An academic paper polishing system and method

Abstract

The application relates to the technical field of artificial intelligence, and discloses an academic paper polishing system and method, which comprises a local computing device and a cloud large language model server. The local computing device maps original text to a semantic subspace and a privacy subspace, generates a privacy protection vector, constructs a semantic correlation graph and encodes the semantic correlation graph into a second vector, fuses the second vector into a privacy protection query vector containing noise mask identification, semantic correlation graph topology and perturbed semantic embedding, and uploads the privacy protection query vector to the cloud large language model server. The cloud side shields noise dimensions through a privacy perception gating mechanism, performs restricted retrieval and collaborative reasoning based on the semantic correlation graph topology, generates a polished feature vector, returns the polished feature vector after injecting result confusion noise, and outputs a polished text after the local computing device eliminates noise, decodes and verifies consistency. The application realizes controllable privacy and semantic fidelity in the academic paper polishing process through semantic-privacy dual-space mapping and hierarchical coding query.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an academic paper polishing system and method. Background Technology

[0002] Academic paper editing techniques have evolved from manual editing to automated tools. Early methods relied on professional editors for manual review, which was inefficient and costly. Subsequently, rule-based natural language processing tools achieved grammatical error correction and simple rewriting, but struggled to handle complex academic contexts. In recent years, large language models, with their powerful semantic understanding and generation capabilities, have significantly improved the quality and efficiency of editing, becoming the mainstream technological approach.

[0003] However, existing cloud-based large language model-based editing solutions pose a risk of privacy breaches. Users upload original papers containing sensitive research data, unpublished experimental results, and author identification information to the cloud, exposing private information to untrusted third parties. Furthermore, the cloud model may infer the original content from the output, lacking an effective privacy protection mechanism. Summary of the Invention

[0004] The main purpose of this application is to provide an academic paper editing system and method, which aims to solve the technical problem of the lack of an effective privacy protection mechanism in the process of academic paper editing.

[0005] To achieve the above objectives, this application provides an academic paper polishing system, which includes: a local computing device and a cloud-based large language model server, wherein the cloud-based large language model server includes a large language model with a privacy-aware gating mechanism in the embedded layer;

[0006] The local computing device is used to acquire the original text of the academic paper to be polished, map the original text to a semantic subspace and a privacy subspace, generate a first vector in the semantic subspace, allocate a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and inject noise into the first vector to obtain a privacy protection vector.

[0007] The local computing device is also used to retrieve academic semantic elements extracted from the semantic subspace from the local knowledge base, construct a semantic association graph between the retrieval results and the original text, and encode the semantic association graph to obtain a second vector;

[0008] The local computing device is further configured to fuse the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and upload it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding.

[0009] The cloud-based large language model server is used to receive the privacy-protected query vector, and parse the noise mask identifier through the privacy-aware gating mechanism to locate and mask the noise dimension in the perturbation semantic embedding to obtain the masked vector.

[0010] The cloud-based large language model server is also used to perform restricted retrieval in the cloud knowledge base based on the semantic association graph topology, and to perform collaborative reasoning with the restricted retrieval results and the masked vector to generate a polished feature vector. The polished feature vector is then injected with result obfuscation noise that matches the noise mask identifier to obtain a privacy-protected result vector and return it to the local computing device.

[0011] The local computing device is further configured to receive the privacy-preserving result vector, eliminate the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtain a refined intermediate vector;

[0012] The local computing device is also used to decode the polishing intermediate vector to obtain polishing candidate text, and after performing academic consistency verification on the polishing candidate text based on the semantic association graph, output the polished paper text.

[0013] Furthermore, to achieve the above objectives, this application also provides an academic paper polishing method, which is applied to an academic paper polishing system. The system includes: a local computing device and a cloud-based large language model server, wherein the cloud-based large language model server includes a large language model with a privacy-aware gating mechanism in its embedded layer; the method includes:

[0014] The local computing device acquires the original text of the academic paper to be polished, maps the original text to a semantic subspace and a privacy subspace, generates a first vector in the semantic subspace, allocates a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and injects noise into the first vector to obtain a privacy protection vector.

[0015] The local computing device retrieves academic semantic elements extracted from the semantic subspace from the local knowledge base, constructs a semantic association graph between the retrieval results and the original text, and encodes the semantic association graph to obtain a second vector;

[0016] The local computing device fuses the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and uploads it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding.

[0017] The cloud-based large language model server receives the privacy-protected query vector and uses the privacy-aware gating mechanism to parse the noise mask identifier to locate and mask the noise dimension in the perturbation semantic embedding, thereby obtaining the masked vector.

[0018] The cloud-based large language model server performs restricted retrieval in the cloud knowledge base based on the semantic association graph topology, and performs collaborative reasoning with the restricted retrieval results and the masked vector to generate a refined feature vector. The refined feature vector is then injected with result obfuscation noise that matches the noise mask identifier to obtain a privacy-protected result vector and return it to the local computing device.

[0019] The local computing device receives the privacy-preserving result vector, eliminates the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtains a refined intermediate vector;

[0020] The local computing device decodes the intermediate vector for polishing to obtain polishing candidate text, and performs academic consistency verification on the polishing candidate text based on the semantic association graph, and then outputs the polished paper text.

