Intelligent retrieval method and system for online document management

Abstract

The invention belongs to the technical field of document retrieval, and provides an intelligent retrieval method and system for online document management. The method comprises the following steps of: dividing a document, and establishing an independent index for each document block in a search engine by taking the document block as a minimum index unit; the method comprises the following steps: receiving a search request input by a user, performing authority verification on the search request, then performing natural language processing and vectorization coding, performing keyword inverted index retrieval and vector semantic retrieval after to-be-retrieved content is obtained, and performing result fusion to generate a result sorting list; dynamically adjusting the result sorting list on the basis of user behavior characteristics, highlighting contents matched with a user search request, and automatically recommending similar documents when the user does not search a target document; and monitoring the state of the document in real time, and updating the dynamic incremental index. According to the method and the device, accurate and efficient intelligent document search can be realized, and personalized search requirements of users are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document retrieval, in particular to an intelligent retrieval method and system for online document management. BACKGROUND

[0002] Currently, online document management systems (such as enterprise network disks, collaborative office platforms, etc.) generally provide full-text retrieval functions to enable users to quickly find target documents; most existing online document management systems use search engines based on inverted indexes to implement document content retrieval, which has the following defects: Firstly, the existing retrieval method stores the entire document as a whole in the search engine, and takes the entire document as the retrieval unit when retrieving, which makes it difficult to directly locate a certain chapter or paragraph, and the user needs to manually search again, which is inefficient; Secondly, the existing retrieval method mainly relies on keyword matching, and lacks semantic understanding of documents, so when the user inputs ambiguous words or synonyms, it is often difficult to accurately match the target document; Thirdly, for multi-version documents, the existing retrieval method cannot retrieve the latest version of the document according to the version dimension, and it is also difficult to break through the language barrier to retrieve documents in foreign languages with local languages; and the final search result sorting mainly depends on the relevance score, lacks dynamic optimization based on user behavior, and cannot meet the user's personalized search needs.

[0003] Therefore, we need to develop an intelligent retrieval method and system for online document management that can accurately retrieve target documents from both temporal and content dimensions, support multi-language version retrieval, and optimize the retrieval results based on user preferences to achieve precise and efficient intelligent document search. SUMMARY

[0004] The present application aims to provide an intelligent retrieval method and system for online document management to solve the problems of low efficiency, inaccuracy and difficulty in meeting user's personalized search needs of the existing retrieval method mentioned in the background.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: According to one aspect of the present application, an intelligent retrieval method for online document management is provided, which comprises the following steps: The document content is parsed and divided to obtain document blocks, and an independent index is established for each document block in the search engine with the document block as the smallest indexing unit; wherein the index records document information and user permission information, and supports multiple language versions; receive a search request input by a user, perform permission verification on the search request based on user permission information in the index, and perform natural language processing and vectorization coding on the content of the search request that passes the permission verification to obtain to-be-retrieved content; based on the to-be-retrieved content, perform double-channel retrieval combining keyword inverted index retrieval and vector semantic retrieval, and fuse the double-channel retrieval results through a fusion algorithm to generate a result ranking list; based on user behavior characteristics, dynamically adjust the result ranking list, highlight the content matched with the user search request, and automatically recommend similar documents when the user does not search for a target document; real-time monitoring of document status and dynamic incremental index updating.

[0006] According to another aspect of the present application, an intelligent retrieval system for online document management is provided, which comprises an index construction module, a request processing module, a document retrieval module, a result display module, and an index updating module. Wherein: The above-mentioned index construction module is used for parsing document content and performing document division, obtaining document blocks, and establishing independent indexes for each document block in a search engine with the document block as the minimum index unit; wherein the index records document information and user permission information, and supports multiple language versions; The request processing module is used for receiving a search request input by a user, performing permission verification on the search request based on user permission information in the index, and performing natural language processing and vectorization coding on the content of the search request that passes the permission verification to obtain to-be-retrieved content; The document retrieval module is used for, based on the to-be-retrieved content, performing double-channel retrieval combining keyword inverted index retrieval and vector semantic retrieval, and fusing the double-channel retrieval results through a fusion algorithm to generate a result ranking list; The result display module is used for, based on user behavior characteristics, dynamically adjusting the result ranking list, highlighting the content matched with the user search request, and automatically recommending similar documents when the user does not search for a target document; The index updating module is used for real-time monitoring of document status and dynamic incremental index updating.

