Document content retrieval method and device, equipment and medium

Abstract

The invention relates to the technical field of intelligent decision, and discloses a document content retrieval method and device, equipment and a medium, and the method comprises the steps: recognizing scene terms in a to-be-retrieved document, and carrying out the multi-level tag association storage of the scene terms to obtain a plurality of multi-level term tag combinations; constructing a tag knowledge graph according to the multi-level term tag combination and the to-be-retrieved document; encoding the text retrieval information into a retrieval vector, and calculating cosine similarity between the retrieval vector and a combination vector corresponding to each multi-level term tag combination; selecting a multi-level term tag combination with cosine similarity greater than a threshold value to obtain a screened term tag combination, and generating a candidate retrieval document set; constructing a user portrait according to the user permission attributes and the historical retrieval features; sorting the documents to be retrieved according to the user portrait to obtain an initial retrieval result; feedback information is obtained, the initial retrieval result is corrected through the feedback information, and a final retrieval result is obtained. According to the method, the document content retrieval efficiency can be improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to a document content retrieval method, apparatus, device, and medium. Background Technology

[0002] In the modern management system of group enterprises, rules and regulations serve as crucial bases for standardizing business processes and ensuring compliant operations. Their management effectiveness directly impacts the enterprise's decision-making efficiency and risk control capabilities. With the continuous expansion of enterprise scale and the diversification of business, the number of policy documents is rapidly increasing, with large group enterprises typically having thousands to tens of thousands of policies. Furthermore, these policy documents often contain a large number of enterprise-specific terms (such as business abbreviations and job-specific expressions) and hierarchical management requirements (such as multi-level permission adaptation between subsidiaries, departments, and positions), making the retrieval and application of these policies extremely complex.

[0003] Current policy retrieval primarily relies on three types of technologies: first, keyword-based search engines; second, semantic optimization using deep learning models; and third, associative retrieval based on knowledge graphs. While these technologies are effective in standardized scenarios, they have shortcomings in group-type enterprise environments. General models cannot recognize enterprise-specific terminology, leading to mismatches between search terms and clause descriptions. This makes it difficult to achieve accurate adaptation by combining organizational hierarchy and job authority, thus reducing retrieval efficiency and intelligence.

[0004] Therefore, existing technologies for retrieving rules and regulations in a group enterprise environment suffer from limited application scenarios for general models and low efficiency in document content retrieval. Summary of the Invention

[0005] This invention provides a document content retrieval method, apparatus, device, and medium to solve the technical problem of document content retrieval.

[0006] Firstly, a document content retrieval method is provided, including: The system identifies scene terms within the document to be retrieved based on a preset target scenario, and performs multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. Based on the scenario terms, obtain the document association relationship between the document to be retrieved and the multi-level term tag combination, and construct a tag knowledge graph by taking the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges; The text retrieval information of the target user obtained in advance is encoded into a retrieval vector, and the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is calculated. Select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain filtered term tag combinations, and gather the documents to be retrieved corresponding to the filtered term tag combinations into a candidate retrieval document set; Construct user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance; Based on the user profile, the documents to be retrieved in the candidate retrieval document set are sorted to obtain initial retrieval results; Based on the initial search results, obtain the feedback information from the target user, and use the feedback information to correct the initial search results to obtain the final search results for the text search information.

[0007] Secondly, a document content retrieval device is provided, comprising: The term tag combination association module is used to identify scene terms in the document to be retrieved according to a preset target scene, and to perform multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. The tag knowledge graph construction module is used to obtain the document association relationship between the document to be retrieved and the multi-level term tag combination based on the scenario terms, and to construct a tag knowledge graph by using the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges. The similarity calculation module is used to encode the pre-acquired text retrieval information of the target user into a retrieval vector, and calculate the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph. The candidate retrieval document set generation module is used to select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain a filtered term tag combination, and to collect the documents to be retrieved corresponding to the filtered term tag combination into a candidate retrieval document set. The user profile building module is used to build user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance. The initial search result sorting module is used to sort the documents to be searched in the candidate search document set according to the user profile to obtain the initial search results; The search result correction module is used to obtain feedback information from the target user based on the initial search results, and use the feedback information to correct the initial search results to obtain the final search results of the text search information.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the document content retrieval method described above.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the document content retrieval method described above.

[0010] The aforementioned document content retrieval method, apparatus, device, and medium can identify scenario terms from the documents to be retrieved through a client, and perform multi-level tag association and storage of scenario terms according to a preset tag hierarchy. A tag knowledge graph is constructed based on the term tag combination and the documents to be retrieved. The user-input retrieval information is extracted, encoded into a retrieval vector, and matched with the tag vector in the knowledge graph. A relevant document set is then selected from the documents to be retrieved. Candidate documents are intelligently sorted based on user profile characteristics. User feedback on the retrieval results is collected, and the retrieval results are optimized, thus improving the efficiency of document content retrieval. In this invention, the user's retrieval intent is accurately matched with scenario-based term tags, quickly filtering out a highly relevant document candidate set. User profiles are integrated for sorting, ensuring that the results match the user's hierarchy and historical retrieval preferences. A user feedback loop is introduced to optimize the retrieval results, solving the problem of inaccurate term identification in traditional keyword retrieval in group enterprise scenarios, and improving the accuracy and efficiency of document content retrieval. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an application environment for a document content retrieval method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a document content retrieval method according to an embodiment of the present invention; Figure 3 yes Figure 2 A flowchart illustrating a specific implementation method of step S1; Figure 4 yes Figure 2 A flowchart illustrating a specific implementation of step S3; Figure 5 This is a schematic diagram of a document content retrieval device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 7This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The document content retrieval method provided in this embodiment of the invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can utilize the client's multi-level tag association and knowledge graph construction to achieve structured semantic organization of the documents to be retrieved, accurately interpreting the user's multi-dimensional search intent. By combining semantic vector matching and user profile ranking, it ensures that search results highly match the user's level, permissions, and historical search preferences, concentrating accurate results at the top of the homepage, improving search efficiency and accuracy. A user feedback-driven closed-loop optimization mechanism is introduced to optimize the search results, further enhancing the practicality and intelligence of document retrieval. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0015] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the document content retrieval method provided in this embodiment of the invention includes the following steps: S1. Identify scene terms in the document to be retrieved according to the preset target scene, and perform multi-level tag association storage on the scene terms according to the preset tag hierarchy relationship to obtain multiple multi-level term tag combinations.