[0021] The technical solution proposed in this application discloses an academic paper polishing system and method. The system includes a local computing device and a cloud-based large language model server. The local computing device maps the original text to a semantic subspace and a privacy subspace, generates a privacy-preserving vector, constructs a semantic association graph and encodes it into a second vector, and fuses them into a privacy-preserving query vector containing noise masking, semantic association graph topology and perturbation semantic embedding. This vector is then uploaded to the cloud-based large language model server. The cloud server uses a privacy-aware gating mechanism to shield the noise dimension, performs restricted retrieval and collaborative reasoning based on the semantic association graph topology, generates a polished feature vector, injects noise to obfuscate the result and returns it, and the local computing device eliminates noise, decodes and verifies consistency before outputting the polished text. This application achieves parallel processing of academic semantic extraction and privacy protection by constructing a dual-space mapping mechanism of semantic subspace and privacy subspace; it utilizes a hierarchical coding structure to fuse noise masking, semantic association graph topology, and perturbation semantic embedding into a privacy-protected query vector, enabling the cloud to retain semantic association information for restricted retrieval while shielding noise dimensions; and it combines dynamic allocation of differential privacy budget with closed-loop elimination of result obfuscation noise to achieve accurate polishing of academic papers by a cloud-based large language model while ensuring the privacy and security of the original text, thus balancing the strength of privacy protection with the quality of polishing. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a structural block diagram of the first embodiment of the academic paper polishing system of this application;

[0025] Figure 2 This is a flowchart illustrating the first embodiment of the academic paper polishing method of this application.

[0026] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0028] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0029] In practical applications, academic paper authors typically rely on cloud-based large language models for text polishing. However, the original paper text contains sensitive content such as author identification information, experimental data, and unpublished research results, posing a privacy risk if directly uploaded to the cloud. Existing solutions either employ local model deployment, resulting in insufficient computing power and limited polishing quality, or use simple anonymization before uploading to the cloud, which compromises the integrity of academic semantics. Furthermore, cloud models cannot utilize private literature resources in local knowledge bases for semantic enhancement. In addition, the polished results returned by the cloud may also expose features of the original text, lacking a two-way privacy protection mechanism.

[0030] Therefore, to overcome the above-mentioned shortcomings, this application provides a solution that separates academic semantics and privacy elements by constructing a semantic subspace and a privacy subspace locally. It dynamically allocates differential privacy budgets based on risk levels and injects structured noise. A hierarchical coding structure is used to fuse noise masking identifiers, semantic association graph topology, and perturbation semantic embedding, enabling the cloud to perform restricted searches based on the semantic association graph while shielding the noise dimension. Finally, a closed-loop elimination of result obfuscation noise achieves bidirectional privacy protection, fully utilizing cloud computing power and local knowledge base resources while ensuring the security of sensitive information, thus achieving high-quality academic paper polishing.

[0031] Based on this, the embodiments of this application provide an academic paper polishing system, referring to... Figure 1 , Figure 1This is a structural block diagram of the first embodiment of the academic paper polishing system of this application. The academic paper polishing system includes: a local computing device 10 and a cloud-based large language model server 20, wherein the cloud-based large language model server 20 includes a large language model with a privacy-aware gating mechanism in the embedded layer.

[0032] The local computing device 10 is used to acquire the original text of the academic paper to be polished, map the original text to a semantic subspace and a privacy subspace, generate a first vector in the semantic subspace, allocate a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and inject noise into the first vector to obtain a privacy-protected vector.

[0033] This embodiment adopts an architecture that coordinates local computing devices with cloud-based large language model servers to resolve the contradiction between the leakage of original text privacy and the efficiency of cloud model utilization during the academic paper polishing process.

[0034] It should be noted that the local computing device is deployed on the user side and can be a personal computer, workstation, or edge server. It holds the user's original academic paper text and local knowledge base resources, possessing the ability to perform text parsing, semantic extraction, privacy detection, noise injection, vector encoding, and result decoding. However, limited by local computing power, it cannot run large-scale language models. The cloud-based large language model server is deployed in a remote data center. It holds a publicly available cloud knowledge base and a large language model pre-trained with massive amounts of academic terminology data. It possesses powerful semantic understanding, knowledge retrieval, and text generation capabilities, but does not directly access the user's original plaintext data. The two interact through privacy-preserving query vectors and privacy-preserving result vectors. The local computing device is responsible for upstream privacy protection processing and downstream result restoration processing, while the cloud-based large language model server is responsible for midstream privacy-aware reasoning and knowledge enhancement. This forms a closed-loop processing flow of local privacy protection – cloud-based secure reasoning – local result restoration, ensuring that the privacy and security of the user's original text are fully controllable throughout the process while fully utilizing cloud computing power and knowledge resources.

[0035] In practice, after the local computing device obtains the original text of the academic paper to be polished, it simultaneously maps the original text to a semantic subspace and a privacy subspace. The semantic subspace is a vector space used to carry the academic expression connotation in the text. Its construction depends on the identification and encoding of language units that embody academic value in the original text, such as discipline-specific terms, descriptions of research methods, experimental design logic, and data analysis conclusions. These language units together constitute academic semantic elements.

[0036] The local computing device encodes the identified academic semantic elements into dense vector representations and aggregates them into a first vector in the semantic subspace. This first vector retains the core academic information of the original text and serves as the main semantic input for subsequent cloud-based models to refine and reason.

[0037] The privacy subspace is an independent vector space used to carry privacy-sensitive information in text. Its construction depends on the identification of content in the original text that may expose the author's identity or research secrets, such as the author's name, affiliation, email address, funding project number, unpublished experimental raw data, and technical details before patent application. These contents constitute privacy-sensitive elements.

[0038] The local computing device performs risk assessments on various privacy-sensitive elements, classifying them into risk levels based on the potential harm caused by their leakage, and allocating differential privacy budgets to different categories of privacy-sensitive elements based on these risk levels. The differential privacy budget is a parameter that quantifies the strength of privacy protection; a higher budget value indicates a greater allowable noise level and stricter privacy protection. The local computing device injects noise into the dimensions corresponding to the privacy-sensitive elements in the first vector according to the allocated differential privacy budget, obtaining a privacy protection vector. This noise-injected privacy protection vector mathematically obscures the original privacy information while retaining sufficient academic semantic information for subsequent processing.

[0039] The local computing device 10 is also used to retrieve academic semantic elements extracted from the semantic subspace from the local knowledge base, construct a semantic association graph between the retrieval results and the original text, and encode the semantic association graph to obtain a second vector.

