Intelligent literature interaction reading assisting system based on privacy protection

Abstract

The invention discloses an intelligent literature interaction reading assisting system based on privacy protection. The system integrates a named entity recognition model with BERT and variants thereof (such as RoBERTa) as a core and a domain knowledge graph, unified analysis, semantic indexing and dynamic tree graph navigation are carried out on PDF, audio and video and other multi-modal data, and a literature structure is recombined in a three-level visualization mode; providing multi-granularity abstracts, questions and answers and personalized recommendation by utilizing an RAG architecture and a BART abstract model; according to the method, the front end Vue3 componentization rendering and the rear end Spring Boot micro-service architecture are combined to realize high-concurrency retrieval and distributed storage in combination with Elasticsearch, Redis and MinIO, and the RBAC right control ensures the user data access security; aES-256 password encryption and private space are isolated and parallel through a'space + file 'two-stage encryption system and localized privacy calculation, and literature full-link hierarchical protection is achieved. The term understanding cost is reduced, the reading efficiency is improved, the sensitive data leakage risk is reduced, and an efficient and safe intelligent solution is provided for the whole scientific research process.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of academic research information security and intelligent knowledge services, and more specifically, to an intelligent document interaction and reading assistance system based on privacy protection. Background Technology

[0002] With the explosive growth in the number of academic documents, researchers face severe challenges in reading, extracting knowledge, and writing academic papers. However, current intelligent reading tools have significant shortcomings in privacy protection, content parsing, terminology recognition, and interactive experience.

[0003] In terms of privacy protection, most mainstream tools employ basic HTTPS transmission encryption or simple file-level encryption, failing to build a two-layer "file-space" protection system. This type of single encryption mode is vulnerable to quantum attacks such as Shor's algorithm and internal keyword guessing attacks, exhibiting high latency and poor flexibility, thus exacerbating the risk of sensitive academic data leakage.

[0004] In terms of multimodal document parsing, general parsers generally have low confidence in OCR recognition of scanned PDFs (e.g., less than 80% in native Apache Tika configuration), and lack accurate timeline correlation for text transcribed from audio and video content, making it difficult to integrate cross-modal data and form a unified knowledge view.

[0005] In terms of academic term recognition, the general BERT-NER model, which has not been fine-tuned for specific fields, generally has an F1 score of less than 85% for term recognition, making it difficult to distinguish between "same name, different meaning" terms. At the same time, it lacks an efficient local caching and cloud knowledge base linkage mechanism, resulting in excessively high response delays for low-frequency term queries.

[0006] In terms of navigation and interactive experience, the linear directory navigation used by traditional tools is in serious conflict with the non-linear thinking habits of researchers, which involve "skipping reading - instant annotation - related notes". It cannot automatically generate an interactive tree-like directory based on the internal logical structure of the document, resulting in a break in the connection between "annotated content - chapter structure - personal notes" and making knowledge tracing difficult.

[0007] In summary, existing technologies need to be improved to address issues such as weak privacy protection, fragmented multimodal parsing, insufficient terminology recognition accuracy, and poor navigation interaction experience. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a privacy-protected intelligent document interactive reading aid system.

[0009] According to a first aspect of the present invention, a privacy-protected intelligent document interaction and reading assistance system is provided, which generally includes a multimodal data unified parsing module, a terminology knowledge graph construction module, a two-level encryption privacy protection module, a dynamic tree diagram navigation module, a semantic retrieval and personalized recommendation module, and a microservice architecture support module.

