three includes:

[0038] S31. Construct semantically enhanced lexical co-occurrence networks for each topic;

[0039] S32. Using the initial query terms of each topic as the central terms, semantics are expanded forward and backward according to the word co-occurrence network to form forward expanded query terms and backward expanded query terms.

[0040] S33. Calculate the semantic contribution score of each word in the forward expansion query term and the backward expansion query term, retain the word with the highest semantic contribution score, and merge it with the initial query term as the expansion query term;

[0041] S34. Using forward and backward extended query terms as new center terms, semantics are expanded unidirectionally forward and backward according to the word co-occurrence network to form new forward and backward extended query terms. The word with the highest semantic contribution score is calculated and retained and added to the extended query terms.

[0042] Optionally, in the above method for constructing thematic collections of knowledge service platforms, S31 specifically includes:

[0043] Calculate the co-occurrence rate of word pairs in the titles of documents under each topic;

[0044] Vocabulary pairs with high co-occurrence rates are labeled with semantic relationship directions to form a semantically enhanced vocabulary co-occurrence network.

[0045] Optionally, in the above-mentioned method for constructing a thematic collection of knowledge service platforms, step S34 can be repeated multiple times in step three.

[0046] Optionally, in the above method for constructing thematic collections for knowledge service platforms, step four includes:

[0047] S41. Vectorize the multiple sets of extended query term sequences for each topic to generate query vectors and perform literature retrieval on the knowledge service platform.

[0048] S42. Organize the retrieved literature and construct thematic collections for each topic.

[0049] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, in S41, based on the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model, the extended query terms of each topic are converted into semantic vectors, normalized, and semantic query vectors are generated. The semantic retrieval model is then used in the knowledge service platform to perform semantic search and retrieve documents with similarity greater than a specified threshold.

[0050] Optionally, in the above-mentioned method for constructing thematic collections of knowledge service platforms, in step S42, documents with earlier publication dates, incomplete metadata, or duplicate content are removed from the retrieved documents. The document value DV is calculated by comprehensively considering the thematic relevance R, document quality Q, and document timeliness T, and then sorted from highest to lowest to construct thematic collections for each theme.

[0051] This application has at least the following beneficial technical effects:

[0052] This paper presents a method for constructing thematic collections on a knowledge service platform. By sensing changes in user literature needs through user search logs on the knowledge service platform, and then capturing user needs based on semantic document clustering, the method further constructs thematic literature collections by building semantically enhanced queries. This method can significantly improve the timeliness and accuracy of thematic collection construction and quickly respond to the dynamic knowledge needs of researchers. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the method for constructing a thematic collection of knowledge service platforms provided in the embodiments of this application.

[0054] To better illustrate this embodiment, some content in the accompanying drawings may be omitted. They are for illustrative purposes only and should not be construed as limiting the scope of this application. Detailed Implementation

[0055] To make the technical solution and advantages of this application clearer, the technical solution of this application will be described in a clearer and more complete manner below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some embodiments of this application, and are only used to explain this application, not to limit this application. It should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, and other related parts can be referred to the general design.

[0056] Furthermore, unless otherwise defined, the technical or scientific terms used in this application description shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The word "comprising" as used in this application description indicates that the concept preceding the word encompasses the concepts listed following the word and their equivalents, without excluding other related concepts.

[0057] A method for constructing thematic collections for a knowledge service platform, such as Figure 1 As shown.

[0058] Step 1: Perceiving changes in literature requirements.

[0059] Analyze user search logs on the knowledge service platform to detect changes in users' literature needs and use this as a signal to trigger subsequent processes.

[0060] S11. Collect and analyze user search logs from the knowledge service platform, and extract the search hot word sets K' and K'' from the two time periods. The search hot word set is a sorted list of search terms.

[0061] When extracting the set of hot search terms, the user search logs are preprocessed, including word segmentation, removal of stop words, and word frequency statistics. The terms are sorted from high to low frequency, and the top 100 search terms are extracted to form the set of hot search terms.

[0062] Word segmentation: Each search term is segmented. During the segmentation process, a Chinese-English word segmentation tool that supports custom domain dictionaries is used, such as jieba. To avoid over-segmentation of domain words and loss of information, a domain-specific dictionary can be introduced. For example, a custom dictionary for the aviation domain can be introduced during word segmentation.