[0007] Based on the foregoing scheme, the document division specifically comprises: parsing document content, identifying the structural characteristics of the document, and dynamically selecting a division granularity according to the document type and the structural characteristics; wherein the division granularity comprises: performing document division with a page as the block unit according to document page symbols or page number information; performing document division with a paragraph as the block unit according to document natural paragraph markers; According to the document title hierarchy, the document is divided into logical chapter units.

[0008] Based on the foregoing scheme, the document information recorded in the index includes: document ID, document chunk ID, document version number, timestamp, and paragraph position information corresponding to the document chunk; wherein, when an independent index is established for each document chunk, it further includes: automatic language recognition of the document, index establishment of multiple language versions through machine translation, so that when subsequent document retrieval is performed, the user input keyword can be automatically mapped to the target language index, and cross-language search is realized.

[0009] Based on the foregoing scheme, the natural language processing and vectorization coding of the search request content that passes the permission verification are specifically as follows: The natural language processing of the search request content that passes the permission verification includes word segmentation, part-of-speech tagging, named entity recognition, and stop word filtering, and extracts core semantic units; The core semantic units are mapped to high-dimensional dense vectors to obtain the to-be-retrieved content.

[0010] Based on the foregoing scheme, before the double-channel retrieval of the document, it further includes: if the user's search request also contains additional search conditions (such as: searching only the latest version, etc.), the system will limit the candidate index range before retrieval according to the document version number or timestamp recorded in the index and other document information, and only query the target document or document chunk that meets the user's additional search conditions.

[0011] Based on the foregoing scheme, the keyword inverted index retrieval is specifically as follows: The to-be-retrieved content is segmented and preprocessed to obtain standardized query terms; According to the standardized query terms, corresponding inverted lists are found in the inverted index to obtain N inverted lists; wherein, the inverted list records document identifiers and term information containing the standardized query terms; According to the query logic, set operations are performed on the N inverted lists to obtain a preliminary matched document chunk set; wherein, the set operation is: intersection, union or difference set; The matching degree score of each candidate document chunk in the preliminary matched document chunk set is calculated, the candidate document chunks are sorted according to the matching degree score, and the keyword retrieval result is obtained.

[0012] Based on the foregoing scheme, the vector semantic retrieval is specifically as follows: In the pre-constructed vector index library, the approximate nearest neighbor search is performed with cosine similarity as the measurement standard, to obtain document vectors similar to the to-be-retrieved content and their semantic similarity scores; Map the document vectors back to the corresponding document chunks and rank them according to the semantic similarity scores to obtain semantic retrieval results.

[0013] Based on the foregoing scheme, the double-channel retrieval results are fused by a fusion algorithm, specifically: Merge the keyword retrieval results and the semantic retrieval results, and perform deduplication processing to obtain a candidate result pool. Based on the matching degree score and the semantic similarity score, calculate the comprehensive score of each candidate result in the candidate result pool, and sort the result ranking list in descending order according to the comprehensive score.

[0014] Based on the foregoing scheme, the result ranking list is dynamically adjusted based on user behavior characteristics, specifically: Collect user behavior data in historical search records, and build a user preference model based on the behavior data; Extract the user behavior characteristics from the user preference model; wherein the user behavior characteristics include: click times, dwell time, historical search interest; Assign a corresponding weight coefficient to each of the user behavior characteristics, and calculate a user preference score of each candidate result in the result ranking list based on the weight coefficient; Sort the result ranking list in descending order according to the user preference score to obtain a final result list, and highlight the top N results in the final result list as matching the user search request; Based on the user behavior data, update the user preference model according to a preset update period.

[0015] Based on the foregoing scheme, when the user does not search for the target document, similar documents are automatically recommended, specifically: If the final result list is empty, or the user's dwell time on the top N results is less than a preset threshold, or the user continuously searches for semantically similar keywords for N times but does not click any result, when any of the conditions is met, it is determined that the user does not search for the target document, triggering the recommendation mechanism, and the system automatically generates a recommendation list and returns it to the user; wherein the recommendation list contains documents semantically similar to the user's historical clicked documents or commonly used documents of the team to which the documents belong.

[0016] Based on the foregoing scheme, when the results are displayed, if there are multiple versions of the same document, the system will provide a version jump function on the result display page according to the document version number recorded in the index, and the user can quickly switch to view other versions of the same document chunk.