[0016] In this embodiment of the invention, the target scenario can be a specific business type and its sub-field involved in enterprise business management, such as "financial reimbursement," "administrative attendance," and "procurement management." Each scenario corresponds to a series of policy documents and business terms. The document to be retrieved can be a multi-source structured text collection related to the target scenario within the enterprise, such as historical policy documents from the past 5 years, standardized business manuals of various departments, and cross-system interface documents (such as a legal compliance system terminology table and OA process node descriptions). The scenario terms are professional terms or phrases used in specific business scenarios, such as business abbreviations (e.g., "standardization"), job-specific terms (e.g., "verification"), and cross-system related terms (e.g., "attendance filing").

[0017] In detail, by organizing scenario terms hierarchically through a structured tagging system, accurate semantic mapping from business type to specific user level can be achieved, improving the hierarchical matching capability of document retrieval and avoiding retrieval bias caused by terminological ambiguity or scenario vagueness.

[0018] In the embodiments of the present invention, see Figure 3 As shown, the step of identifying scene terms within the document to be retrieved based on a preset target scene includes: S31. Label the text content of the document to be retrieved within the preset target scene with a tag sequence to obtain a tag sequence; S32. Extract the characters corresponding to the preset type labels in the label sequence to generate a candidate term set; S33. Calculate the semantic similarity between all terms in the candidate term set, and combine the terms whose semantic similarity is greater than a preset threshold to obtain multiple term combinations; S34. Obtain the terminology mapping table in the target scene, map multiple terms in the terminology combination to standard terms according to the terminology mapping table, and summarize all standard terms to obtain the scene terms in the target scene.

[0019] In detail, a pre-defined sequence labeling model (such as BiLSTM-CRF) can be used to traverse and label documents. The BiLSTM layer fully captures the contextual semantic dependencies of terms in long texts through bidirectional encoding, solving the boundary recognition problem of complex expressions such as "subsidiary A travel exception". The CRF layer ensures that the labeled sequence conforms to the terminology structure rules based on predefined label transfer constraints (such as B-SCENE must be followed by I-SCENE or E-SCENE). The sequence labeling model is jointly trained with the company's historical labeled corpus (such as 30,000 labeled policy documents) and general corpus. By adjusting the hidden layer dimension (such as 128 dimensions) and the CRF regularization coefficient (such as 0.01), the recognition accuracy of complex business terms and long text expressions is significantly improved (F1≥0.92), thereby achieving automated and high-precision preliminary localization of potential terms in the target scenario.

[0020] Specifically, the preset type tags are a predefined entity classification system designed to cover the core knowledge units that need to be identified and managed in the policy documents. Examples include SCENE (business scenario), such as "attendance exception"; ORG (organizational structure), such as "subsidiary A"; POST (job role), such as "cashier"; PROC (business process), such as "project approval process"; and TERM (business terminology / abbreviation), such as "standard adoption". When applying the model, the B (start), I (internal), E (end), S (single word), and O (external) annotation methods are combined with the above types to form specific tags such as B-SCENE, I-SCENE, and E-SCENE, used to structurally extract the corresponding type of terminology entities from the text. Based on the annotation results (such as continuous segments of B-SCENE / I-SCENE / E-SCENE), the corresponding characters are merged to form initial term units (such as "attendance exception" and "travel reimbursement"). The semantic similarity between terms in the candidate term set is further calculated. For example, the Word2Vec model (window size 5, vector dimension 200) or the Sentence-BERT model is used to map the terms to semantic vectors and calculate the cosine similarity. Terms with semantic similarity greater than a preset threshold (such as 0.85) are clustered to obtain multiple synonym or near-synonymous term combinations (such as "reimbursement", "expense reimbursement", and "expense verification" are grouped together).

[0021] Next, a terminology mapping table for the target scenario is obtained. Based on the terminology mapping relationships in the table, multiple terms in the terminology combination are mapped to standard terms. This terminology mapping table includes a pre-defined "business abbreviation-standard term" lookup table (e.g., "adopting standards" → "procurement standard system"), a job-specific terminology mapping table, and a normalized table of approved synonyms. The terminology combination is iterated. If an abbreviation or synonym defined in the mapping table exists, it is uniformly replaced with the corresponding standard term. For new abbreviations not included in the mapping table (e.g., abbreviations for emerging businesses), the mapping table is dynamically updated after review by relevant departments through contextual semantic analysis or by triggering a user feedback process, ensuring that the terminology mapping relationship is synchronized with the company's business evolution.

[0022] Furthermore, the mapped standard terms are merged with the initially identified non-abbreviated scenario terms, and duplicates are removed to form a structured target scenario terminology library. The source document, location of occurrence, and frequency information for each term are recorded. In addition, new terms confirmed by relevant departments can be added to the terminology library in real time through a review process, ultimately generating an iterative and associative scenario terminology knowledge system, providing an accurate and consistent terminology foundation for subsequent hierarchical retrieval and ranking.

[0023] Furthermore, according to the tag hierarchy of "Level 1 Business Type → Level 2 Sub-Scenario → Level 3 Position", corresponding tags are matched to scenario terms one by one to form tag combinations. The preset tag hierarchy can be a three-level progressive tag system pre-built by the enterprise, namely "Level 1 Business Type → Level 2 Sub-Scenario → Level 3 Position". Level 1 tags cover 8-12 core business areas, Level 2 tags cover 30-50 sub-business processes, and Level 3 tags correspond to 50-100 specific positions. Multi-level term tag combinations refer to the structured tag combinations formed after mapping scenario terms according to the preset tag hierarchy. The same scenario term can generate multiple tag combinations due to its applicability to multiple positions (e.g., "travel expense reimbursement" corresponds to "Finance → Reimbursement → Subsidiary B Cashier" and "Finance → Reimbursement → Subsidiary A Accountant").