[0063] Stop word removal: To avoid interference from prepositions, conjunctions, auxiliary words, articles, etc. with low meaning, the Harbin Institute of Technology stop word list can be used to delete stop words in the corpus.

[0064] Word frequency statistics: The Counter function in Python's collections library can be used to efficiently count word frequencies.

[0065] S12. The Sorted Jaccard Distance (S-JD) metric is used to measure the difference between the search hot word sets K' and K'' in the two time periods before and after, to determine whether the user's demand for literature has changed significantly.

[0066] The rank of hot word h in the ordered set K is defined as sort(K, h) = |K| - i, where |K| is the maximum assignable rank value of the ordered set K, and i is the position of h in the ordered set K.

[0067] The sort(K'∩K'') is calculated as the intersection of the search hot word sets for the two time periods. Its value is not a simple count, but rather the sum of the smaller ranking values ​​of each hot word h in the intersection across the two sets.

[0068] sort(K'∩K'') = Σ(h∈K'∩K'') min[sort(K', h), sort(K'',h) ].

[0069] Calculate the sorted union of the search hot word sets for the two time periods, sort(K'∪K''). For each hot word h in the union, sum its average rank in both sets:

[0070] sort(K'∪K'')=Σ(h∈K'∪K'')[sort(K',h)+sort(K'',h)] / 2.

[0071] Calculate the S-JD distance of the search hot keyword set in the two time periods:

[0072] S-JD(K',K'') =1- [sort(K'∩K'') / sort(K'∪K'') ].

[0073] If S-JD(K',K'') is greater than the distance threshold δ, it is determined that the user's demand for literature has changed significantly, and subsequent literature demand content analysis steps are triggered. The distance threshold δ can be set according to actual conditions. This mechanism enables the knowledge service platform to promptly capture newly emerging literature demands or old literature demands that have regained attention.

[0074] Step 2: Analysis of literature requirements.

[0075] After determining that users' needs for literature have changed significantly, we analyze the literature that the knowledge service platform has interacted with during the interval of demand change, including literature that has been browsed, downloaded, collected, and shared. We then explore the core content of users' needs for literature during the interval of demand change, providing themes and initial search terms for the subsequent construction of thematic collections.

[0076] S21. Vectorize the titles of documents in the knowledge service platform and construct a document title vector matrix.

[0077] Since knowledge service platforms may contain multiple languages, the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model, which supports multilingual processing and is provided by Sentence-BERT, is used. A batch processing method is employed based on the scale of the document data on the knowledge service platform to transform the titles of the documents into high-dimensional semantic vectors, achieving text-based semantic representation of the discourse. Because the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model can obtain semantically meaningful sentence vectors, it can effectively solve the problem of sparse semantic features in document titles by calculating the similarity of document titles.

[0078] S22. Perform dimensionality reduction on the document title vector matrix.

[0079] The UMAP algorithm is used to reduce the dimensionality of the high-dimensional semantic vector matrix of document titles. UMAP is a highly effective and scalable dimensionality reduction algorithm that can map high-dimensional probability distributions to a low-dimensional space while preserving more global structural information of document titles, effectively reducing the complexity of subsequent clustering calculations.

[0080] It can create IVF-PQ indexes, optimize query speed and memory usage for document title vectors, and form a semantic retrieval model.

[0081] S23. Cluster the titles of documents whose vectorized representations were used by the knowledge service platform and users during the interval of demand changes to form multiple title clusters.

[0082] Experiments were conducted to compare the non-hierarchical k-means algorithm and the hierarchical clustering Ward algorithm. The hierarchical clustering Ward algorithm was selected to cluster the titles of the vectorized documents that the knowledge service platform had interacted with users during the interval of demand changes. Each title cluster represents a potential topic.

[0083] S24. Generate the topic of each title cluster and the initial query terms for that topic.

[0084] Calculate the TF-IDF value of each word within the heading cluster, where TF is the frequency of the word in the current heading cluster, and IDF is the reciprocal of the frequency of the word in all heading clusters.