[0010] The dual-level encryption privacy protection module is used to construct a dual-level protection system combining "file-level encryption" and "space-level access control" to encrypt and store user document data and perform secondary verification of user access permissions, achieving end-to-end privacy protection for document data. The dual-level encryption privacy protection module performs the following preprocessing: a front-end form interceptor generates a multi-bit random salt value, which, combined with the user's password hash value, generates a key using the PBKDF2 algorithm; the document data is encrypted in blocks, with an encryption identifier, initialization vector, and salt value information embedded in the file header; when the private space is created, the Spring Security filter generates a JWT token containing user role information, and upon user access, secondary verification is performed, including password verification and token validity check. The unified multimodal data parsing module is used to uniformly parse and standardize academic materials in various formats such as PDF, Word, and audio / video. It extracts structured information such as text, tables, and formulas, and converts audio / video content into indexable text data, forming a unified standardized dataset to provide a data foundation for upper-layer intelligent services. The unified multimodal data parsing module performs the following steps: it calls the Apache Tika parser to identify PDF or Word document types, using magic number detection to eliminate the risk of disguised files; it uses PDFBox to parse the page layout and marks formula content using regular expressions, combining Apache POI to extract table data; it calls decoding for audio / video files, extracts the audio stream, and then uses the Whisper API to transcribe it into timestamped text segments; the parsing results are then segmented and uniformly stored in the standardized dataset. The terminology knowledge graph construction module is used to accurately identify professional terms from standardized data based on a domain-adaptive artificial intelligence model. Through a linkage mechanism between local caching and a cloud-based knowledge base, it constructs an interactive domain knowledge graph, enabling visualization and related queries of terminology knowledge. The terminology knowledge graph construction module executes as follows: a pre-trained BERT model is fine-tuned on a domain corpus to obtain terminology recognition results; the recognition results are first queried from the local Redis cache, and if no match is found, the CNKI open API is called to obtain term definitions and related literature; the front-end uses the Vue3 custom directive `v-term-highlight` to mark text terms, and clicking triggers a floating window component to load the knowledge graph's associated node data. The dynamic tree diagram navigation module automatically generates an interactive tree-like navigation view based on the inherent logical structure of the document, and dynamically associates user annotations, notes, and other actions with the navigation structure, forming an integrated reading and knowledge management interface that integrates "structure-content-notes". The dynamic tree diagram navigation module performs the following steps: parsing the hierarchical relationship of document titles to generate a tree-like data structure; using Vue3 component-based rendering to achieve node collapse or expansion; and synchronously recording user operation logs to an encrypted personalized note library. When a user annotates text, it automatically associates the annotation with the current tree node ID, forming a three-element association index of structure-content-notes. The semantic retrieval and personalized recommendation module integrates keyword matching and deep semantic vector matching, combined with knowledge graphs and user behavior profiles, to provide accurate semantic retrieval results and personalized document content recommendations. The semantic retrieval and personalized recommendation module executes as follows: it uses Elasticsearch to build an inverted index, generates text semantic vectors through pre-trained models such as BERT and RoBERTa, and optimizes the results by integrating ranking models such as LightGBM; when a user searches, the input text is segmented and simultaneously triggers keyword retrieval and vector similarity matching, and the results are re-ranked based on entity associations in the knowledge graph.

[0011] The microservice architecture support module is used to decouple the core functions of the system into independent microservices. Through containerization and cluster orchestration technologies, it achieves high availability, elastic scaling, and fault tolerance and self-healing of the system. The microservice architecture support module executes as follows: it uses Spring Boot to split the system into authentication, parsing, and retrieval services, employs Docker containerization deployment and Kubernetes automatic scaling, and automatically adds service instances when concurrent parsing tasks exceed a set threshold.

[0012] Through the collaborative work described above, each module works together to achieve a full-process reading assistance function, from secure document uploading, intelligent parsing, knowledge association to personalized interaction.

[0013] According to a second aspect of the present invention, a privacy-preserving intelligent document interaction aid method is provided. The method includes: S1, Two-level encryption and access control steps: The system receives literature data uploaded by users, generates a random salt value, and uses a key derivation algorithm with the user's password to generate an encryption key. The system uses the key to encrypt the literature data in blocks and embeds an initialization vector and salt value in the file header. At the same time, it generates an access token containing the scope of literature access and operation permissions. When a user accesses encrypted data or private space, the system performs secondary verification through a password and the access token. S2, Unified Analysis and Standardization Steps for Multimodal Data: The system parses encrypted PDF, Word documents, and audio / video multimodal data, extracts text, tables, and formula structures, performs optical character recognition on scanned PDFs, and decodes audio and video data into timestamped text segments. All parsing results are then segmented to generate a unified, standardized text dataset. S3, Steps for Terminology Recognition and Knowledge Graph Construction: A neural network model finely tuned with domain-specific terminology corpus is used to identify technical terms from the standardized text dataset; the identification results are first queried from the local cache, and if no match is found, the cloud knowledge base application programming interface is called to obtain the term definitions and related documents; the terms and their relationships are constructed into an interactive knowledge graph; S4, Dynamic Tree Graph Navigation and Behavior Awareness Steps: It analyzes the logical hierarchical structure of documents and automatically generates an interactive tree-like navigation directory; it records user annotations, notes, and browsing behavior on tree nodes and associates them with corresponding node identifiers to form a mapping index of structure, content, and notes; S5, Semantic Retrieval and Personalized Recommendation Steps: Upon receiving user query requests, the system simultaneously performs keyword retrieval based on inverted indexes and similarity matching based on semantic vectors, and re-ranks the search results by combining entity associations from the knowledge graph; based on user historical behavior profiles, it performs personalized filtering and recommendation of document content. S6, Microservice Deployment and System Fault Tolerance Steps: The functions of steps S1 to S5 are broken down into independent microservices and deployed in containers; automatic scaling of services is achieved through cluster orchestration tools; when the error rate of a service instance exceeds the threshold, the application interface gateway automatically triggers the circuit breaker mechanism and returns cached data in a downgraded manner.