[0017] Based on the foregoing scheme, the document state is monitored in real time, and dynamic incremental index updating is performed, specifically: When detecting the document editing event, compare the current document version with the content difference of the last index version, identify N target blocks that have content changes; For each target block, a new independent index is established in the search engine, and the document block index that has not changed is kept unchanged.

[0018] From the above technical solution, compared with the prior art, the present application has at least the following advantages and positive effects: (1) The present application divides the document and establishes an independent index for each document block, records the document information and user permission information in the index, and supports multiple language versions, which can quickly and accurately locate the target document or target document paragraph during subsequent document search, and also realizes cross-language document retrieval.

[0019] (2) The present application adopts a double-channel retrieval method combining keyword inverted index retrieval and vector semantic retrieval, which can meet the accuracy of keyword retrieval and optimize the search results through the recall ability of semantic retrieval to realize more efficient and accurate intelligent search.

[0020] (3) The present application dynamically adjusts the search results based on user behavior characteristics, and when it is detected that the user has not searched for the target document, triggers the recommendation mechanism to automatically recommend similar documents, thereby meeting the personalized search needs of users.

[0021] (4) When the document content changes, the present application only performs incremental update on the changed block content, which can reduce the index pressure caused by the overall update of large documents, and for the document being edited, the user can still retrieve the unmodified block content during search, ensuring the real-time and continuity of search and improving the search efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only relate to some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 A flowchart of an intelligent retrieval method for online document management provided by an embodiment of the present application is shown; Figure 2 A method flowchart for generating a result ranking list based on a double-channel retrieval method provided by an embodiment of the present application is shown; Figure 3A flow chart of a method for keyword inverted index retrieval is shown according to an embodiment of the present application; Figure 4 A flow chart of a method for vector semantic retrieval is shown according to an embodiment of the present application; Figure 5 A flow chart of a method for dynamically adjusting a result ranking list based on user behavior characteristics is shown according to an embodiment of the present application; Figure 6 A flow chart of a method for dynamically incrementally updating a document index is shown according to an embodiment of the present application; Figure 7 A structural schematic diagram of an intelligent retrieval system for online document management is shown according to an embodiment of the present application; In the drawings, Figure 7 The reference signs in the drawings are explained as follows: 700 - an intelligent retrieval system for online document management; 701 - an index construction module, 7011 - a document division unit, 7012 - an index construction unit; 702 - a request processing module, 7021 - an authority verification unit, 7022 - a request processing unit; 703 - a document retrieval module, 7031 - a keyword retrieval unit, 7032 - a semantic retrieval unit, 7033 - a result fusion unit; 704 - a result display module, 7041 - a result adjustment unit, 7042 - a document recommendation unit; 705 - an index update module, 7051 - a document monitoring unit, 7052 - an index update unit. DETAILED DESCRIPTION

[0024] In order to more clearly illustrate the object, technical solutions and advantages of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The example embodiments can be implemented in various forms, and should not be understood as being limited to the examples described herein. On the contrary, these embodiments are provided so that the present application will be more comprehensive and complete, and the ideas of the example embodiments will be fully conveyed to those skilled in the art.

[0025] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to give a sufficient understanding of the embodiments of the present application. However, those skilled in the art will realize that the technical solutions of the present application can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be used. In other cases, well-known methods, devices, implementations or operations are not shown or described in detail to avoid obscuring the aspects of the present application.

[0026] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0028] The present invention will now be described in detail with reference to specific embodiments: Example

[0029] like Figure 1 As shown in the figure, this embodiment of the invention provides an intelligent retrieval method for online document management. The specific steps of the method are as follows: S1: Parse the document content and divide the document into blocks. Using each block as the smallest index unit, establish an independent index for each document block in the search engine. The index records document information and user permission information and supports multiple language versions. Unlike existing retrieval methods that index the entire document, this invention automatically divides the document into paragraphs, pages, or logical chapters, using blocks as the smallest index unit, and establishing a mapping relationship between blocks and documents, version numbers, etc. Specifically, in this embodiment, the document division is as follows: The document content is parsed, the structural features of the document are identified, and the granularity of the division is dynamically selected based on the document type and the structural features. The granularity of the division includes: dividing the document into blocks based on page breaks or page number information, using pages as blocks; dividing the document into blocks based on paragraphs as blocks; and dividing the document into blocks based on logical chapters as blocks, according to the document title hierarchy.