[0024] S2. Obtain the document association relationship between the document to be retrieved and the multi-level term tag combination based on the scenario terms, and construct a tag knowledge graph by taking the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges.

[0025] In this embodiment of the invention, the document association relationship refers to the correspondence between the document to be retrieved and the multi-level term tag combination established based on scenario terms. The tag knowledge graph refers to a knowledge graph constructed with multi-level term tag combinations, the document to be retrieved, and scenario terms as core nodes, and document association relationships, tag hierarchy relationships, and term-tag mapping relationships as edges.

[0026] In detail, by constructing a multi-dimensional and interconnected tag knowledge graph, the isolation of terms, tags, and documents in traditional retrieval is avoided, and deep binding and hierarchical storage of the three are achieved, thereby improving the efficiency of related queries during retrieval.

[0027] In detail, the process involves traversing the documents to be retrieved, extracting the scenario terms contained within, and binding documents containing the same term with all corresponding multi-level term tag combinations to form preliminary associations. Text semantic analysis is then used to identify the core chapter positions (e.g., titles) of the scenario terms within the documents, and the association strength between the terms and documents is marked (core association weight 0.8, ordinary association weight 0.3). Simultaneously, considering the document's effective date and applicable organization, associations that do not match the scope of application of the tag combinations are filtered out (e.g., excluding subsidiary B's tag combination from subsidiary A's document). A structured document association table is generated (including document ID, tag combination ID, term ID, association strength, and scope of application). This method avoids the problems of blindly binding documents and tag combinations and lacking business adaptability in conventional methods, ensuring the accuracy and effectiveness of the associations.

[0028] Specifically, "multi-level term tag combination nodes" (including tag combination ID, hierarchical information, and scope of application) and "document nodes to be retrieved" (including document ID, effective date, organizational affiliation, and core terminology list) can be created in the Neo4j graph database. A connection between the two nodes is established using document association as an edge, and "scenario terminology nodes" (including terminology ID, normalized expression, and ambiguity annotation) are created. The node attributes and associated data are imported in batches using Cypher statements to construct parent-child hierarchical edges of "first-level tag - second-level tag - third-level tag", mapping edges of "third-level tag - scenario terminology", and inclusion edges of "scenario terminology - document to be retrieved", forming a multi-level associative tag knowledge graph of "business - scenario - hierarchy - terminology - document".

[0029] The above method avoids the problems of missing node attributes, simple relationships, and lack of hierarchical constraints in conventional methods. It ensures that the graph has both structured query capabilities and supports semantic matching requirements, thereby improving the efficiency and accuracy of subsequent retrieval.

[0030] S3. Encode the pre-acquired text retrieval information of the target user into a retrieval vector, and calculate the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph.

[0031] In the embodiments of the present invention, the text retrieval information refers to the retrieval content input by the target user to obtain relevant documents, which can be text input (such as "Subsidiary B cashier reimbursement process"), or the user's voice converted into text. The text retrieval information is the core data reflecting the user's retrieval intention and may contain information such as business abbreviations and scenario descriptions.

[0032] In the embodiments of the present invention, refer to Figure 4 As shown, encoding the pre-obtained text retrieval information of the target user into a retrieval vector includes: S41. Extract the terms in the pre-obtained text retrieval information of the target user that match the scenario terms as the first retrieval terms; S42. Identify the abbreviated terms in the text retrieval information according to the preset term mapping table, and map the abbreviated terms to the second retrieval terms according to the term mapping relationship in the term mapping table; S43. Perform semantic encoding on the first retrieval terms and the second retrieval terms to obtain the retrieval vector of the target user.

[0033] Specifically, perform word segmentation on the retrieval text, automatically remove stop words such as "of", "process", and "application" that have no substantial business meaning, and further extract the complete business vocabulary that matches the scenario term library as the first retrieval terms. At the same time, identify and map the abbreviated terms in the text based on the preset term mapping table to obtain the corresponding second retrieval terms. For voice recognition conversion or incompletely expressed inputs, the system automatically calls the term completion module to complete the abbreviated terms into standardized expressions based on the context. For example, complete "adopt standards" into "procurement standard system process" to ensure the integrity and semantic coherence of the retrieval intention.

[0034] Specifically, in the process of mapping abbreviated terms to standard terms using the term mapping table, if the retrieved abbreviation is not matched in the table, start the Word2Vec semantic similarity calculation model, select the standard terms with a similarity not lower than 0.85 as the mapping result, and push the unmatched terms to the maintenance closed-loop for dynamic update of the term list. Through the above method, the matching problem of new terms or uncollected terms is solved, and a dynamic iterative closed-loop is also constructed. That is, the system will record all the abbreviated terms that are not successfully matched and automatically push them to the term library maintenance team for continuous update and improvement of the term mapping table, so as to ensure the timeliness and coverage of the term system.

[0035] Next, splice the normalized first retrieval terms and the second retrieval terms into a complete query text, input it into the fine-tuned language model for semantic encoding, and output a retrieval vector with a fixed dimension to provide a structured semantic representation for subsequent hierarchical intention matching and similarity calculation.

[0036] In this embodiment of the invention, the combined vector refers to a semantic vector generated by combining multi-level term tags through fine-tuning the language model encoding. The vector integrates hierarchical semantic information of "first-level business type + second-level subdivision scenario + third-level position".

[0037] In this embodiment of the invention, calculating the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph includes: According to the preset tag hierarchy, the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is split to obtain multiple hierarchical sub-vectors of the combination vector; Calculate the cosine similarity between the retrieval vector and the multiple hierarchical sub-vectors respectively; The multiple hierarchical sub-vectors are weighted and summed according to the preset hierarchical weights to obtain the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph.