[0040] It should be understood that while generating privacy-preserving vectors, local computing devices retrieve relevant literature fragments from a local knowledge base based on academic semantic elements extracted from the semantic subspace. The local knowledge base stores users' private academic resources, such as published papers, internal technical reports from research groups, and core references in the field. These resources are not publicly available and cannot be directly accessed by the cloud-based model.

[0041] The local computing device performs semantic association analysis on the relevant literature fragments retrieved and the original text, identifies the citation relationships, concept inheritance relationships, method borrowing relationships, experimental comparison relationships, etc. between the literature fragments and the original text, and constructs a semantic association graph based on these relationships.

[0042] It should be noted that a semantic association graph is a graph-structured data representation where nodes represent academic semantic elements in the original text or relevant document fragments retrieved, edges represent semantic relationships between nodes, and edge weights reflect the strength of these relationships. The semantic association graph encodes the contextual location information of the original text within the local knowledge base, providing the cloud model with an extended semantic background beyond the original text itself. The local computing device encodes this semantic association graph into a second vector, which complements the privacy-preserving vector: the privacy-preserving vector carries the semantics of the original text after noise processing, while the second vector carries the structural information of the association between the original text and local private documents.

[0043] The local computing device 10 is further configured to fuse the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and upload it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding.

[0044] In its implementation, the local computing device heterogeneously fuses the privacy-preserving vector with the second vector to generate a privacy-preserving query vector. Heterogeneous fusion refers to integrating two vector representations from different sources and with different structures but related semantics into a unified data structure, enabling the cloud model to simultaneously parse the semantic content and contextual information of the original text. The privacy-preserving query vector adopts a hierarchical encoding structure, logically divided into three layers: the upper layer is a noise mask identifier, used to record whether each dimension of the privacy-preserving vector has been injected with noise, as well as the type and intensity of the noise, serving as the key basis for the cloud model to identify and mask noise dimensions; the middle layer is a semantic association graph topology, used to carry the node connection relationships and edge weight information of the semantic association graph, forming the structural foundation for the cloud model to perform restricted retrieval; the lower layer is a perturbation semantic embedding, i.e., the privacy-preserving vector itself after noise injection, which is the core input for the cloud model to perform semantic reasoning.

[0045] The local computing device uploads the generated privacy-preserving query vector to the cloud-based large language model server via the network. Due to the presence of noise masking identifiers, even if the privacy-preserving query vector is intercepted during transmission, attackers cannot accurately distinguish between the noise dimension and the true semantic dimension, thus ensuring transmission security.

[0046] The cloud-based large language model server 20 is used to receive the privacy-protected query vector, parse the noise mask identifier through the privacy-aware gating mechanism to locate and mask the noise dimension in the perturbation semantic embedding, and obtain the masked vector.

[0047] It should be understood that after receiving the privacy-preserving query vector, the cloud-based large language model server first parses it through a privacy-aware gating mechanism set in the embedding layer. The privacy-aware gating mechanism is a dedicated functional module in the large language model's embedding layer, designed to identify and process special encoding structures in the privacy-preserving query vector from the local computing device. Specifically, the privacy-aware gating mechanism parses the noise mask identifier from the upper layer, locates the noise dimensions injected into the perturbed semantic embedding, and masks these noise dimensions to obtain the masked vector.

[0048] The essence of masking is to isolate noise interference in the vector space, ensuring that subsequent inference is based solely on genuine academic semantic information. After masking, the vectors retain the core academic semantics of the original text but eliminate the negative impact of differential privacy noise on model inference, a prerequisite for high-quality refinement of cloud-based models.

[0049] The cloud-based large language model server 20 is also used to perform restricted retrieval in the cloud knowledge base based on the semantic association graph topology, and to perform collaborative reasoning with the restricted retrieval results and the masked vector to generate a polished feature vector. The server then injects result obfuscation noise that matches the noise mask identifier into the polished feature vector to obtain a privacy-protected result vector and returns it to the local computing device.

[0050] It's important to note that restricted retrieval refers to limiting the search scope to a subset of documents directly related to the semantic association graph topology, rather than performing a global traversal of the cloud-based knowledge base. The cloud-based knowledge base stores publicly available academic literature, pre-training corpora, domain knowledge graphs, and other resources, and its scale is far larger than that of the local knowledge base. Guided by the semantic association graph topology, the cloud-based model can quickly locate cloud-based document fragments highly relevant to the original text's research topic, avoiding interference from irrelevant information and improving retrieval efficiency and result relevance.

[0051] Furthermore, the cloud-based large language model server performs collaborative reasoning with the restricted search results and the masked vectors. Collaborative reasoning refers to the cloud model simultaneously utilizing two types of input information: the masked vectors carry the semantic content of the original text, while the restricted search results carry the related knowledge of the original text in a broader academic context. The combination of these two allows the model to understand the deeper academic connotations of the original text and refine it precisely. The reasoning process is completed by the multi-layered Transformer structure of the large language model, generating refined feature vectors that encode the academic language expressions optimized by the cloud model.

[0052] The local computing device 10 is further configured to receive the privacy protection result vector, eliminate the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtain a refined intermediate vector.

[0053] It should be understood that the result obfuscation noise is a pseudo-random noise sequence that matches the noise mask identifier, and its dimensionality distribution and intensity characteristics correspond to the locally injected differential privacy noise. The purpose of injecting result obfuscation noise is to prevent attackers from inferring the original text content by analyzing the polished results returned from the cloud. Even if an attacker intercepts the privacy-protected result vector, they will not be able to distinguish between the real polished semantics and the obfuscation noise, thereby protecting user privacy from the cloud model output side.

[0054] The cloud-based large language model server returns the privacy-preserving result vector to the local computing device. Throughout the entire cloud processing process, the original text exists in a noisy and encoded form, and the cloud model can never directly access the original plaintext.

[0055] The local computing device 10 is also used to decode the polishing intermediate vector to obtain polishing candidate text, and after performing academic consistency verification on the polishing candidate text based on the semantic association graph, output the polished paper text.