[0030] Furthermore, after document segmentation, each document block is used as the smallest index unit. An independent index is built in the search engine for each document block, and the index records document information and user permission information. The document information includes: document ID, document block ID, document version number, timestamp, and paragraph position information corresponding to the document block. Furthermore, when building an independent index for each document block, the process also includes: automatic language recognition of the document and the creation of multilingual indexes through machine translation. This enables users to automatically map keywords to the target language index during subsequent document retrieval, achieving cross-language search.

[0031] S2: receiving a search request input by a user, performing permission verification on the search request based on user permission information in the index, and performing natural language processing and vectorization coding on the search request content for which the permission verification is passed, to obtain to-be-retrieved content; Further, when receiving a search request input by a user, the permission of the search request of the user is verified based on the user permission information recorded in the index, and for a user without permission, filtering is directly performed at the index level, thereby improving the search efficiency while improving the text security. Further, if the search request of the user further contains an additional search condition (such as: searching only the latest version), the system will limit the candidate index range before retrieval according to the document version number or timestamp and other document information recorded in the index, and only query the target document or document block that meets the additional search condition of the user.

[0032] Further, the search request content for which the permission verification is passed is subjected to natural language processing and vectorization coding, specifically including: The search request content for which the permission verification is passed is subjected to natural language processing, including word segmentation, part-of-speech tagging, named entity recognition, and stop word filtering, to extract core semantic units. The core semantic units are mapped into high-dimensional dense vectors to obtain the to-be-retrieved content.

[0033] Wherein, for a complex query, dependency syntax analysis, semantic role labeling and other methods can also be used, and the specific processing content can be routinely adjusted according to actual needs, which is not limited by the embodiments of the present application.

[0034] S3: based on the to-be-retrieved content, performing double-channel retrieval combining keyword inverted index retrieval and vector semantic retrieval, and fusing the double-channel retrieval results through a fusion algorithm to generate a result ranking list; Further, after obtaining the to-be-retrieved content, document retrieval is performed in a manner combining keyword inverted index retrieval and vector semantic retrieval, as shown in Figure 2 The result ranking list is generated based on the double-channel retrieval manner, specifically including the following steps: S201: based on the to-be-retrieved content, performing keyword inverted index retrieval to obtain keyword retrieval results; wherein, as shown in Figure 3 The keyword inverted index retrieval specifically includes the following steps: S301: performing word segmentation and preprocessing on the to-be-retrieved content to obtain standardized query terms; S302: according to the standardized query terms, searching for corresponding inverted lists in the inverted index to obtain N inverted lists; wherein, the inverted list records document identifiers and term information containing the standardized query terms; S303: According to the query logic, perform set operations on the N inverted lists to obtain a preliminary matching set of document chunks; wherein, the set operation is: taking the intersection, taking the union, or taking the difference; Specifically, let's further illustrate the content of set operations with concrete examples: Assuming the search query is "AI in healthcare" or "AI AND healthcare", the system searches the inverted index for the inverted lists of the terms "AI" and "healthcare" respectively. For example, the inverted list for the term "AI" is: [Doc1, Doc3, Doc5, Doc7] (meaning documents 1, 3, 5, and 7 contain the term "AI"); the inverted list for the term "healthcare" is: [Doc2, Doc3, Doc5, Doc6] (meaning documents 2, 3, 5, and 6 contain the term "healthcare"). Since the query logic is "AND", an intersection operation is performed to find the document IDs that appear in both inverted lists, resulting in a preliminary matching list: [Doc3, Doc5,]. Assuming the search term is "Artificial Intelligence OR Medical Care", since the query logic is "OR", a union operation is performed to merge the two inverted lists and remove duplicate document IDs, resulting in the preliminary matching list: [Doc1, Doc2, Doc3, Doc5, Doc6, Doc7]. Assuming the search term is "AI NOT Medical", since the query logic is "NOT", a difference operation is performed to remove the document IDs that also appear in the inverted list of the term "medical" from the inverted list of the term "AI", resulting in a preliminary matching list: [Doc1, Doc7].

[0035] If the content to be searched is "medical", the system will find the inverted list of the term "medical" from the inverted index as: [Doc2, Doc3, Doc5, Doc6]. Since there is only one inverted list, the set operation step is skipped.

[0036] S304: Calculate the matching score of each candidate document block in the preliminary matched document block set, sort the candidate document blocks according to the matching score, and obtain the keyword retrieval results; Specifically, the matching score can be calculated based on factors such as word frequency and position. The specific settings can be made according to actual needs, and the embodiments of the present invention do not impose any limitations.