[0014] In one embodiment, in step S1: The random salt value is 128 bits; the key derivation algorithm is the PBKDF2 algorithm, which iterates 10,000 times and uses the SHA-256 hash function; the encryption is AES-256 encryption and the block size is 4096 bytes.

[0015] In one embodiment, step S1 further includes a periodic security hardening step: The key is automatically updated at a preset time every week. After deriving a new subkey from the master key, the stored encrypted documents are re-encrypted in batches. The Merkle tree algorithm is used to verify the integrity of the encrypted documents. If the verification fails, the local backup and recovery mechanism is triggered.

[0016] In one embodiment, in step S2: Optical character recognition is performed on scanned PDFs, specifically by calling the Tesseract OCR engine to perform text recognition, setting the recognition confidence threshold to 0.85, and generating a searchable text layer; For Office 2003 and earlier versions of documents, convert them to OOXML format using LibreOffice and then perform parsing.

[0017] In one embodiment, in step S3: The neural network model is the BERT model, which is fine-tuned on a domain corpus containing 100,000 academic literature summaries. The fine-tuning parameters include: a learning rate of 5e-5, a batch size of 16, and 10 training rounds. At a set time each day, domain terminology data is incrementally synchronized from the cloud knowledge base. The synchronized data is then compared with the local MD5 hash and the local cache and graph database are updated.

[0018] In one embodiment, in step S5: The search results are filtered by field to prioritize the display of literature that matches the user's research field; and a literature quality score is generated based on the user's reading time and annotation frequency, with literature with a score of ≥7 marked as "high-quality recommendation"; The cosine similarity threshold for semantic vector similarity matching is set to 0.75.

[0019] In one embodiment, the specific process of the circuit breaker mechanism in step S6 is as follows: When the error rate of a service instance is detected to be ≥5% for 5 consecutive minutes, the API gateway will automatically trigger a circuit breaker. After the circuit breaker state lasts for 10 minutes, it will enter a half-open state, allowing 10% of requests to attempt to call the real service. If the success rate is ≥90%, the service will be restored to normal; otherwise, it will re-enter the circuit breaker state.

[0020] In one embodiment, step S4, predicting potential interest nodes based on user behavior data, specifically involves: The system collects user behavior data, such as dwell time and scrolling frequency at each tree node. After anonymization, the data is input into the LSTM model for real-time analysis to predict potential interest nodes and highlight them.

[0021] Compared with existing technologies, the advantages of this invention are as follows: The privacy-preserving intelligent document interaction and reading assistance system provided reduces the risk of sensitive data leakage by constructing a two-level encryption system; it effectively shortens the multimodal parsing time and reduces terminology query response latency through unified parsing and standardization of multimodal data, terminology recognition, and knowledge graph construction, thereby improving reading efficiency; dynamic tree-graph navigation, behavior perception, semantic retrieval, and personalized recommendations help shorten chapter location time and improve the accuracy of personalized recommendations; and through the design of microservice deployment and system fault tolerance mechanisms, it ensures that the system can automatically scale up when there are ≥50 concurrent tasks and maintain the availability of core functions when some services are abnormal. This invention integrates natural language processing (NLP), privacy computing, distributed storage, and microservice architecture, and is suitable for researchers' intelligent reading, knowledge extraction, and privacy protection scenarios for multimodal academic literature.

[0022] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0024] Figure 1 This is a schematic diagram of the overall architecture of a privacy-preserving intelligent document interactive reading aid system according to an embodiment of the present invention; Figure 2 This is a flowchart of a two-level encryption and access control method according to an embodiment of the present invention; Figure 3 This is a flowchart of a method for unified parsing and standardized processing of multimodal data according to an embodiment of the present invention; Figure 4 This is a flowchart of a terminology recognition and knowledge graph interaction method according to an embodiment of the present invention; Figure 5 This is a flowchart of a dynamic tree diagram navigation and user behavior perception method according to an embodiment of the present invention; Figure 6 This is a flowchart of a semantic retrieval and personalized recommendation method according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a microservice fault tolerance and circuit breaker mechanism according to an embodiment of the present invention. Detailed Implementation

[0025] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0026] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0027] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0028] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0029] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0030] In general, the privacy-protected intelligent document interaction and reading assistance system provided by this invention includes a multimodal data unified parsing module, a terminology knowledge graph construction module, a two-level encryption privacy protection module, a dynamic tree diagram navigation module, a semantic retrieval and personalized recommendation module, and a microservice architecture support module. Each module achieves intelligent document processing and privacy protection through the following collaborative steps.