[0038] In detail, based on the preset semantic boundaries of the tag hierarchy, the combined vector is split into three independent sub-vectors according to three preset semantic levels (business layer, scenario layer, and job level). For example, dimensions 1-256 correspond to the business layer sub-vectors, responsible for representing the semantic information of the first-level tags; dimensions 257-512 correspond to the scenario layer sub-vectors, representing the semantic content of the second-level tags; and dimensions 513-768 correspond to the user level layer sub-vectors, representing the semantic features of the third-level tags. Through this structured split, the system ensures that each sub-vector can accurately and independently carry the semantic expression of its corresponding level, avoiding mutual interference between semantic information of different levels and improving the hierarchical clarity and semantic integrity of the vector representation. This method avoids the problems of overall vector semantic ambiguity and inability to distinguish hierarchical requirements in conventional methods, ensuring accurate matching of the hierarchical structure of the user's search intent.

[0039] Specifically, the retrieval vector is split into three corresponding retrieval-level sub-vectors according to the same semantic boundaries. The cosine similarity between the retrieval vector and the sub-vectors of each term tag combination at the business, scenario, and job levels is calculated respectively. The cosine similarity formula is (sim(a,b)=a・b / (||a||×||b||)). This method avoids the misjudgment problem of "a mismatch at a certain level but a high overall score" caused by overall vector matching in conventional methods, ensuring accurate matching of requirements at each level.

[0040] Furthermore, the similarity scores of the three levels are weighted and summed according to preset hierarchical weights, such as a weighting scheme of 0.3 for the business level, 0.3 for the scenario level, and 0.4 for the user level, to highlight the priority of user level adaptation in the overall matching, thus deriving the final cosine similarity. Simultaneously, the system introduces a weight penalty mechanism: if the user level similarity is lower than a preset threshold (e.g., 0.5), the final weighted score is additionally multiplied by 0.5 as a penalty, forcibly reducing the ranking position of user level mismatched combinations in the search results. This mechanism ensures the dominant position of the core need level (user level) in similarity calculation, thereby effectively improving the relevance of search results to the actual user level needs, achieving more accurate and business-oriented document filtering and ranking.

[0041] S4. Select the multi-level term tag combination with a cosine similarity greater than a preset similarity threshold to obtain the filtered term tag combination, and gather the documents to be retrieved corresponding to the filtered term tag combination into a candidate retrieval document set.

[0042] In this embodiment of the invention, the filtered term tag combinations are term tag hierarchical structures that are highly relevant to the user's search intent after semantic matching and filtering, typically containing term tag paths from first level to multiple levels. The candidate retrieval document set refers to the set of all documents to be retrieved that are associated with the filtered term tag combinations, obtained by back-mapping document association relationships in the tag knowledge graph.

[0043] In detail, by filtering by similarity threshold, the search scope is narrowed and documents that do not match the user's needs are removed, which not only improves the efficiency of subsequent sorting, but also avoids irrelevant documents occupying display resources.

[0044] Specifically, the final cosine similarity of each multi-level term tag combination is compared with a preset similarity threshold (e.g., a similarity threshold of 0.75 based on enterprise scenarios). Combinations with similarity greater than the threshold are selected. Next, the system obtains the target user's level ID through the permission management interface and matches it with the third-level level tags in the multi-level term tag combinations. Combinations containing tags inconsistent with the user's level are eliminated (e.g., when the user's level is "Subsidiary B Cashier," combinations containing non-matching level tags such as "Subsidiary A Accountant" are automatically excluded). From the remaining matching combinations, they are sorted by similarity from high to low, and only the top three combinations are retained as the core filtering results, yielding the filtered term tag combinations. This method avoids the problems of mismatched search results and irrelevant combinations caused by simply filtering based on similarity in conventional methods, ensuring that the filtered combinations match the user's search intent.

[0045] Furthermore, based on the combination of filtering term tags, all associated documents to be retrieved are extracted from the knowledge graph. Filtering and optimization operations are then performed on all documents, such as deduplication based on document ID to prevent duplicate appearances of the same document due to multiple tag associations. Documents inaccessible to the target user are removed based on their organizational level and functional scope. For example, in a group policy retrieval scenario, sensitive group-level documents inaccessible to subsidiary users are excluded. Expired or abolished documents are automatically filtered based on their effective date and repeal status, ensuring only currently valid document text is retained. The final candidate document set records the corresponding combination of filtering term tags, similarity score, and association strength in the knowledge graph for each document. This method avoids redundancy in candidate documents and the inclusion of invalid or inaccessible documents, ensuring the conciseness and effectiveness of the candidate set.

[0046] S5. Construct a user profile based on the user permission attributes and historical search characteristics of the target user obtained in advance.

[0047] In this embodiment of the invention, the user permission attributes are basic attributes of the target user obtained from the permission management system, including the organization identifier (e.g., "subsidiary policy document B"), user level (e.g., "administrative specialist" or "cashier supervisor"), and access permissions (e.g., group-level policy document / subsidiary-level policy document / department-level policy document). These are core data defining the user's search permissions and scope of application. The historical search features refer to behavioral features extracted from the user's search logs over the past three months, including high-frequency search terms (e.g., "attendance filing" or "travel expense reimbursement"), user behavior features (click count, dwell time ≥ 30 seconds considered valid interaction, and collection operations), and search scenario features (e.g., administrative users' "attendance policy document / travel policy document" scenario accounts for 70% of policy documents). The user profile refers to a 128-dimensional structured feature vector generated by weightedly fusing user permission attributes and historical search features through an attention mechanism. This vector comprehensively represents the user's personalized search needs, permission boundaries, and search characteristics.