[0085] Calculate the cosine distance (cosim) between each word in the title cluster and the cluster centroid.

[0086] Calculate the comprehensive evaluation value α(TF-IDF)+(1-α) cosim for each word within the title cluster, where α is a linear combination parameter. Experiments have shown that when α is 0.5, the comprehensive evaluation value is most representative.

[0087] By combining the topic blacklist, non-compliant or non-domain technical terms are deleted from the title cluster. The term with the highest comprehensive evaluation value is selected as the topic of the title cluster. The top 5 terms with the highest comprehensive evaluation value in the title cluster are selected as the initial query terms for the topic.

[0088] The document titles under each topic are segmented into words, stop words are removed, and word frequency is statistically analyzed.

[0089] Step 3: Expanding the literature search terms.

[0090] Considering the data quality of knowledge service platforms, using semantic retrieval to directly use single keywords lacking contextual information as query terms for similarity calculation results in poor semantic matching. It is usually necessary to enrich the information content of keywords.

[0091] Literature searches using initial query terms for each topic are too short. They only expand the information from the perspective of word meaning without supplementing the word meaning or semantics itself. The search information is not complete enough and it is difficult to obtain satisfactory literature search results. Therefore, it is necessary to semantically expand the initial query terms for each topic to obtain an expanded query term sequence.

[0092] S31. Construct semantically enhanced lexical co-occurrence networks for each topic.

[0093] Calculate the co-occurrence rate of word pairs in the titles of documents under each topic:

[0094] rate(a,b) = co(a,b) / [fre(a) + fre(b) - co(a,b)], where co(a,b) is the number of times the word pair (a,b) appears together in the document title, and fre(a) and fre(b) are the number of times words a and b appear individually in the document title.

[0095] Vocabulary pairs with high co-occurrence rates are labeled with semantic relationship directions to form a semantically enhanced vocabulary co-occurrence network.

[0096] For word pairs with a co-occurrence rate greater than 80%, the semantic relationship direction can be marked. The order in which word pairs (a,b) appear in the document title can be counted. If word a appears more frequently than word b, the semantic relationship direction of word pair (a,b) is marked as 1, indicating that word a tends to appear before word b. If word b appears more frequently than word a, the semantic relationship direction of word pair (a,b) is marked as -1, indicating that word a tends to appear after word b. If word a and word b are the same word, they are marked as 0.

[0097] After constructing semantically enhanced lexical co-occurrence networks for each topic, based on the principle of transitivity, the initial query terms are expanded in both forward and backward directions to obtain expanded query terms. This improves the contextual information of the query terms and enhances the recall rate of semantic retrieval.

[0098] S32. Using the initial query terms of each topic as the central terms, semantics are expanded forward and backward according to the word co-occurrence network to form forward expanded query terms and backward expanded query terms.

[0099] S33. Calculate the semantic contribution score of each word in the forward and backward extended query terms, retain the word with the highest semantic contribution score, and merge it with the initial query term as the extended query term.

[0100] S(Ci) = rate(Qi,Ci) × g(QI,Ci);

[0101] in,

[0102] S(Ci) is the semantic contribution score of the i-th word in the forward expansion query term and the backward expansion query term;

[0103] rate(Ci,Qi) is the co-occurrence rate of the i-th word in the forward expansion query term and the backward expansion query term with the current query term;

[0104] g(QI,Ci) is the association strength function between the center word and the i-th word in both the forward and backward expansion query terms.

[0105] If co(QI, Ci) > 0, then g(QI,Ci) = 1 / rank(QI,Ci); otherwise, g(QI,Ci) = 0. Here, co(QI,Ci) is the number of times the i-th word in the forward and backward expanded query terms and the head word co-occur in the document title; rank(QI,Ci) is the position value of the i-th word in the forward and backward expanded query terms in the initial query term co-occurrence list after sorting by co-occurrence frequency from high to low.

[0106] S34. Using forward and backward extended query terms as new center terms, semantics are expanded unidirectionally forward and backward according to the word co-occurrence network to form new forward and backward extended query terms. The word with the highest semantic contribution score is calculated and retained and added to the extended query terms.