[0056] The local computing device generates a noise mask identifier and records the differential privacy budget allocation strategy in the above steps, which contains complete information for eliminating obfuscated noise in the results. The elimination operation is the inverse process of the injection operation. Through mathematical calculation, obfuscated noise components are stripped from the privacy-preserving result vector to restore the true polished semantic representation and obtain the polished intermediate vector.

[0057] Furthermore, the local computing device decodes the intermediate vectors used for polishing, mapping them back from the vector space to the natural language text space to obtain polished candidate text. This candidate text is an optimized expression generated by the cloud model based on academic semantic reasoning, but it needs to be verified locally to ensure its academic consistency with the original text.

[0058] Based on the semantic association graph constructed in the above steps, the local computing device performs academic consistency verification on the candidate text for polishing. Verification includes checking whether the polished academic terminology matches the research field of the original text, whether the experimental data description retains its original meaning, whether the citation relationships are accurately maintained, and whether the description of research methods has deviated. If the verification passes, the local computing device outputs the polished paper text; if inconsistencies are found, a partial re-reasoning or manual review process can be triggered. The final polished text retains the academic core of the original text while achieving professional optimization of language expression, and at the same time, no original privacy information is exposed to the cloud throughout the process.

[0059] As one implementation, the cloud-based large language model server 20 is also used to extract node identifiers and edge weights by parsing the semantic association graph topology.

[0060] After receiving the privacy-preserving query vector, the cloud-based large language model server parses the semantic association graph topology of its middle layer. This topology records the unique identifier of each node in the graph (such as the DOI, local number, or title hash value of a document fragment) and the edge weights between nodes in the form of structured data. For example, the parsed list of node identifiers is [Node_A, Node_B, Node_C], and the edge weights are recorded as AB: 0.85, AC: 0.62, and BC: 0.91.

[0061] The cloud-based large language model server 20 is also used to perform a same-domain search in the cloud knowledge base using the node identifier as the search key, and return cloud document fragments with a matching degree higher than a threshold with the node identifier.

[0062] It's important to note that the cloud-based knowledge base stores a massive amount of publicly available academic literature. The server converts each node identifier into a corresponding search keyword or document identifier, enabling precise matching or semantic retrieval within the cloud-based knowledge base. Same-domain retrieval means limiting the search scope to a subset of documents within the same academic field as the node identifier. For example, if the node identifier represents a document about "graph neural networks," the system will only search within the computer science sub-database of the cloud-based knowledge base.

[0063] Returns all cloud-based document fragments with a matching degree higher than a preset threshold (e.g., 0.75) with respect to the node identifier. The matching degree can be calculated based on title similarity, abstract vector cosine similarity, or citation relationship overlap.

[0064] The cloud-based large language model server 20 is also used to sort and filter the cloud-based document fragments according to the edge weights, and retain the cloud-based document fragments that have a direct semantic relationship with the original text as limited search results.

[0065] For each returned cloud-based document fragment, it is associated with one or more original nodes. The cloud-based large language model server sorts these fragments according to the edge weights of the nodes in the semantic association graph: the higher the edge weight, the stronger the semantic association between the node and the original text, and the more relevant the corresponding search results. For example, if the average edge weight of node A is 0.85 and the average edge weight of node C is 0.62, then the cloud-based document fragments retrieved by node A will be ranked higher. During filtering, fragments from nodes with higher edge weights are retained, or an edge weight threshold (such as 0.7) is set to retain only cloud-based document fragments whose associated node edge weights are greater than this threshold, ultimately resulting in restricted search results.

[0066] As one implementation, the cloud-based large language model server 20 is also used to construct a semantic association strength matrix between nodes based on the edge weights, and to mark the nodes connected by edges whose edge weights are lower than a preset association threshold as weakly associated nodes, and to mark the nodes connected by edges whose edge weights are higher than the preset association threshold as strongly associated nodes.

[0067] The cloud-based large language model server first constructs a square matrix based on the edge weights of all nodes in the semantic association graph. The element in the i-th row and j-th column of the matrix represents the edge weight between node i and node j (0 is filled in if there is no direct edge). A preset association threshold is set, for example, 0.5.

[0068] For each node, calculate the average or maximum edge weight between it and all its neighbors: if the maximum value is less than 0.5, the node is marked as a weakly associated node; if there is an edge weight greater than or equal to 0.5, it is marked as a strongly associated node. For example, node B has an edge weight of 0.91 with node A and 0.85 with node C, both of which are greater than 0.5, therefore node B is a strongly associated node; node D only has an edge with node C with a weight of 0.3, therefore node D is a weakly associated node.

[0069] The cloud-based large language model server 20 is also used to perform preliminary screening of the cloud-based document fragments, removing cloud-based document fragments originating from the weakly related nodes, and retaining a set of candidate cloud-based document fragments originating from the strongly related nodes.

[0070] In the returned candidate cloud document fragments, each fragment corresponds to one or more original nodes. The system checks the strength of these nodes: if all source nodes of a cloud document fragment are weakly related nodes, the fragment is directly removed; if at least one source node is a strongly related node, the fragment is retained and added to the candidate set.

[0071] The cloud-based large language model server 20 is also used to calculate the cross-document semantic coherence score between each cloud document fragment in the candidate cloud document fragment set and the original text, and retain the cloud document fragments with cross-document semantic coherence scores higher than the coherence threshold as restricted search results.

[0072] It should be understood that for each cloud-based document fragment retained after the initial screening, the system calculates its cross-document semantic coherence with the original text. This score is evaluated using an additional semantic model: the overall semantic vector of the original text and the summary vector of the cloud-based document fragment are input into a pre-trained coherence estimator, which outputs a value between 0 and 1, reflecting the degree of coherence between the two in academic logic. For example, if the original text discusses "the application of Transformers in machine translation" and the cloud-based document fragment describes "variants of attention mechanisms," the coherence score might be 0.82; while another fragment describing "the application of convolutional neural networks in image recognition" might only have a coherence score of 0.21. A coherence threshold is set, for example, 0.6, and only cloud-based document fragments with scores higher than this threshold are retained as the final restricted search results.