[0031] The dual-level encryption privacy protection module performs a preprocessing process, including: a front-end form interceptor generates a 128-bit random salt value, which, combined with the user's password hash value, generates an AES-256 key using the PBKDF2 algorithm (10,000 iterations, SHA-256 hash function); the document data is encrypted in 4096-byte blocks, with an encryption identifier, initialization vector (IV), and salt value information embedded in the file header; when the private space is created, the Spring Security filter generates a JWT token containing user role information (valid for 1 hour), requiring users to complete a two-stage verification process: password verification + token validity check. For example, the JWT token includes custom declaration fields: doc_scope limits the range of accessible document IDs, operate_perm indicates the allowed operations (read-only / edit / share), and device_id binds to the unique identifier of the device upon first login.

[0032] In one embodiment, the two-level encryption privacy protection module also includes a periodic security hardening step. For example, a key update is automatically performed every Monday at 3:00 AM, and new subkeys are derived from the master key to perform batch re-encryption of stored files; a Merkle tree algorithm is used for integrity verification, and a local backup and recovery mechanism is triggered if the verification fails.

[0033] The multimodal data unified parsing module performs the following steps: It calls the Apache Tika parser to identify PDF / Word document types and uses magic number detection to eliminate the risk of fake files; it uses PDFBox to parse the page layout and marks formula content with regular expressions, and combines Apache POI to extract table data; it calls FFmpeg to decode MP4 / MP3 audio and video files (sampling rate 44.1kHz, bit rate 128kbps), extracts the audio stream, and then uses Whisper API to transcribe it into timestamped text segments; all parsing results are segmented by HanLP and then uniformly stored in a standardized dataset.

[0034] In one embodiment, the multimodal data unified parsing module performs text recognition (with a recognition confidence level ≥ 0.85) on scanned PDFs using Tesseract OCR and generates a searchable text layer; and parses Office 2003 and earlier versions of documents after converting them to OOXML format using LibreOffice.

[0035] The terminology knowledge graph construction module is executed as follows: a pre-trained BERT model is fine-tuned on a domain corpus (containing 100,000 academic literature abstracts), and the expected F1 score for terminology recognition can reach 92%; the recognition results are first queried in the local Redis cache (the cache is valid for 24 hours), and if no match is found, the CNKI open API is called to obtain the terminology definition and related literature; the front end uses the Vue3 custom directive v-term-highlight to mark text terms, and clicking triggers a floating window component to load the knowledge graph associated node data (node ​​depth ≤ 3 layers).

[0036] In one embodiment, the terminology knowledge graph construction module incrementally synchronizes domain terminology data (incremental JSON format) from the cloud knowledge base at 2:00 AM daily. The synchronized data is then updated in the Redis cache and graph database after being compared with the local MD5 hash.

[0037] The dynamic tree diagram navigation module executes the following: it parses the hierarchical relationship of document titles to generate a tree-like data structure, and implements node collapse / expansion through Vue3 component-based rendering; it synchronously records user operation logs to an encrypted personalized note library, and automatically associates them with the current tree node ID when the user annotates text, forming a "structure-content-note" three-element association index.

[0038] In one embodiment, the dynamic tree map navigation module records user dwell time, scrolling frequency and other behavioral data (with anonymized processing), predicts potential interest nodes of the user based on the LSTM model and highlights them with light red.

[0039] Semantic retrieval and personalized recommendation module execution: Elasticsearch is used to build an inverted index, and text vectors (768 dimensions) are generated through the BERT model to build a vector index; when a user searches, the input text is segmented and keyword retrieval and vector similarity matching (threshold ≥ 0.75) are triggered simultaneously, and the results are re-ranked based on the entity association degree of the knowledge graph.

[0040] In one embodiment, the semantic retrieval and personalized recommendation module performs domain filtering on the results, prioritizing the display of documents that match the user's research field (identified through historical reading records); and generates a document quality score (out of 10) based on the user's reading time and annotation frequency, with documents scoring ≥7 points marked as "high-quality recommendations".