[0107] Step S34 can be repeated multiple times, but in order to ensure the accuracy of predicting extended terms based on the initial query term, an appropriate number of repetitions should be selected. To prevent semantic overgeneralization, the number of times step S34 is repeated can be set to 3 times, forming multiple sets of extended query term sequences in the form of <......, preceding extended term 1, central term, subsequent extended term 1, ......>.

[0108] Step 4: Constructing a thematic collection.

[0109] Using multiple sets of extended query terms for each topic, semantic retrieval is performed on a knowledge service platform, and the obtained documents are used to construct thematic collections.

[0110] S41. Vectorize the multiple sets of extended query term sequences for each topic to generate query vectors, and then perform document retrieval on the knowledge service platform.

[0111] Based on the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model, the extended query terms of each topic are transformed into semantic vectors, normalized, and semantic query vectors are generated. Semantic search is then performed on the knowledge service platform using a semantic retrieval model to retrieve and select documents with similarity greater than a specified threshold.

[0112] S42. Organize the retrieved literature and construct thematic collections for each topic.

[0113] Considering that the literature is usually displayed in a list format on the platform, we can first remove the literature that was published earlier, has incomplete metadata, or has duplicate content. Then, we design a literature sorting strategy, taking into account the relevance of the literature to the topic (R), the quality of the literature (Q), and the timeliness of the literature (T). We calculate the value of the literature (DV) using a comprehensive weighted scoring method and sort them from highest to lowest to construct thematic collections for each topic.

[0114] DV = α × R + β × Q + γ × T;

[0115] in,

[0116] α, β, and γ are weighting coefficients that satisfy α + β + γ = 1. Based on experience, α = 0.5, β = 0.3, and γ = 0.2 are usually set.

[0117] The topic relevance R can be represented by the similarity score from semantic retrieval, which is used to ensure that the content of the document is highly matched with the topic.

[0118] The quality Q of a document can be preset on the knowledge service platform based on the source of the document. It is used to reflect the value of the document content and can be enumerated, such as ICAS=1.0, AIAA=0.95, DOD=0.8, and others=0.6.

[0119] The timeliness T of the document can be calculated as follows:

[0120] T=e^(-λ×(current_year-publish_year));

[0121] in,

[0122] λ is the attenuation coefficient, which is usually taken as 0.1;

[0123] current_year is the current year;

[0124] publish_year is the year the document was published;

[0125] e is a natural constant, taken as 2.71828.

[0126] Ultimately, this results in a high-quality collection of documents that are highly relevant to the topic, namely, a thematic anthology.

[0127] The knowledge service platform thematic document collection construction method disclosed in the above embodiments perceives changes in user document needs through user search logs of the knowledge service platform, then captures user needs based on semantic document clustering, and finally constructs a thematic document collection by constructing semantically enhanced queries.

[0128] The knowledge service platform thematic collection construction method disclosed in the above embodiments can be implemented through programming on the knowledge service platform, realizing a complete closed loop from sensing changes in demand, intelligently analyzing demand topics to automatically constructing thematic collections. This can significantly improve the timeliness and accuracy of thematic collection construction, and quickly respond to the dynamic knowledge needs of researchers.