[0073] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on the above Figure 1 The illustrated embodiment presents a second embodiment of the academic paper polishing system of this application.

[0074] In this embodiment, the local computing device 10 is also used to perform word segmentation and part-of-speech tagging on the original text, and to identify at least one type of academic semantic element in academic terms, research method descriptions, and experimental data.

[0075] In practice, after obtaining the original academic paper text to be polished, the local computing device first executes a text preprocessing pipeline. Word segmentation employs a sequence labeling model based on a bidirectional long short-term memory network and conditional random fields, capable of adapting to scenarios with unclear boundaries of compound words, abbreviations, and proper nouns commonly found in academic papers. Part-of-speech tagging utilizes a fine-grained academic tagging set, distinguishing not only common nouns and verbs but also identifying special grammatical categories such as methodological markers and data descriptive adverbs.

[0076] Furthermore, based on the results of word segmentation and part-of-speech tagging, the local computational setup further filters out three categories of academically valuable content: academic terminology refers to standardized conceptual terms within a specific research field, such as "density functional theory" and "Markov decision process"; research method descriptions refer to verb phrases or prepositional structures characterizing experimental design, algorithmic steps, or analytical processes, such as "optimizing parameters through gradient descent" and "adopting a double-blind controlled design"; experimental data includes all quantified observations, statistical indicators, and mathematical expressions, such as "mean absolute error is 0.32" and "correlation coefficient r=0.87". The recognition process can utilize a domain-adaptive named entity recognition model, supplemented by regular expression matching of numerical patterns. The output of this step is a set of semantic elements with location and category labels, providing accurate source data for subsequent encoding while excluding non-core semantic content such as acknowledgments and references from the paper.

[0077] The local computing device 10 is further configured to encode the academic semantic elements into semantic vectors based on a word embedding model, construct a semantic subspace using the spanned space of the semantic vectors, and generate a first vector in the semantic subspace.

[0078] It should be understood that each identified academic semantic element exists in the form of natural language fragments and cannot directly participate in vector operations. The local computing device loads a word embedding model pre-trained on a corpus of tens of millions of academic papers. This model can map each word or phrase into a dense real-valued vector of fixed length, and semantically similar terms are close in distance in the vector space.

[0079] For a single term, its word embedding vector is directly extracted. For multi-word segments such as research method descriptions or experimental data, a Transformer-based sentence encoder is used to encode them as a single segment vector, or an attention-weighted average is applied to the word vectors within the segment, with the weights determined by the information entropy of the word in the academic context. All resulting semantic vectors reside in the same high-dimensional space, but there are correlations between them. The local computing device performs principal component analysis or orthogonal decomposition on these semantic vectors, using a set of orthonormal bases spanned by their space as the coordinate axes of the semantic subspace. This subspace is essentially a low-dimensional linear manifold, capable of preserving most of the variance information in the set of semantic vectors while eliminating redundant dimensions. Based on this, the system merges the semantic vectors according to their order of appearance in the original text and preset academic importance weights, for example, using a self-attention pooling mechanism, to output a single comprehensive vector, i.e., the first vector.

[0080] The local computing device 10 is also used to synchronously identify at least one type of privacy-sensitive element in the original text, including author identity information, institutional affiliation, funding project number, and unpublished experimental data, and to assess the semantic exposure risk level of various privacy-sensitive elements based on a preset privacy rule base.

[0081] In parallel with the extraction of the aforementioned academic semantic elements, the local computing device runs another processing chain on the same raw text, specifically for discovering privacy-sensitive content. Author identification is achieved by detecting co-occurrence patterns of email format, name, and contextual keywords such as "author" and "corresponding author." Institutional affiliation is typically represented by the name of a university, research institute, or laboratory, along with its subordinate departments. Funding project numbers have a fixed prefix and number combination format, such as "No.123456" or "Grant2024-001." Unpublished experimental data includes unpublished raw measurements, intermediate calculation results in experimental logs, and any data blocks marked "to be published" or "unpublished." After identification, the system calls a pre-set privacy rule base for risk assessment. This rule base is a knowledge system containing multi-layered reasoning logic, with each rule in the form of "If a privacy element is of type X and appears in context Y, then its exposure risk level is Z."

[0082] For example, the rules stipulate that when author identification information and unpublished experimental data appear simultaneously in the same paragraph, the combined risk level is upgraded to high; when a grant number appears alone, it is rated as low risk by default, given that it is usually publicly searchable; however, if the grant number is adjacent to the description of the technical solution in an ongoing patent application, it is upgraded to medium risk. The evaluation results assign a high-risk, medium-risk, or low-risk label to each type of privacy-sensitive element. This label directly determines the allocation weight of the subsequent differential privacy budget, ensuring effective protection against combined privacy leakage scenarios.

[0083] The local computing device 10 is further configured to encode the privacy-sensitive elements into privacy vectors, construct a privacy subspace using the spanned space of the privacy vectors, and allocate a differential privacy budget in the privacy subspace based on the semantic exposure risk level.

[0084] In practice, the local computing device numerically encodes each type of privacy-sensitive element. The encoding method varies depending on the element type: for discrete symbols such as author names and institution names, sparse one-hot encoding or hash embedding is used, with each possible value corresponding to a high-dimensional vector; for grant project numbers, their numerical sequence is preserved and mapped to a continuous value vector; for unpublished experimental data, the numerical value itself is used directly with the addition of a dimensional indicator bit.

[0085] All privacy-encoded vectors reside in the same original space. The local computing device performs a linear transformation on them, spanning a privacy subspace that is orthogonal or approximately orthogonal to the aforementioned semantic subspace. This orthogonality is achieved by pre-removing components along the privacy direction when generating the first vector, and by projecting the privacy vectors onto the orthogonal complement of the semantic subspace, thus ensuring that adding noise to the privacy dimension does not significantly affect the value of the academic semantic dimension.