[0041] In another embodiment, the semantic retrieval and personalized recommendation module can employ a multi-model fusion strategy to further improve performance. For example, semantic vector generation can utilize a RoBERTa model trained on a larger corpus to obtain richer contextual representations; in the personalized recommendation stage, in addition to collaborative filtering and knowledge graph association, a LightGBM gradient boosting decision tree model can be introduced to comprehensively analyze user historical behavior features, document metadata, and real-time contextual information to predict click-through rate (CTR) and filter high-quality content, thereby achieving more accurate personalized recommendations.

[0042] The microservice architecture supports module execution: Spring Boot is used to split the system into independent microservices such as authentication service, parsing service, and retrieval service. Docker containerization and Kubernetes automatic scaling are adopted for deployment. When the number of concurrent parsing tasks is ≥50, service instances are automatically added.

[0043] In one embodiment, the microservice architecture support module adopts a service circuit breaker mechanism: when the error rate of a service is ≥5% for 5 consecutive minutes, the API gateway automatically triggers the circuit breaker and returns the cached result; after the circuit breaker state lasts for 10 minutes, it enters a half-open state, allowing 10% of requests to attempt to call, and the normal service is restored when the success rate is ≥90%.

[0044] Specifically, see Figure 1 As shown, the privacy-protected intelligent document interaction and reading assistance system adopted a three-layer architecture of "front-end presentation layer - back-end microservice layer - data storage layer".

[0045] The front-end presentation layer is built on the Vue 3 framework and uses the Element Plus component library to implement a responsive user interface. This layer is mainly responsible for rendering dynamic tree diagram navigation components, integrating terminology highlighting and floating window parsing components, and communicating with the backend via API through Axios.

[0046] The backend microservice layer is built on the Spring Boot framework and is divided into several independent microservices, including user authentication service, document parsing service, semantic search service, and terminology knowledge service. These services communicate via RESTful APIs, with Spring Cloud Gateway serving as the unified API gateway.

[0047] The data storage layer employs a multi-source heterogeneous storage solution. For example, a MySQL database is used to store user information, document metadata, and permission relationships; an Elasticsearch cluster is used to store full-text and vector indexes of documents; Redis is used to cache terminology data and frequently accessed documents; and MinIO distributed object storage is used to store encrypted original document files and intermediate parsing results. The system is deployed using Docker containers and managed uniformly by a Kubernetes cluster, enabling automatic scaling and failover of services.

[0048] Figure 2 It is a two-level encryption and access control process, specifically including the following steps: Step S201: When a user uploads a document, the front-end interceptor automatically generates a 128-bit random salt value.

[0049] Step S202: The system calls the PBKDF2 algorithm to perform a hash calculation on the salt value and the user's password (e.g., 10,000 iterations, hash function SHA-256) to generate a 256-bit AES encryption key.

[0050] Step S203: Using the AES key, combined with a randomly generated initialization vector (IV), the document data is encrypted in blocks (4096 bytes each).

[0051] Step S204: Write the encryption identifier, IV and salt value into the file header and store them together with the encrypted ciphertext in MinIO.

[0052] Step S205: When a user creates or accesses a private space, the system generates a JWT token containing custom claims (such as doc_scope, operate_perm) through Spring Security, for example, with an expiration time set to 1 hour.

[0053] Step S206: When a user accesses encrypted documents or a private space, they must provide both a valid password and an unexpired JWT token. The gateway filter performs secondary verification, and data can only be decrypted or access granted after successful verification.

[0054] Figure 3 This is a schematic diagram illustrating the unified analysis and standardization process of multimodal data. The multimodal analysis process includes the following steps: Step S301: The system identifies the MIME type of uploaded files using the Apache Tika parser and eliminates the risk of spoofed files through magic number detection.

[0055] Step S302: For PDF documents, call the PDFBox library to parse the page text flow and layout, use regular expressions to identify and mark embedded formulas, and extract embedded table data through Apache POI.

[0056] Step S303: For scanned PDFs, the Tesseract OCR engine is called to perform text recognition. Through preprocessing methods such as image binarization and noise reduction, the recognition confidence is improved to above 0.85, and a searchable transparent text layer is generated.

[0057] Step S304: For MP4 / MP3 and other audio and video files, FFmpeg is called to decode them (sampling rate 44.1kHz, bit rate 128kbps). After extracting the audio stream, speech recognition is performed through the Whisper API to generate text segments with precise timestamps.

[0058] Step S305: For all parsed text, use the HanLP word segmentation tool to perform word segmentation and part-of-speech tagging, filter out stop words, and store them uniformly in a structured JSON dataset to lay the foundation for subsequent processing.

[0059] Figure 4 This describes the construction and interaction process of a terminology knowledge graph, specifically including the following steps: Step S401: The BERT-NER model (learning rate 5e-5, batch size=16, epochs=10), which was finely tuned on 100,000 academic abstract corpora, was used to perform named entity recognition on the standardized text. It is good at extracting academic terms and the F1 score reached 92%.