[0037] S202: Based on the content to be retrieved, perform vector semantic retrieval to obtain semantic retrieval results; Among them, such as Figure 4 As shown, vector semantic retrieval specifically includes the following steps: S401: In the pre-constructed vector index library, the cosine similarity is taken as a measurement standard to perform an approximate nearest neighbor search to obtain a document vector similar to the content to be searched and a semantic similarity score thereof; S402: The document vector is mapped back to a corresponding document block, and the semantic similarity score is used for sorting to obtain a semantic search result; In the embodiment, the introduction of the vector library semantic similarity search can compensate for the problem of inaccurate search results caused by the lack of semantic understanding of documents in the traditional document search method, thereby improving the accuracy of document search.

[0038] S203: The keyword search result and the semantic search result are merged and de-duplicated to obtain a candidate result pool; S204: Based on the matching degree score and the semantic similarity score, a comprehensive score of each candidate result in the candidate result pool is calculated, and the candidate result pool is sorted in descending order according to the comprehensive score to generate a result sorting list; Specifically, the calculation formula of the comprehensive score is: Q 总 = α × Q1+ β × Q2; wherein α and β are adjustable parameters for balancing accurate matching and semantic recall; Q 总 is the comprehensive score, Q1 is the matching degree score, and Q2 is the semantic similarity score.

[0039] In the embodiment, the dual-channel search mode combining the keyword inverted index search and the vector semantic search can meet the accuracy of keyword search, and can also optimize the search result through the recall ability of semantic search to realize more efficient and accurate intelligent search.

[0040] S4: The result sorting list is dynamically adjusted based on user behavior characteristics, the content matched with the user search request is highlighted, and when the user does not search for a target document, similar documents are automatically recommended; Further, after obtaining the result sorting list, further adjustment is performed based on user behavior characteristics, as shown in Figure 5 The specific steps include the following steps: S501: User behavior data in historical search records is collected, and a user preference model is constructed based on the behavior data; S502: The user behavior characteristics are extracted from the user preference model; wherein the user behavior characteristics include: click times, dwell time, and historical search interest; S503: Each user behavior characteristic is assigned a corresponding weight coefficient, and a user preference score of each candidate result in the result sorting list is calculated based on the weight coefficient; Specifically, the calculation formula of the user preference score is: Q偏 = W1xT1+W2xT2+...+W n xT n ; wherein, Q 偏 is the user preference score, W n is the weight coefficient corresponding to the user behavior feature, and satisfies W1+W2+...+W n =1; T n is the normalized function corresponding to the user behavior feature. Specifically, when assigning the corresponding weight coefficient to each user behavior feature, different feature weights can be set by empirical values or dynamically adjusted by training algorithms, such as increasing the weight of documents with high click counts, increasing the weight of documents with long dwell times, and the like, which are not limited by the embodiments of the present application.

[0041] S504: The result ranking list is sorted in descending order according to the user preference score to obtain a final result list, and the first N results in the final result list are regarded as content matched with the user search request and are highlighted for display. S505: Based on the user behavior data, the user preference model is updated according to a preset update period.

[0042] Further, if it is detected that the user has not searched for the target document, a recommendation mechanism is triggered to automatically recommend similar documents; specifically, if the final result list is empty, or the dwell time of the user on the first N results filtered out is lower than a preset threshold, or the user has searched for semantically similar keywords for N consecutive times but has not clicked any result, when any of the conditions is met, it is determined that the user has not searched for the target document, the recommendation mechanism is triggered, a recommendation list is automatically generated by the system, and is returned to the user; wherein the recommendation list contains documents semantically similar to the user's historical clicked documents or commonly used documents of the team to which the documents belong.

[0043] Further, when the results are displayed, if there are multiple versions of the same document, the system will provide a version jump function on the result display page according to the document version number recorded in the index, and the user can quickly switch to view other versions of the same document block.

[0044] S5: Real-time monitoring of document status and dynamic incremental index updating.

[0045] Further, the document status is monitored in real time, and when the content of the document is identified to have changed, dynamic incremental index updating is performed; as shown in Figure 6 , the dynamic incremental updating of the index specifically includes the following steps: S601: When a document editing event is detected, the content difference between the current document version and the last index version is compared, and N target blocks with content changes are identified; S602: For each of the target blocks, a new independent index is established in the search engine, and the document block index that has not changed in content is kept unchanged.