[0129] The technical solution of this application has been described in conjunction with the preferred embodiments shown in the accompanying drawings. Those skilled in the art should understand that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for constructing a thematic collection of articles for a knowledge service platform, characterized in that, include: Step 1: Perceiving Changes in Literature Needs Analyze user search logs on knowledge service platforms to understand changes in users' document needs; Step Two: Literature Requirements Analysis After determining that users' needs for literature have changed, the knowledge service platform analyzes the literature that users have interacted with during the interval of the change in needs, and mines the core content of users' needs for literature during the interval of the change in needs, so as to provide themes and initial query terms for the construction of thematic collections. Step 3: Expanding the literature search terms: Semantic expansion is performed on the initial query terms for each topic to obtain an expanded query term sequence; Step Four: Constructing a Thematic Collection: Using multiple sets of extended query word sequences for each topic, semantic retrieval is performed on a knowledge service platform, and the obtained documents are used to construct thematic collections. Step one includes: S11. Collect and analyze user search logs from the knowledge service platform, and extract the search hot word sets K' and K'' from the two time periods. The search hot word set is a sorted list of search terms. S12. The ranking Jaccard distance index is used to measure the difference between the search hot word sets K' and K'' in the two time periods to determine whether the user's demand for literature has changed. Step three includes: S31. Construct semantically enhanced lexical co-occurrence networks for each topic; S32. Using the initial query terms of each topic as the central terms, semantics are expanded forward and backward according to the word co-occurrence network to form forward expanded query terms and backward expanded query terms. S33. Calculate the semantic contribution score of each word in the forward expansion query term and the backward expansion query term, retain the word with the highest semantic contribution score, and merge it with the initial query term as the expansion query term; S34. Using forward-expanded query terms and backward-expanded query terms as new center terms, semantics are expanded forward and backward in one direction according to the word co-occurrence network to form new forward-expanded query terms and backward-expanded query terms. The words with the highest semantic contribution scores are calculated and retained and added to the expanded query terms. S31 specifically includes: Calculate the co-occurrence rate of word pairs in the titles of documents under each topic; Vocabulary pairs with high co-occurrence rates are labeled with semantic relationship directions to form a semantically enhanced vocabulary co-occurrence network; In step three, step S34 is repeated multiple times.

2. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 1, characterized in that, In S11, when extracting the set of hot search terms, the user search logs are preprocessed, including word segmentation, removal of stop words, and word frequency statistics. The words are sorted from high to low frequency, and the top 100 search terms are extracted to form the set of hot search terms.

3. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 2, characterized in that, In S12, if the S-JD distance S-JD(K',K'') of the search hot word set in two consecutive time periods is greater than the distance threshold δ, it is determined that the user's demand for literature has changed.

4. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 3, characterized in that, In step two, the documents that users browsed, downloaded, collected, and shared on the knowledge service platform during the intervals of demand changes are analyzed.

5. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 4, characterized in that, Step two includes: S21. Vectorize the titles of documents in the knowledge service platform and construct a document title vector matrix; S22. Reduce the dimensionality of the document title vector matrix; S23. Cluster the titles of the vectorized documents that the knowledge service platform has interacted with users during the interval of demand changes to form multiple title clusters; S24. Generate the topic of each title cluster and the initial query terms for that topic.

6. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 5, characterized in that, In S21, the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model is used to convert the titles of documents in the knowledge service platform into high-dimensional semantic vectors. In S22, the UMAP algorithm is used to reduce the dimensionality of the high-dimensional semantic vector matrix of the document titles.

7. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 6, characterized in that, In S23, the Ward algorithm based on hierarchical clustering is used to cluster the titles of the vectorized representations of documents that the knowledge service platform and users have interacted with during the interval of demand changes. S24 specifically includes: Calculate the TF-IDF value of each word within the title cluster, where TF is the frequency of the word in the current title cluster, and IDF is the reciprocal of the frequency of the word in all title clusters. Calculate the cosine distance (cosim) between each word in the title cluster and the cluster centroid. Calculate the comprehensive evaluation value α(TF-IDF)+(1-α) cosim for each word within the title cluster, where α is the linear combination parameter, which is taken as 0.5; The words with the highest overall evaluation value are selected as the topic of the title cluster, and the words with the highest overall evaluation value within the title cluster are selected as the initial query terms for the topic.

8. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 7, characterized in that, Step four includes: S41. Vectorize the multiple sets of extended query term sequences for each topic to generate query vectors and perform literature retrieval on the knowledge service platform. S42. Organize the retrieved literature and construct thematic collections for each topic.

9. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 8, characterized in that, In S41, based on the paraphrase-multilingual-MiniLM-L12-v2 pre-trained language model, the extended query terms of each topic are transformed into semantic vectors, normalized, and semantic query vectors are generated. Semantic search is performed on the knowledge service platform using the semantic retrieval model to retrieve and select documents with similarity greater than a specified threshold.

10. The method for constructing a thematic collection of documents for a knowledge service platform according to claim 9, characterized in that, In S42, the thematic relevance R, quality Q, and timeliness T of the documents are comprehensively considered to calculate the document value DV, which is then sorted from largest to smallest to construct thematic collections for each theme.

Images