[0060] Step S402: The identified terms are first queried in the local Redis cache (cache validity period is 24 hours). If the cache is not found, the authoritative definition and relevant literature citations of the terms are obtained by calling the CNKI open API or domain knowledge base, and the cache is updated.

[0061] Step S403: The front-end uses the Vue3 custom directive v-term-highlight to automatically render the terms in the literature as highlighted.

[0062] Step S404: When the user clicks on the highlighted term, the floating window component is triggered. This component asynchronously requests the term service to obtain and visualize the related nodes of the term in the knowledge graph (depth limited to 3 layers or less), forming a closed loop of "identification-parsing-association".

[0063] Figure 5 This is a flowchart of the dynamic tree diagram navigation and user behavior awareness method. The implementation process of dynamic tree diagram navigation includes the following steps: Step S501: The system parses the title hierarchy structure of the document (such as H1, H2, H3...) and converts it into a tree-like JSON data structure.

[0064] Step S502: The front-end uses Vue3 components to recursively render the tree structure and provides interactive interfaces for expanding / collapsed nodes.

[0065] Step S503: The system records the user's operation logs at each node of the tree diagram (such as clicks, expansions, and dwell time), and associates them with the user's text annotations and notes in the corresponding chapters. A three-element mapping index of "structure-content-notes" is formed through the node ID. All notes are stored after being encrypted.

[0066] Step S504: Based on the anonymized user behavior data (stay time, scrolling frequency), use the LSTM model for real-time analysis to predict potential user interest nodes, and highlight them with a light red background on the interface to guide users to read.

[0067] Figure 6 It is a semantic retrieval and personalized recommendation method. The semantic retrieval and recommendation process includes the following steps: Step S601: The user inputs the query text, and the system triggers two retrieval modes simultaneously: on the one hand, keyword matching and Boolean retrieval are performed through Elasticsearch; on the other hand, the BERT model is used to convert the query text into a 768-dimensional semantic vector, and similarity matching is performed in the Elasticsearch vector index (cosine similarity threshold is set to 0.75).

[0068] Step S602: Merge the results of the two retrieval modes, and use the knowledge graph to calculate the relevance between entities and queries in the retrieval results, and perform comprehensive re-ranking.

[0069] Step S603: The recommendation module analyzes the user's historical reading records and identifies their research field through a collaborative filtering algorithm.

[0070] Step S604: The system automatically calculates a document quality score (out of 10) based on the user's reading time, annotation frequency, and other behaviors. Documents with a score of ≥7 are marked as "high-quality recommendations" and are given priority in the search results and recommendation list.

[0071] Figure 7 This refers to the fault tolerance and circuit breaker mechanism for microservices. The fault tolerance process in a microservice architecture is implemented through the following steps: Step S701: The system monitors the resource metrics of each microservice through Kubernetes' HPA (Horizontal Pod Auto-Scaling) mechanism. When the number of concurrent tasks in the parsing service is detected to be ≥50 for 2 consecutive minutes, a new service instance is automatically added to share the load.

[0072] Step S702: The API gateway continuously monitors the error rate of all microservice interfaces. When the error rate of a service (such as the terminology service) is found to be ≥5% for 5 consecutive minutes, the circuit breaker is automatically triggered to enter the OPEN state.

[0073] Step S703: During the circuit breaker period, all requests sent to the service are intercepted by the gateway, and the cached results generated based on historical data are returned immediately to ensure that the user experience is not affected.

[0074] Step S704: After the circuit breaker remains in the HALF-OPEN state for 10 minutes, it enters the HALF-OPEN state, allowing 10% of requests to pass and attempt to call the real service. If the success rate of these requests is ≥90%, the service is considered to have been restored, the circuit breaker closes, and traffic returns to normal; otherwise, the circuit breaker re-enters the OPEN state.