[0046] In the embodiment, when the document changes in content, only the changed block content is incrementally updated, which can reduce the index pressure caused by the overall update of the large document, and for the document being edited, the unmodified block content can still be retrieved when the user searches, ensuring the real-time and continuity of the search and improving the search efficiency.

[0047] The intelligent retrieval method provided in the embodiment can not only break through the language barrier and realize cross-language document retrieval, but also realize more efficient and accurate intelligent search in a combination of keyword retrieval and semantic retrieval. Meanwhile, the search results are dynamically adjusted based on user behavior characteristics, and when it is detected that the user has not searched for the target document, a recommendation mechanism is triggered to automatically recommend similar documents, which can meet the personalized search needs of the user. When the document changes in content, only the changed block content is incrementally updated, which can reduce the overhead of index update and storage while ensuring real-time.

[0048] Embodiment 2

[0049] As shown in Figure 7 The embodiment of the present application provides an intelligent retrieval system 700 for online document management, which comprises an index construction module 701, a request processing module 702, a document retrieval module 703, a result display module 704 and an index update module 705; wherein: The index construction module 701 is configured to parse the document content and divide the document to obtain document blocks, and then establish an independent index for each of the document blocks in the search engine as the minimum index unit; wherein the index records the document information and the user permission information, and supports multiple language versions. The index construction module 701 comprises a document division unit 7011 and an index construction unit 7012; wherein: The document division unit 7011 is configured to parse the document content, identify the structural features of the document, dynamically select the division granularity according to the document type and the structural features, and divide the document; wherein the division granularity comprises: dividing the document according to the document page symbol or page number information to obtain the page as the block unit; dividing the document according to the document natural paragraph mark to obtain the paragraph as the block unit; and dividing the document according to the document title level to obtain the logical chapter as the block unit.

[0050] The index construction unit 7012 is configured to take the document block divided by the document division unit 7011 as the minimum index unit, and establish an independent index for each document block in the search engine.

[0051] The request processing module 702 is configured to receive a search request input by a user, perform authority verification on the search request based on user authority information in the index, and perform natural language processing and vectorization coding on the content of the search request that passes the authority verification to obtain to-be-retrieved content. The request processing module 702 includes an authority verification unit 7021 and a request processing unit 7022. The authority verification unit 7021 is configured to receive a search request input by a user, and perform authority verification on the search request based on user authority information in the index. The request processing unit 7022 is configured to perform natural language processing and vectorization coding on the content of the search request that passes the authority verification in the authority verification unit 7021 to obtain to-be-retrieved content. The document retrieval module 703 is configured to perform double-channel retrieval of keyword inverted index retrieval combined with vector semantic retrieval on the to-be-retrieved content, and perform fusion of the double-channel retrieval results by using a fusion algorithm to generate a result ranking list. The document retrieval module 703 includes a keyword retrieval unit 7031, a semantic retrieval unit 7032, and a result fusion unit 7033. The keyword retrieval unit 7031 is configured to perform keyword inverted index retrieval on the to-be-retrieved content to obtain a keyword retrieval result, and specifically includes the following steps: Performing word segmentation and preprocessing on the to-be-retrieved content to obtain standardized query terms. According to the standardized query terms, searching for corresponding inverted lists in the inverted index to obtain N inverted lists, wherein the inverted list records document identifiers and term information containing the standardized query terms. According to query logic, performing set operations on the N inverted lists to obtain a set of preliminarily matched document blocks, wherein the set operations are intersection, union, or difference. Calculating a matching degree score of each candidate document block in the set of preliminarily matched document blocks, and sorting the candidate document blocks according to the matching degree score to obtain a keyword retrieval result.

[0052] The semantic retrieval unit 7032 is configured to perform vector semantic retrieval on the to-be-retrieved content to obtain a semantic retrieval result, and specifically includes the following steps: In the pre-constructed vector index library, a cosine similarity is taken as a measurement standard, an approximate nearest neighbor search is performed, and a document vector similar to the content to be searched and a semantic similarity score thereof are obtained; The document vector is mapped back to a corresponding document block, and is sorted according to the semantic similarity score, so as to obtain a semantic search result.

[0053] The result fusion unit 7033 is configured to merge the keyword search result and the semantic search result to generate a result ranking list, and specifically includes the following steps: The keyword search result and the semantic search result are merged and de-duplicated to obtain a candidate result pool; Based on the matching degree score and the semantic similarity score, a comprehensive score of each candidate result in the candidate result pool is calculated, and the result ranking list is sorted in descending order according to the comprehensive score to generate a result ranking list.