[0086] Within the privacy subspace, the local computing device allocates a differential privacy budget to each privacy vector based on the risk level determined in the preceding steps. The budget value ε reflects the strength of privacy protection; the larger ε is, the greater the subsequent noise injection amplitude. The specific mapping rule is as follows: high-risk levels are allocated the maximum budget ε_high, medium-risk levels are allocated ε_mid, and low-risk levels are allocated ε_low, with ε_high ≥ ε_mid ≥ ε_low. For example, for a combined high-risk scenario involving both author identity and unpublished data, ε_high is set to 1.0, corresponding to a larger Laplace noise scale parameter; while for isolated grant project numbers, ε_low can be as low as 0.1, adding only a weak perturbation. This budget value is recorded in the noise mask identifier and used to guide the process of injecting structured noise into the corresponding dimension of the first vector, ultimately generating a privacy-protected vector. Through this differentiated budget allocation, the system strongly protects highly sensitive information while avoiding excessive noise eroding academic semantics, achieving refined utilization of the privacy budget.

[0087] In one implementation, the risk levels of the privacy-sensitive elements include high-risk privacy-sensitive elements, medium-risk privacy-sensitive elements, and low-risk privacy-sensitive elements. The allocation of differential privacy budgets includes a first differential privacy budget, a second differential privacy budget, and a third differential privacy budget. The local computing device 10 is further configured to allocate the high-risk privacy-sensitive elements to the first differential privacy budget, allocate the medium-risk privacy-sensitive elements to the second differential privacy budget, and allocate the low-risk privacy-sensitive elements to the third differential privacy budget, wherein the first differential privacy budget is greater than the second differential privacy budget, and the second differential privacy budget is greater than the third differential privacy budget.

[0088] The local computing device 10 is further configured to inject structured noise of corresponding intensity into the first vector according to the differential privacy budget to obtain a privacy-preserving vector, wherein the structured noise is Gaussian noise or Laplace noise, and the noise intensity of the privacy-preserving vector is positively correlated with the differential privacy budget.

[0089] In practice, structured noise can be Gaussian or Laplace noise. The scale parameter of the noise is inversely proportional to the differential privacy budget; therefore, the larger the budget, the higher the noise intensity. The resulting privacy-preserving vector satisfies the definition of differential privacy, effectively preventing attackers from reverse-engineering the original privacy information while retaining sufficient academic semantics for cloud processing.

[0090] In one implementation, the local computing device 10 is also used to retrieve relevant document fragments from a local knowledge base based on the academic semantic elements extracted from the semantic subspace, and to calculate the semantic similarity between the relevant document fragments and the original text.

[0091] The local computing device uses academic terms extracted from the original text (such as "graph neural network" and "knowledge graph embedding") as query terms to perform vector retrieval in a locally stored private collection of documents. Specifically, each query term and the title or abstract of each document fragment are encoded into a fixed-length vector, and then cosine similarity is calculated. For example, if the original text contains the term "Transformer-XL," and the local knowledge base contains a previously downloaded paper titled "An Improved Method for Transformer-XL," the similarity between its abstract vector and the term vector is 0.82. The system uses this similarity as the semantic similarity score between the document fragment and the original text.

[0092] The local computing device 20 is also used to construct a semantic association graph by taking relevant document fragments with semantic similarity exceeding a preset threshold as nodes and co-occurrence citation relationships or semantic dependency relationships as edges. The node weight reflects the association strength between the document fragment and the original text, and the edge weight reflects the logical association strength between the document fragments.

[0093] In practice, a similarity threshold of 0.7 can be set, and only document fragments with a score ≥ 0.7 are selected as nodes. The node weight is directly taken from this similarity score. There are two ways to construct edges:

[0094] Co-occurrence citation relationship: If two document fragments cite the same third-party document, an edge is established between them. For example, if fragments A and B both cite "AttentionIsAllYouNeed", the edge weight is set to the number of co-citations, 2.

[0095] Semantic dependency: Calculate the dot product similarity of the academic semantic vectors corresponding to the two segments. If the similarity is greater than 0.6, an edge is created, and the edge weight is the similarity value. For example, segment C discusses "multi-head attention" and segment D discusses "self-attention mechanism". The similarity between their vectors is 0.75, so an edge with a weight of 0.75 is added.

[0096] The original text serves as the central node, and an edge is automatically added between it and each document fragment node. The edge weight is equal to the node weight of that fragment.

[0097] As one implementation, the cloud-based large language model server 20 is also used to parse the noise mask identifier through the privacy-aware gating mechanism to extract the noise dimension label sequence. The noise dimension label sequence records the position index of the noise dimension injected with structured noise in the perturbation semantic embedding and the corresponding noise type identifier.

[0098] The privacy-aware gating mechanism reads the upper-level structure of the privacy-protected query vector, namely the noise mask identifier. This identifier is a list, recording the dimensional position index of each injected noise and the corresponding noise type identifier. For example, the mask identifier content is: [{dim:5,type:"Gaussian"},{dim:12,type:"Laplace"},{dim:23,type:"Gaussian"}], indicating that Gaussian noise was added to the 5th and 23rd dimensions, and Laplace noise was added to the 12th dimension.

[0099] The cloud-based large language model server 20 is also used to call the corresponding denoising decoder based on the noise type identifier, perform typed noise stripping on the noise dimension, and obtain the denoised dimension value. The denoising decoder includes at least one of a Bayesian posterior estimator for Gaussian noise, a sparse reconstructor for Laplace noise, and an adaptive gated fusion unit for mixed noise.

[0100] For the Gaussian noise dimension, the Bayesian posterior estimator is invoked. Based on the preset prior distribution (such as a Gaussian prior with a mean of 0 and a variance of σ²) and the observed noisy values, the posterior expectation is calculated as the denoised dimension value.

[0101] For the Laplace noise dimension, the sparse reconstructor is invoked, and L1 regularization is used to optimize and recover the true value, since Laplace noise corresponds to sparse prior.