[0075] Accordingly, this invention also provides a privacy-preserving intelligent document interaction and reading assistance method. The method includes: Step S1, two-level encryption and access control. The system receives document data uploaded by the user, generates a random salt value, and derives an AES-256 encryption key using the PBKDF2 algorithm with the user's password; the key is used to encrypt the document data in blocks, and an initialization vector and salt value are embedded in the file header to complete file-level encryption. Simultaneously, a JWT token containing the document access scope and operation permissions is created for the request. When the user accesses encrypted data or the private space, a second verification using the password and token is required to achieve space-level isolation. Step S2, unified parsing and standardization of multimodal data. An integrated parsing engine is invoked to parse multimodal data including PDF, Word, and audio / video: for documents, text, tables, and formula structures are extracted, and OCR recognition is performed on scanned PDFs; for audio / video, the data is decoded and transcribed into timestamped text fragments; finally, all parsing results are segmented and denoised to generate a unified standardized text dataset. Step S3, terminology recognition and knowledge graph construction. A BERT-NER model, fine-tuned with domain-specific terminology corpus, is used to identify specialized terms from a standardized dataset. The identification results are first queried from the local cache; if no match is found, the cloud-based knowledge base API is invoked to retrieve term definitions and related documents. An interactive knowledge graph is constructed from the terms and their relationships, and the terms are highlighted on the front end, responding to user clicks and displaying related node information. Step S4: Dynamic Tree Graph Navigation and Behavior Awareness. Based on the logical hierarchical structure of the documents, an interactive tree-like navigation directory is automatically generated; user annotations, notes, and browsing behavior on tree graph nodes are recorded synchronously and associated with corresponding node IDs, forming a "structure-content-note" mapping index; furthermore, potential interest nodes are predicted based on user behavior data and visual prompts are provided. Step S5: Semantic Retrieval and Personalized Recommendation. User query requests are received, and keyword retrieval based on the Elasticsearch inverted index and similarity matching based on BERT semantic vectors are performed synchronously, and the search results are re-ranked based on knowledge graph relevance. Based on the user's historical behavior profile, document content is quality-assessed and personalized, generating and pushing a customized recommendation list. Step S6, Microservice Deployment and System Fault Tolerance. The functions described in steps S1 to S5 are broken down into independent microservices and deployed in containers; automatic scaling of services is achieved through cluster orchestration tools; when the error rate of a service instance exceeds a threshold, the API gateway automatically triggers a circuit breaker mechanism, degrading and returning cached data to ensure high availability of core processes.

[0076] Furthermore, step S1 also includes a periodic security hardening step: automatically performing key updates according to a preset cycle, deriving new subkeys from the master key, and batch reencrypting the stored encrypted documents; using a Merkle tree to perform integrity verification on the encrypted documents, and triggering a local backup and recovery mechanism when the verification fails.

[0077] Furthermore, in step S2, OCR recognition is performed on the scanned PDF. Specifically, the Tesseract OCR engine is called, a recognition confidence threshold is set, and image preprocessing and re-recognition are performed on areas below the threshold until an acceptable confidence level is reached and a searchable text layer is generated.

[0078] Furthermore, in step S4, potential interest nodes are predicted based on user behavior data. Specifically, the user's dwell time, scrolling frequency, and annotation history at each tree node are collected, and after desensitization processing, they are input into the LSTM model for real-time analysis to predict the nodes that the user may pay attention to and highlight them.

[0079] Furthermore, in step S6, the specific process of the circuit breaker mechanism is as follows: when the error rate of a certain service is ≥5% for 5 consecutive minutes, the API gateway automatically triggers the circuit breaker, rejects requests and directly returns cached results within the following 10 minutes; after the circuit breaker state lasts for 10 minutes, it enters a half-open state, allowing 10% of requests to attempt to call the real service. If the success rate is ≥90%, the normal service is restored; otherwise, it re-enters the circuit breaker state.

[0080] In summary, the privacy-preserving intelligent document interaction and reading assistance system provided by this invention can be applied to the fields of academic research information security and intelligent knowledge services. This system integrates a named entity recognition model centered on BERT and its variants (such as RoBERTa) with a domain knowledge graph. It performs unified parsing, semantic indexing, and dynamic tree-graph navigation for multimodal materials such as PDFs, audio, and video, and restructures the document structure through three-level visualization. It utilizes the RAG architecture and BART summarization model to provide multi-granular summaries, question answering, and personalized recommendations. The front-end uses Vue3 component-based rendering, and the back-end uses a Spring Boot microservice architecture, combined with Elasticsearch, Redis, and MinIO to achieve high-concurrency retrieval and distributed storage. RBAC access control ensures user data access security. Through a two-level encryption system of "space + file" and localized privacy computation, AES-256 encryption and private space isolation are implemented in parallel, achieving hierarchical protection across the entire document chain. This invention can reduce the cost of terminology comprehension, effectively improve reading efficiency, reduce the risk of sensitive data leakage, and provide an efficient and secure intelligent solution for the entire scientific research process.