[0048] In this embodiment of the invention, the step of constructing a user profile based on the pre-acquired user permission attributes and historical search features of the target user includes: Obtain the target user's organization identifier, user level, and access permissions; weight and fuse the vectors corresponding to the organization identifier, user level, and access permissions according to preset attribute weights to generate the target user's user permission attribute vector. Extract the target user's historical search data for the searched document, and identify the search terms, user behavior characteristics, and search scenario characteristics in the historical search data; The vectors corresponding to the search terms, user behavior features, and search scenario features are weighted and fused according to preset search weights to obtain the historical search feature vector of the target user. The user permission attribute vector and the historical retrieval feature vector are weighted and fused according to preset user feature weights to obtain the user profile of the target user.

[0049] In detail, three core attributes of the target user—organizational identifier, user level, and access permissions—are extracted from the access control system. A tagging technique is used to convert these non-numerical attributes into numerical vectors; for example, the job title "Administrative Specialist" is coded as the number 15. Then, using a preset weighting of "user level 0.4, organization identifier 0.3, access permissions 0.3," an attention mechanism is employed to weight and fuse the three vectors, generating a one-dimensional user permission attribute vector. The user level represents the target user's job position, responsibilities, and management authority within the enterprise's organizational structure. Different user levels correspond to different document viewing, editing, approval, and retrieval permissions.

[0050] Specifically, historical search data of target users is extracted from the search log database. Relevant features are extracted using term frequency statistics and behavioral statistics methods. For example, a 30-day sliding window is used to extract users' search records for the past three months. Term frequency statistics are used to identify search terms that appear ≥3 times. Behavioral statistics are used to extract user behavior features such as the number of times users click on documents, the duration of their stay, and their favorites. Based on the high-frequency search terms and the tag attributes of the documents clicked by users, the user's preference coefficient in various business scenarios, i.e., the search scenario feature, is further calculated. For example, the "attendance scenario preference coefficient" equals the number of times the user interacts with attendance-related documents divided by the total number of effective interactions. This method avoids the problems of preference lag and single feature dimensions caused by the excessively broad time range of historical feature extraction in conventional methods, ensuring accurate capture of users' recent search preferences.

[0051] Next, the extracted features are vectorized and fused. For example, the search terms are converted into semantic vectors through the Word2Vec model; the user behavior features are normalized by Min-Max and mapped to the 0-1 interval. The above vectors are weighted and summed according to the preset weights of "search scenario features 0.4, user behavior features 0.3, and search term semantic vectors 0.3" to generate historical search feature vectors.

[0052] Furthermore, the user permission attribute vector and the historical retrieval feature vector are weighted and fused to generate the final user profile. This is done by configuring the user feature weights as "user permission attribute 0.4, historical retrieval feature 0.6" and using an attention layer to weight and sum the two vectors. The fusion result is then standardized (mapped to the 0-1 range) to generate a user profile feature vector.

[0053] S6. Sort the documents to be retrieved in the candidate retrieval document set according to the user profile to obtain the initial retrieval results.

[0054] In this embodiment of the invention, the initial retrieval result is a list of documents formed by sorting the candidate retrieval document set and arranging them in descending order of score.

[0055] In this embodiment of the invention, the step of sorting the documents to be retrieved in the candidate retrieval document set according to the user profile to obtain initial retrieval results includes: The number of terms with the same scene terms in the candidate document set and the text search information is counted, and the number of terms is multiplied by a preset matching score to obtain a first matching score; Based on the user profile, calculate the second matching score between the combination vector of the multi-level term tag combination corresponding to the document to be retrieved and the retrieval vector of the target user; Calculate the third matching score between the document to be retrieved in the candidate retrieval document set and the historical retrieval document corresponding to the historical retrieval features in the user profile; The first matching score, the second matching score, and the third matching score are weighted and fused according to the preset ranking weights to obtain the retrieval score of the document to be retrieved in the candidate retrieval document set; The documents to be retrieved are sorted according to the retrieval scores to obtain initial retrieval results.

[0056] In detail, each document in the candidate document set is traversed, and the number of scenario terms that are completely consistent with the user's text search information is counted. Each matching term is added 0.2 points according to preset rules. At the same time, the number of synonyms with a semantic similarity of ≥0.85 determined by the Word2Vec model is also included in the total score (the first matching score for a single document does not exceed 1 point). Finally, the first matching score is obtained. This method avoids the problem of conventional text matching methods that only count the number of keywords and ignore the importance of terms and synonyms. It ensures that the first matching score accurately reflects the textual relevance between the document and the search information, thus improving the accuracy of text matching.

[0057] Specifically, by comparing the organizational identifier in the user profile with the organizational identifier corresponding to the user level of the candidate document, the degree of overlap of the two tag combinations at the level is calculated. If they completely overlap, 1 point is awarded; if they only overlap to the second-level tag, 0.7 points are awarded; and if they only overlap to the first-level tag, 0.3 points are awarded. If the applicable organization of the document is the same as the organization to which the user belongs, an additional 0.5 points are awarded. The upper limit of the second matching score is 1.5 points, and the final second matching score is obtained.

[0058] Next, the top 10 most clicked documents are extracted from the user profile as references. The semantic similarity between the candidate documents and these documents is calculated, and documents with a dwell time of more than 3 minutes are given double weight. Finally, the third matching score is controlled between 0 and 1 points to accurately reflect the user's preference for similar documents and make the ranking results closer to the user's real usage habits and focus.

[0059] Furthermore, the score is weighted according to the ratio of "first matching score weight 0.25, second matching score weight 0.35, and third matching score weight 0.4", and the final search score is proportionally amplified to a range of 0-100 points, thereby ensuring that the comprehensive score is highly consistent with the user's personalized needs, user level and business scenario.

[0060] Next, the candidate documents are sorted from highest to lowest based on their search scores, and the sorting results are validated for accuracy and labeled. An initial search results list is generated to ensure that more than 80% of the top 5 documents on the homepage fully match the user's level, organization, and scenario needs. At the same time, obsolete documents are marked with a red "obsolete" label and moved to the end of the list, and documents with high core clause matching are marked with a "core clause match" label. Finally, a sorted list with multi-dimensional labels is output, which significantly improves the user's search efficiency and the readability of the results.