[0102] For mixed noise dimensions (such as partial Gaussian plus partial Laplacian), an adaptive gated fusion processor is invoked. This fusion processor consists of a small neural network that dynamically learns the weighting coefficients of the two denoising results and outputs the fused dimension value.

[0103] For example, the fifth dimension has Gaussian noise, with an observed value of 0.78, and the Bayesian estimator calculates a denoised value of 0.92.

[0104] The cloud-based large language model server 20 is also used to perform attention weighted interpolation on the denoised dimension value and the adjacent unperturbed dimension to compensate for the semantic loss introduced by noise stripping and obtain the masked vector.

[0105] It should be understood that the denoising process may lose some semantic information, therefore compensation is needed using the unperturbed values ​​of adjacent dimensions. For each denoised dimension, take its two adjacent dimensions before and after it in the vector that have not been injected with noise, and calculate the attention weights between these adjacent dimensions and the denoised dimension. The weights are obtained by taking the dot product of the values ​​of each adjacent dimension after a linear transformation and the value of the denoised dimension, and then normalizing them using softmax.

[0106] The final interpolation result is the weighted average of the denoised value and its adjacent dimensions. For example, if the denoised value of the 5th dimension is 0.92, and the adjacent unperturbed values ​​of the 4th and 6th dimensions are 0.85 and 0.88 respectively, with attention weights of 0.4 and 0.6, then the interpolated result is 0.92 × 1.0 (with its own weight retained) + 0.85 × 0.4 + 0.88 × 0.6, which is then re-normalized. In the actual implementation, the value itself also participates in the attention calculation. The vector output from this step is the masked vector, used for subsequent inference.

[0107] As one implementation, the cloud-based large language model server 20 is also used to generate a pseudo-random noise sequence that matches the dimensional distribution of the noise mask identifier, wherein the noise intensity of the pseudo-random noise sequence is inversely proportional to the differential privacy budget.

[0108] After generating the refined feature vectors, the cloud-based large language model server needs to inject obfuscation noise into these vectors to protect the returned results from reverse engineering attacks. The noise mask identifier, already read in the previous parsing steps, records the dimensional position index of the structured noise initially injected into the local computing device. Based on this dimensional distribution information, the cloud generates a pseudo-random noise sequence whose length and dimensional position are identical to the mask identifier. The pseudo-random sequence is generated using a deterministic random number generator; the seed can be based on a session identifier or timestamp, but does not need to be shared locally.

[0109] Noise intensity is inversely proportional to the differential privacy budget: a larger differential privacy budget indicates a higher requirement for native privacy protection, but the intensity of cloud-injected obfuscation noise is actually lower. This is to avoid excessive degradation of polishing quality by strong dual noise. For example, when the locally allocated high-risk privacy budget ε=1.0, the cloud-injected noise intensity is set to σ=0.1; while when the low-risk budget ε=0.1, the cloud-injected noise intensity σ=0.5. The specific mapping relationship can be expressed as σ=k / ε, where k is a preset constant.

[0110] The cloud-based large language model server 20 is also used to superimpose the pseudo-random noise sequence onto the corresponding dimension of the polished feature vector to generate a privacy-preserving result vector.

[0111] The cloud-based system, according to the dimension positions recorded by the mask identifier, sequentially adds generated pseudo-random noise to the same dimensions of the refined feature vector. Dimensions not recorded by the mask identifier retain their original values. The addition method is used: new dimension value = original refined dimension value + corresponding noise value. For example, if the mask identifier indicates that dimensions 3, 7, and 15 need noise, and the cloud generates three noise values ​​of +0.05, -0.12, and +0.08 respectively, these will be added to dimensions 3, 7, and 15 of the refined feature vector. After the addition is complete, the resulting vector is the privacy-preserving result vector, which is returned to the local computing device.

[0112] In one implementation, the local computing device 10 is further configured to map the polished candidate text to the semantic association graph and detect the consistency of the academic terms in the polished candidate text with the terminology used in the node document fragments in the semantic association graph.

[0113] After the local computing device removes obfuscation noise from the privacy-preserving result vector and decodes it, it obtains the polished candidate text.

[0114] Furthermore, the text is mapped back to the previously constructed semantic association graph. The mapping method is as follows: extract all academic terms (such as "Generative Adversarial Network" and "Batch Normalization") from the polished candidate text and compare them with the terms contained in the document fragments of each node in the semantic association graph. Detecting the consistency of term usage means checking whether the same concept is used in different positions. For example, the original text uses "Generative Adversarial Network", and whether the polished text is incorrectly changed to "Generative Adversarial Network". Detecting citation conflicts means checking whether the documents cited in the polished text match the citation relationships of the nodes in the semantic association graph. For example, the original text cites a key document [1], and whether the polished text incorrectly deletes the citation or changes it to an irrelevant document.

[0115] The local computing device 10 is also used to, if term drift or citation conflict is detected, re-retrieve alternative document fragments based on the edge weight paths of the semantic association graph, and trigger the cloud-based large language model server to perform local re-inference until the verification is passed.

[0116] When term drift is detected, such as when the polished text mistakenly writes "Long Short-Term Memory Network" as "Long Short-Term Memory Model," the local computing device searches for the most relevant node document fragment in the semantic association graph.

[0117] Specifically, standard terms from the document fragment are extracted as the basis for correction by following the path with higher edge weights in the graph (for example, starting from the original text node and passing through an edge with a weight greater than 0.8 to reach a certain document node).

[0118] When a citation conflict is detected, such as when the polished text cites a document not found in the semantic association graph, the system utilizes the co-occurrence citation relationships in the graph structure to retrieve an alternative, semantically similar, and correct document fragment via edge weight paths. This triggers a local re-inference on the cloud-based large language model server: the polished result is regenerated only for the conflicting sentence or paragraph, rather than the entire text being rewritten. The re-inference request includes locally provided information on the correct terminology or alternative document fragment.