[0054] The result display module 704 is configured to dynamically adjust the result ranking list based on user behavior characteristics, highlight display the content matched with the user search request, and automatically recommend similar documents when the user does not search for a target document. The result display module 704 includes a result adjustment unit 7041 and a document recommendation unit 7042. The result adjustment unit 7041 is configured to dynamically adjust the result ranking list generated by the document search module 703 based on user behavior characteristics, and highlight display the content matched with the user search request. User behavior data in historical search records is collected, and a user preference model is constructed based on the behavior data. The user behavior characteristics are extracted from the user preference model, and the user behavior characteristics include the number of clicks, the dwell time, and the historical search interest. A corresponding weight coefficient is assigned to each user behavior characteristic, and a user preference score of each candidate result in the result ranking list is calculated based on the weight coefficient. The result ranking list is sorted in descending order according to the user preference score to obtain a final result list, the first N results in the final result list are regarded as the content matched with the user search request, and the content is highlighted displayed. The user preference model is updated according to a preset update period based on the user behavior data.

[0055] The document recommendation unit 7042 is configured to trigger a recommendation mechanism to automatically recommend similar documents when it is detected that the user does not search for a target document. If the final result list is empty, or the user's dwell time on the top N results is lower than a preset threshold, or the user successively searches for semantically similar keywords for N times but does not click any result, when any of the conditions is met, it is determined that the user has not searched for the target document, a recommendation mechanism is triggered, a recommendation list is automatically generated by the system, and is returned to the user; wherein the recommendation list contains documents semantically similar to the user's historical clicked documents or commonly used documents of the team to which the documents belong.

[0056] The index updating module 705 is configured to monitor the document state in real time and perform dynamic incremental index updating. The index updating module 705 comprises a document monitoring unit 7051 and an index updating unit 7052. The document monitoring unit 7051 is configured to monitor the document state in real time, and when a document editing event is detected, compare the content difference between the current document version and the last index version, and identify N target blocks that have content changes. The index updating unit 7052 is configured to, for the N target blocks identified by the document monitoring unit 7051, establish a new independent index for each target block in the search engine, and keep the index of the document blocks that have not changed unchanged.

[0057] In the embodiment, the index construction module 701 is used to divide the documents and establish independent indexes for each document block, and the document retrieval module 703 is used to retrieve the documents by combining the double-channel retrieval mode of keyword inverted index retrieval and vector semantic retrieval, which can not only meet the accuracy of keyword retrieval, but also optimize the search results through the recall ability of semantic retrieval. Meanwhile, the search results are dynamically adjusted based on the user behavior characteristics in the result display module 704, and when it is detected that the user has not searched for the target document, the recommendation mechanism is triggered to automatically recommend similar documents, which not only realizes more efficient and accurate intelligent search, but also meets the personalized search needs of the user.

[0058] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims. It is understood that the application is not limited to the precise structures and methods described herein and that various modifications can be made therein without departing from the scope of the application. The scope of the application is limited only by the claims that follow.

Claims

1. An intelligent retrieval method for online document management, characterized in that, Includes the following steps: The document content is parsed and the document is divided into blocks. Each block is used as the smallest index unit, and an independent index is built in the search engine for each document block. The index records document information and user permission information and supports multiple language versions. The system receives a search request input by a user, verifies the user's permission information based on the index, and performs natural language processing and vectorization encoding on the search request content that has passed the permission verification to obtain the content to be retrieved. Based on the content to be searched, a dual-channel search combining keyword inverted index search and vector semantic search is performed, and the dual-channel search results are merged through a fusion algorithm to generate a sorted list of results; The result ranking list is dynamically adjusted based on user behavior characteristics, highlighting content that matches the user's search request, and automatically recommending similar documents when the user does not find the target document. Monitor document status in real time and perform dynamic incremental index updates.

2. The intelligent retrieval method according to claim 1, characterized in that, The document division is as follows: The document content is parsed, its structural features are identified, and the granularity of the partitioning is dynamically selected based on the document type and the structural features; wherein, the granularity of the partitioning includes: Divide the document into blocks based on page breaks or page number information, using pages as the unit of division. The document is divided into sections based on paragraph tags. The document is divided into sections based on the document title hierarchy and logical chapters.