[0061] S7. Obtain feedback information from the target user based on the initial search results, and use the feedback information to correct the initial search results to obtain the final search results of the text search information.

[0062] In this embodiment of the invention, the feedback information refers to the target user's evaluation and suggestions on the initial search results, including search success information, search failure information (clicking the corresponding button on the results page), terminology missing markers (unrecognized terms are reported through a dedicated entry point), supplementary explanations, etc., and is the core data for iteratively optimizing search results, terminology database, and ranking model. The final search results refer to the list of documents that accurately match the user's needs after correcting the initial search results based on user feedback.

[0063] In this embodiment of the invention, the step of using the feedback information to correct the initial search result to obtain the final search result of the text search information includes: If the feedback information is a retrieval failure message, the ranking weight corresponding to the initial retrieval result is adjusted according to the feedback information to obtain an updated ranking weight; The initial search results are determined based on the updated ranking weights, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information is missing term information, then the updated term is extracted from the feedback information and incorporated into the scene terminology to obtain the updated scene terminology; The initial search results are determined based on the updated scenario terms, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information indicates a successful retrieval, the initial retrieval result will be determined as the final retrieval result.

[0064] In detail, by recognizing users' continuous feedback of marking documents as useless (such as "useless" or "irrelevant"), the system automatically lowers the ranking weight of the document. If a document is marked as useless three or more times in consecutive searches, the feedback information is recorded as a search failure message, and the system will dynamically lower its weight parameter in the ranking model. At the same time, the reason, time and magnitude of the adjustment are recorded in the weight adjustment log to ensure that the adjustment process is traceable. The system continuously optimizes the ranking strategy through the feedback mechanism to avoid the recurrence of low-quality documents, thereby improving the accuracy of search results and user satisfaction.

[0065] Specifically, when users provide specific feedback such as "job mismatch," "scenario mismatch," or "institution mismatch," the feature weights in the ranking model are adjusted accordingly based on the feedback type. For example, feedback "user level mismatch" lowers the weight of the job suitability feature by 0.2; feedback "scenario mismatch" lowers the weight of the scenario tag overlap rate in the second matching degree by 0.15; and feedback "institution mismatch" lowers the weight of the institution matching degree by 0.1. Each adjustment is recorded synchronously in the weight adjustment log, and the ranking model parameters are updated in real time to ensure that the model can dynamically adapt to the actual needs of users and changes in business scenarios, improving the personalization and usability of search results. The candidate search document set is then re-ranked using the updated ranking weights to generate new initial search results.

[0066] Next, terms not recognized by the current scenario terminology database are extracted from user feedback and added to the database to improve the knowledge system. Updated terms can be extracted from feedback using text extraction algorithms (such as regular expression keyword matching), incrementally validated and fine-tuned using a BiLSTM-CRF model, and then pushed to the relevant business department for review. If approved, the system adds the updated terms to the scenario terminology database and simultaneously updates the terminology mapping table and tag mapping relationships; if not approved, modification suggestions are returned and the user who provided feedback is notified, thus ensuring the timeliness and accuracy of the terminology database and supporting more precise document association and retrieval.

[0067] Furthermore, for initial search results marked as "useful" by the user, the feedback information is recorded as a successful search, and this initial search result is determined as the final search result. Simultaneously, the weights of relevant features (such as user preference and tag overlap rate) of this document in the ranking model are increased by 0.1. At the same time, the document ID, tag combination ID, user profile features, and search score are associated and stored in the iteration log library. This effectively improves user trust in the search system and enhances the user experience.

[0068] As can be seen, the above solution achieves accurate structured representation of the document to be retrieved by storing multi-level tags associated with scene terms and constructing a tag knowledge graph. It also filters candidate document sets by combining cosine similarity calculation, sorts them based on user profiles, and corrects the search results by combining user feedback in a closed loop. This not only improves the matching accuracy and efficiency of text retrieval, but also dynamically adapts to user search needs, ensuring that the final search results are highly relevant. It effectively solves the problems of fuzzy tag hierarchy and insufficient matching of user needs in traditional retrieval.

[0069] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0070] In one embodiment, a document content retrieval device is provided, which corresponds one-to-one with the document content retrieval methods described in the above embodiments. For example... Figure 5 As shown, the document content retrieval device 100 includes a term tag combination association module 101, a tag knowledge graph construction module 102, a similarity calculation module 103, a candidate retrieval document set generation module 104, a user profile construction module 105, an initial retrieval result ranking module 106, and a retrieval result correction module 107. Detailed descriptions of each functional module are as follows: The term tag combination association module 101 is used to identify scene terms in the document to be retrieved according to a preset target scene, and to perform multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. The tag knowledge graph construction module 102 is used to obtain the document association relationship between the document to be retrieved and the multi-level term tag combination based on the scenario terms, and to construct a tag knowledge graph by using the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges. The similarity calculation module 103 is used to encode the pre-acquired text retrieval information of the target user into a retrieval vector, and calculate the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph. The candidate retrieval document set generation module 104 is used to select the multi-level term tag combination with a cosine similarity greater than a preset similarity threshold, obtain the filtered term tag combination, and collect the documents to be retrieved corresponding to the filtered term tag combination into a candidate retrieval document set. User profile building module 105 is used to build user profiles based on the user permission attributes and historical search features of the target user obtained in advance. The initial search result sorting module 106 is used to sort the documents to be searched in the candidate search document set according to the user profile to obtain the initial search results; The search result correction module 107 is used to obtain feedback information from the target user based on the initial search results, and use the feedback information to correct the initial search results to obtain the final search results of the text search information.

[0071] In one embodiment, the terminology tag combination association module 101, when performing the task of identifying scene terms in the document to be retrieved based on a preset target scene, is used to: Tag sequences are generated by labeling the text content in the document to be retrieved within the preset target scene. Extract the characters corresponding to the preset type tags in the tag sequence to generate a candidate term set; Calculate the semantic similarity between all terms in the candidate term set, and combine terms whose semantic similarity is greater than a preset threshold to obtain multiple term combinations; Obtain a terminology mapping table for the target scene, map multiple terms in the terminology combination to standard terms based on the terminology mapping table, and summarize all standard terms to obtain the scene terms in the target scene.