[0081] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0082] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0083] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0084] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0085] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0086] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0087] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be well known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0088] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A privacy-preserving intelligent document interaction and reading assistance system, comprising a multimodal data unified parsing module, a terminology knowledge graph construction module, a two-level encryption privacy protection module, a dynamic tree graph navigation module, a semantic retrieval and personalized recommendation module, and a microservice architecture support module, wherein: The dual-level encryption privacy protection module performs the following preprocessing: the front-end form interceptor generates a multi-bit random salt value, which is combined with the user password hash value to generate a key using the PBKDF2 algorithm; the document data is encrypted in blocks, and the encryption identifier, initialization vector and salt value information are embedded in the file header; when the private space is created, the Spring Security filter generates a JWT token containing user role information, and when the user accesses it, a second verification is performed to verify the password and the token's validity. The multimodal data unified parsing module performs the following steps: it calls the Apache Tika parser to identify PDF or Word document types and uses magic number detection to eliminate the risk of fake files; it uses PDFBox to parse the page layout and marks formula content with regular expressions, and combines Apache POI to extract table data; it calls decoding for audio and video files, extracts the audio stream, and then uses WhisperAPI to transcribe it into time-stamped text fragments; the parsing results are then segmented and uniformly stored in a standardized dataset. The terminology knowledge graph construction module works as follows: a pre-trained BERT model is fine-tuned on a domain corpus to obtain terminology recognition results; the recognition results are first queried in the local Redis cache, and if no match is found, the CNKI open API is called to obtain terminology definitions and related literature; the front end uses the Vue3 custom directive v-term-highlight to mark text terms, and clicking triggers a floating window component to load the knowledge graph associated node data. The dynamic tree diagram navigation module executes the following: it parses the hierarchical relationship of document titles to generate a tree-like data structure, and uses Vue3 component-based rendering to realize node collapse or expansion; it synchronously records user operation logs to an encrypted personalized note library, and automatically associates them with the current tree node ID when the user annotates text, forming a structure-content-note three-element association index; Semantic retrieval and personalized recommendation module execution: Elasticsearch is used to build an inverted index, and BERT or RoBERTa pre-trained models are used to generate text semantic vectors to build a vector index; when a user searches, the input text is segmented and keyword retrieval and vector similarity matching are triggered simultaneously; in the result ranking stage, machine learning models are combined to refine the search results, and personalized filtering and recommendations are performed based on user behavior profiles; The microservice architecture supports module execution: Spring Boot is used to split the service into authentication, parsing, and retrieval services, and Docker containerization and Kubernetes automatic scaling are adopted. When the concurrent parsing tasks exceed the set threshold, service instances are automatically added.

2. The system according to claim 1, characterized in that, The dual-level encryption privacy protection module also performs the following: automatically updating the key at set time points, deriving a new subkey from the master key, and then batch re-encrypting the stored files; using the Merkle tree algorithm for integrity verification, and triggering a local backup and recovery mechanism when the verification fails.

3. The system according to claim 1, characterized in that, The multimodal data unified parsing module uses Tesseract OCR to perform text recognition on scanned PDFs and generate a search text layer; it also parses Office 2003 and earlier versions of documents after converting them to OOXML format using LibreOffice.

4. The system according to claim 1, characterized in that, The terminology knowledge graph construction module incrementally synchronizes domain terminology data from the cloud knowledge base at set time points. After the synchronized data is compared with the local MD5 hash, the Redis cache and the graph database are updated.

5. The system according to claim 1, characterized in that, The semantic search and personalized recommendation module filters the results by domain, prioritizing the display of literature that matches the user's research field; it generates a literature quality score based on the user's reading time and annotation frequency, and marks literature with a score greater than a set threshold as "high-quality recommendation".

6. The system according to claim 1, characterized in that, The semantic retrieval and personalized recommendation module is also used to: generate enhanced semantic vectors using the RoBERTa model, and refine the retrieval results using the LightGBM ranking model.

7. The system according to claim 1, characterized in that, The microservice architecture support module adopts a service circuit breaker mechanism: when the error rate of a service exceeds a threshold for a continuous set time period, the API gateway automatically triggers the circuit breaker and returns the cached result; after the circuit breaker state lasts for a set time, it enters a half-open state, allowing a set proportion of requests to attempt to call, and the normal service is restored when the success rate is greater than the set proportion threshold.

8. The system according to claim 1, characterized in that, The dynamic tree map navigation module is also used to: predict potential interest nodes of the user based on the user's dwell time and scrolling frequency using an LSTM model and highlight them.

9. The system according to claim 1, characterized in that, The JWT token contains custom declaration fields: doc_scope limits the range of document IDs that can be accessed, operate_perm identifies the allowed operations, and device_id binds to the unique identifier of the device that was logged in for the first time.

Images