3. The intelligent retrieval method according to claim 1, characterized in that, The process of performing natural language processing and vector encoding on the search request content that has passed permission verification specifically includes: Natural language processing is performed on search request content that has passed permission verification, including word segmentation, part-of-speech tagging, named entity recognition, and stop word filtering, to extract core semantic units; The core semantic unit is mapped to a high-dimensional dense vector to obtain the content to be retrieved.

4. The intelligent retrieval method according to claim 1, characterized in that, The keyword inverted index retrieval is specifically as follows: The content to be retrieved is segmented and preprocessed to obtain standardized query terms; Based on the standardized query terms, the corresponding inverted lists are searched in the inverted index to obtain N inverted lists; wherein, the inverted lists record the document identifiers and term information containing the standardized query terms; Based on the query logic, set operations are performed on the N inverted lists to obtain a preliminary set of matching document chunks; wherein, the set operations are: taking the intersection, taking the union, or taking the difference. Calculate the matching score for each candidate document block in the initially matched document block set, sort the candidate document blocks according to the matching score, and obtain the keyword retrieval results.

5. The intelligent retrieval method according to claim 1, characterized in that, The vector semantic retrieval specifically includes: In a pre-built vector index library, an approximate nearest neighbor search is performed using cosine similarity as the metric to obtain document vectors similar to the content to be retrieved and their semantic similarity scores. The document vectors are mapped back to the corresponding document blocks, and sorted according to the semantic similarity scores to obtain the semantic retrieval results.

6. The intelligent retrieval method according to claim 1, characterized in that, The fusion algorithm is used to fuse the dual-channel retrieval results, specifically as follows: The keyword search results and semantic search results are merged and deduplicated to obtain a candidate result pool. Based on the matching score and semantic similarity score, a comprehensive score is calculated for each candidate result in the candidate result pool, and the results are sorted in descending order according to the comprehensive score to generate the result sorting list.

7. The intelligent retrieval method according to claim 1, characterized in that, The dynamic adjustment of the result ranking list based on user behavior characteristics specifically involves: Collect user behavior data from historical search records, and build a user preference model based on the behavior data; The user behavior features are extracted from the user preference model; wherein, the user behavior features include: number of clicks, dwell time, and historical search interests; Assign a corresponding weight coefficient to each of the user behavior features, and calculate the user preference score for each candidate result in the result ranking list based on the weight coefficient; The result sorting list is sorted in descending order based on the user preference score to obtain the final result list. The top N results in the final result list are regarded as the content that matches the user's search request and are highlighted. Based on the user behavior data, the user preference model is updated according to a preset update cycle.

8. The intelligent retrieval method according to claim 1, characterized in that, The automatic recommendation of similar documents when the user does not find the target document specifically includes: If the final results list is empty, or the user's dwell time on the first N filtered results is less than a preset threshold, or the user searches for semantically similar keywords N times consecutively but does not click on any results, then if any of these conditions are met, it is determined that the user did not find the target document, triggering the recommendation mechanism. The system automatically generates a recommendation list and returns it to the user. The recommendation list includes documents that are semantically similar to the documents the user has previously clicked or commonly used documents from the user's team.

9. The intelligent retrieval method according to claim 1, characterized in that, The real-time monitoring of document status and the dynamic incremental index update are specifically as follows: When a document editing event is detected, the content differences between the current document version and the previous indexed version are compared to identify N target blocks where the content has changed; For each target segment, a new independent index is created in the search engine, while the indexes of document segments whose content has not changed remain unchanged.

10. An intelligent retrieval system for online document management, characterized in that, include: The index building module is used to parse document content and divide the document into blocks. After obtaining document blocks, each block is used as the smallest index unit to build an independent index in the search engine. The index records document information and user permission information and supports multiple language versions. The request processing module is used to receive search requests input by users, verify the permissions of the search requests based on the user permission information in the index, and perform natural language processing and vectorization encoding on the search request content that has passed the permission verification to obtain the content to be retrieved. The document retrieval module is used to perform dual-channel retrieval based on the content to be retrieved, combining keyword inverted index retrieval and vector semantic retrieval, and to merge the dual-channel retrieval results through a fusion algorithm to generate a sorted list of results; The results display module is used to dynamically adjust the result sorting list based on user behavior characteristics, highlight the content that matches the user's search request, and automatically recommend similar documents when the user does not find the target document. The index update module is used to monitor the document status in real time and perform dynamic incremental index updates.

Images