[0072] In one embodiment, the similarity calculation module 103, when encoding the pre-acquired text retrieval information of the target user into a retrieval vector, is used to: Extract the terminology that matches the scenario terminology from the pre-acquired text retrieval information of the target user and use it as the first retrieval term; The abbreviations in the text retrieval information are identified according to a preset term mapping table, and the abbreviations are mapped to second retrieval terms according to the term mapping relationship in the term mapping table. Semantic encoding is performed on the first search term and the second search term to obtain the search vector of the target user.

[0073] In one embodiment, the similarity calculation module 103, when calculating the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph, is used to: According to the preset tag hierarchy, the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is split to obtain multiple hierarchical sub-vectors of the combination vector; Calculate the cosine similarity between the retrieval vector and the multiple hierarchical sub-vectors respectively; The multiple hierarchical sub-vectors are weighted and summed according to the preset hierarchical weights to obtain the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph.

[0074] In one embodiment, the user profile building module 105, when performing the process of building a user profile based on the pre-acquired user permission attributes and historical search features of the target user, is used to: Obtain the target user's organization identifier, user level, and access permissions; weight and fuse the vectors corresponding to the organization identifier, user level, and access permissions according to preset attribute weights to generate the target user's user permission attribute vector. Extract the target user's historical search data for the searched document, and identify the search terms, user behavior characteristics, and search scenario characteristics in the historical search data; The vectors corresponding to the search terms, user behavior features, and search scenario features are weighted and fused according to preset search weights to obtain the historical search feature vector of the target user. The user permission attribute vector and the historical retrieval feature vector are weighted and fused according to preset user feature weights to obtain the user profile of the target user.

[0075] In one embodiment, the initial search result sorting module 106, when performing the sorting of documents to be retrieved in the candidate search document set according to the user profile to obtain initial search results, is used to: The number of terms with the same scene terms in the candidate document set and the text search information is counted, and the number of terms is multiplied by a preset matching score to obtain a first matching score; Based on the user profile, calculate the second matching score between the combination vector of the multi-level term tag combination corresponding to the document to be retrieved and the retrieval vector of the target user; Calculate the third matching score between the document to be retrieved in the candidate retrieval document set and the historical retrieval document corresponding to the historical retrieval features in the user profile; The first matching score, the second matching score, and the third matching score are weighted and fused according to the preset ranking weights to obtain the retrieval score of the document to be retrieved in the candidate retrieval document set; The documents to be retrieved are sorted according to the retrieval scores to obtain initial retrieval results.

[0076] In one embodiment, the retrieval result correction module 107, when correcting the initial retrieval result using the feedback information to obtain the final retrieval result of the text retrieval information, is configured to: If the feedback information is a retrieval failure message, the ranking weight corresponding to the initial retrieval result is adjusted according to the feedback information to obtain an updated ranking weight; The initial search results are determined based on the updated ranking weights, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information is missing term information, then the updated term is extracted from the feedback information and incorporated into the scene terminology to obtain the updated scene terminology; The initial search results are determined based on the updated scenario terms, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information indicates a successful retrieval, the initial retrieval result will be determined as the final retrieval result.

[0077] This invention provides a document content retrieval device that achieves accurate structured representation of the document to be retrieved through multi-level tag association storage of scene terms and construction of a tag knowledge graph. It combines cosine similarity calculation to filter candidate document sets, and then completes the sorting based on user profiles. The retrieval results are corrected in a closed loop based on user feedback. This not only improves the matching accuracy and efficiency of text retrieval, but also dynamically adapts to user retrieval needs, ensuring that the final retrieval results are highly relevant. It effectively solves the problems of fuzzy tag hierarchy and insufficient matching of user needs in traditional retrieval.

[0078] For specific limitations regarding the document content retrieval device, please refer to the limitations on the document content retrieval method above, which will not be repeated here. Each module in the aforementioned document content retrieval device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0079] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a document content retrieval method on the server side.

[0080] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of a document content retrieval method.

[0081] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The system identifies scene terms within the document to be retrieved based on a preset target scenario, and performs multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. Based on the scenario terms, obtain the document association relationship between the document to be retrieved and the multi-level term tag combination, and construct a tag knowledge graph by taking the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges; The text retrieval information of the target user obtained in advance is encoded into a retrieval vector, and the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is calculated. Select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain filtered term tag combinations, and gather the documents to be retrieved corresponding to the filtered term tag combinations into a candidate retrieval document set; Construct user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance; Based on the user profile, the documents to be retrieved in the candidate retrieval document set are sorted to obtain initial retrieval results; Based on the initial search results, obtain the feedback information from the target user, and use the feedback information to correct the initial search results to obtain the final search results for the text search information.

[0082] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The system identifies scene terms within the document to be retrieved based on a preset target scenario, and performs multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. Based on the scenario terms, obtain the document association relationship between the document to be retrieved and the multi-level term tag combination, and construct a tag knowledge graph by taking the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges; The text retrieval information of the target user obtained in advance is encoded into a retrieval vector, and the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is calculated. Select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain filtered term tag combinations, and gather the documents to be retrieved corresponding to the filtered term tag combinations into a candidate retrieval document set; Construct user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance; Based on the user profile, the documents to be retrieved in the candidate retrieval document set are sorted to obtain initial retrieval results; Based on the initial search results, obtain the feedback information from the target user, and use the feedback information to correct the initial search results to obtain the final search results for the text search information.

[0083] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0086] It should be noted that if any software tools or components not belonging to our company appear in the embodiments of this application, they are merely for illustrative purposes and do not represent actual use.

[0087] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A document content retrieval method, characterized in that, include: The system identifies scene terms within the document to be retrieved based on a preset target scene, and performs multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship, resulting in multiple multi-level term tag combinations. Based on the scenario terms, obtain the document association relationship between the document to be retrieved and the multi-level term tag combination, and construct a tag knowledge graph by taking the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges; The text retrieval information of the target user obtained in advance is encoded into a retrieval vector, and the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is calculated. Select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain filtered term tag combinations, and gather the documents to be retrieved corresponding to the filtered term tag combinations into a candidate retrieval document set; Construct user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance; Based on the user profile, the documents to be retrieved in the candidate retrieval document set are sorted to obtain initial retrieval results; Based on the initial search results, obtain the feedback information from the target user, and use the feedback information to correct the initial search results to obtain the final search results for the text search information.

2. The document content retrieval method as described in claim 1, characterized in that, The step of identifying scene terms within the document to be retrieved based on a preset target scene includes: Tag sequences are generated by labeling the text content in the document to be retrieved within the preset target scene. Extract the characters corresponding to the preset type tags in the tag sequence to generate a candidate term set; Calculate the semantic similarity between all terms in the candidate term set, and combine terms whose semantic similarity is greater than a preset threshold to obtain multiple term combinations; Obtain a terminology mapping table for the target scene, map multiple terms in the terminology combination to standard terms based on the terminology mapping table, and summarize all standard terms to obtain the scene terms in the target scene.

3. The document content retrieval method as described in claim 1, characterized in that, The step of encoding the pre-acquired text retrieval information of the target user into a retrieval vector includes: Extract the terminology that matches the scenario terminology from the pre-acquired text retrieval information of the target user and use it as the first retrieval term; The abbreviations in the text retrieval information are identified according to a preset term mapping table, and the abbreviations are mapped to second retrieval terms according to the term mapping relationship in the term mapping table. Semantic encoding is performed on the first search term and the second search term to obtain the search vector of the target user.

4. The document content retrieval method as described in claim 1, characterized in that, The calculation of the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph includes: According to the preset tag hierarchy, the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph is split to obtain multiple hierarchical sub-vectors of the combination vector; Calculate the cosine similarity between the retrieval vector and the multiple hierarchical sub-vectors respectively; The multiple hierarchical sub-vectors are weighted and summed according to the preset hierarchical weights to obtain the cosine similarity between the retrieval vector and the combined vector corresponding to each multi-level term tag combination in the tag knowledge graph.

5. The document content retrieval method as described in claim 1, characterized in that, The step of constructing a user profile based on the pre-acquired user permission attributes and historical search features of the target user includes: Obtain the target user's organization identifier, user level, and access permissions; weight and fuse the vectors corresponding to the organization identifier, user level, and access permissions according to preset attribute weights to generate the target user's user permission attribute vector. Extract the target user's historical search data for the searched document, and identify the search terms, user behavior characteristics, and search scenario characteristics in the historical search data; The vectors corresponding to the search terms, user behavior features, and search scenario features are weighted and fused according to preset search weights to obtain the historical search feature vector of the target user. The user permission attribute vector and the historical retrieval feature vector are weighted and fused according to preset user feature weights to obtain the user profile of the target user.

6. The document content retrieval method as described in claim 1, characterized in that, The step of sorting the documents to be retrieved in the candidate document set according to the user profile to obtain the initial search results includes: The number of terms with the same scene terms in the candidate document set and the text search information is counted, and the number of terms is multiplied by a preset matching score to obtain a first matching score; Based on the user profile, calculate the second matching score between the combination vector of the multi-level term tag combination corresponding to the document to be retrieved and the retrieval vector of the target user; Calculate the third matching score between the document to be retrieved in the candidate retrieval document set and the historical retrieval document corresponding to the historical retrieval features in the user profile; The first matching score, the second matching score, and the third matching score are weighted and fused according to the preset ranking weights to obtain the retrieval score of the document to be retrieved in the candidate retrieval document set; The documents to be retrieved are sorted according to the retrieval scores to obtain initial retrieval results.

7. The document content retrieval method as described in claim 1, characterized in that, The step of using the feedback information to correct the initial search results to obtain the final search results for the text search information includes: If the feedback information is a retrieval failure message, the ranking weight corresponding to the initial retrieval result is adjusted according to the feedback information to obtain an updated ranking weight; The initial search results are determined based on the updated ranking weights, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information is missing term information, then the updated term is extracted from the feedback information and incorporated into the scene terminology to obtain the updated scene terminology; The initial search results are determined based on the updated scenario terms, and the process returns to the step of obtaining feedback information from the target user based on the initial search results. If the feedback information indicates a successful retrieval, the initial retrieval result will be determined as the final retrieval result.

8. A document content retrieval device, characterized in that, include: The term tag combination association module is used to identify scene terms in the document to be retrieved according to a preset target scenario, and to perform multi-level tag association storage on the scene terms according to a preset tag hierarchy relationship to obtain multiple multi-level term tag combinations. The tag knowledge graph construction module is used to obtain the document association relationship between the document to be retrieved and the multi-level term tag combination based on the scenario terms, and to construct a tag knowledge graph by using the multi-level term tag combination and the document to be retrieved as nodes and the document association relationship as edges. The similarity calculation module is used to encode the pre-acquired text retrieval information of the target user into a retrieval vector, and calculate the cosine similarity between the retrieval vector and the combination vector corresponding to each multi-level term tag combination in the tag knowledge graph. The candidate retrieval document set generation module is used to select multi-level term tag combinations with a cosine similarity greater than a preset similarity threshold to obtain a filtered term tag combination, and to collect the documents to be retrieved corresponding to the filtered term tag combination into a candidate retrieval document set. The user profile building module is used to build user profiles based on the user permission attributes and historical search characteristics of the target users obtained in advance. The initial search result sorting module is used to sort the documents to be searched in the candidate search document set according to the user profile to obtain the initial search results; The search result correction module is used to obtain feedback information from the target user based on the initial search results, and use the feedback information to correct the initial search results to obtain the final search results of the text search information.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the document content retrieval method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the document content retrieval method as described in any one of claims 1 to 7.

Images