[0119] Repeat the above verification process until the candidate text passes the consistency check, and then output the final polished paper text. If the process fails after more than a preset number of retries, the system marks that part as awaiting manual review.

[0120] For ease of understanding, the following examples are provided, but they are not intended to limit the academic paper editing system used in this application.

[0121] In a practical academic paper polishing application scenario, researcher A from a research institution completed a draft of a paper on time series prediction, which included a section of original text awaiting polishing: "This study uses an LSTM network to predict time series data. The experiment uses the Adam optimizer with a learning rate of 0.001. Author: Zhang San, affiliated with the School of Computer Science at a university, grant number: 2024-ZD-001. Accuracy on the unpublished test set reached 92.3%." Researcher A input this text into the academic paper polishing system. The specific steps are as follows:

[0122] I. Processing of local computing devices

[0123] The local computing device (A's personal workstation) first performs word segmentation and part-of-speech tagging on the original text, identifying academic semantic elements: LSTM network (academic term), time series data prediction (research method description), Adam optimizer (academic term), learning rate = 0.001 (experimental data), and test set accuracy of 92.3% (experimental data). It also identifies privacy-sensitive elements: Zhang San (author identity, high risk), School of Computer Science of a certain university (institutional affiliation, medium risk), grant number 2024-ZD-001 (funding project, low risk), and unpublished test set accuracy of 92.3% (unpublished experimental data, high risk).

[0124] The system invokes a 128-dimensional word embedding model pre-trained on an arXiv paper to encode each academic semantic element as a vector. The vector corresponding to the LSTM network is [0.12, -0.34, 0.56, …, 0.78], the Adam optimizer vector is [0.45, 0.21, -0.13, …, 0.55], the learning rate vector is [0.01, 0.03, 0.98, …, 0.22], and the accuracy vector is [0.67, -0.09, 0.34, …, 0.89]. An attention pooling mechanism is used with weights of 0.3, 0.2, 0.3, and 0.2, respectively. The weighted sum is used to obtain the first vector V_sem = [0.25, -0.10, 0.48, …, 0.65] (128 dimensions).

[0125] In the privacy subspace, risk levels are assessed based on a pre-defined privacy rule base: author names and unpublished experimental data are marked as high-risk, allocated a differential privacy budget ε=1.0; institutional affiliation is medium-risk, allocated ε=0.5; and grant number is low-risk, allocated ε=0.1. Privacy-sensitive elements are encoded as privacy vectors (orthogonal to the semantic subspace). Structured noise is injected into the positions corresponding to the privacy dimensions in the first vector: the 10th dimension, corresponding to the author name, adds Laplace noise with a scale parameter b=1 / ε=1.0, the original value of 0.23 becomes 1.10 after adding noise of 0.87; the 25th dimension, corresponding to the institution, adds Gaussian noise with a standard deviation σ=1 / ε=2.0, the original value of -0.10 becomes -0.32 after adding noise of -0.22; and the 40th dimension, corresponding to the grant number, adds Gaussian noise with σ=10.0, the original value of 0.65 becomes 2.00 after adding noise of 1.35. The remaining dimensions remain unchanged. This yields the privacy protection vector V_priv = [1.10, -0.10, 0.48, -0.32, …, 2.00].

[0126] II. Constructing a Semantic Relationship Graph

[0127] The local computing device uses the aforementioned academic semantic elements to search the local knowledge base. This knowledge base contains six relevant documents from research group A. The system calculates the semantic similarity between each document fragment and the original text: fragment A, "Application of LSTM in Time Series Forecasting," scores 0.85; fragment B, "Convergence Analysis of Adam Optimizer," scores 0.78; fragment C, "Privacy Protection Techniques in Deep Learning," scores 0.45. The scores of the other three fragments are all below the threshold of 0.7, so only fragments A and B are retained as nodes, with node weights of 0.85 and 0.78, respectively. The co-occurrence citation relationship between the two fragments is analyzed: fragments A and B both cite the same third-party document, "A Review of Recurrent Neural Networks," with a co-occurrence count of 1. Therefore, an edge is established between A and B, with an edge weight of 1.0. Edges are established between the original text nodes and fragments A and B, respectively, with weights set to the corresponding node weights of 0.85 and 0.78. Thus, the semantic association graph contains three nodes and three edges. The system uses a graph isomorphic network encoder to embed the graph into a 128-dimensional second vector V_graph = [0.45, -0.22, 0.33, …, -0.11].

[0128] III. Integrated Upload

[0129] The local computing device fuses the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure. The upper-layer noise mask identifies the dimension index of the injected noise and its corresponding noise type and budget: [(10, "Laplace", 1.0), (25, "Gaussian", 2.0), (40, "Gaussian", 10.0)]. The middle layer is a semantic association graph topology, storing the edge weights between nodes using a 3×3 adjacency matrix. The lower layer is a perturbation semantic embedding V_priv. This query vector is uploaded to a cloud-based large language model server via the network.

[0130] IV. Cloud-based parsing and noise reduction

[0131] After receiving the query vector, the cloud server uses a privacy-aware gating mechanism in its embedded layer to parse the upper-layer noise mask identifier and extract the noise dimension label sequence. For the 10th dimension Laplace noise, a sparse reconstructor is invoked for denoising. This reconstructor is optimized based on the L1 norm, and with an input observation value of 1.10, the estimated true value is 0.21. For the 25th dimension Gaussian noise (standard deviation 2.0), a Bayesian posterior estimator is invoked. Assuming the prior distribution of the true value is a standard normal distribution, the posterior expectation is calculated as (observation value / 4 + 0) / (1 / 4 + 1) = 1.10 / 5 = 0.22. For the 40th dimension Gaussian noise (standard deviation 10.0), a similar calculation yields a posterior expectation of approximately 0.01. After denoising, the system performs attention-weighted interpolation between each denoised dimension and its adjacent unperturbed dimensions: taking the 9th dimension (value 0.35) and the 11th dimension (value 0.42), the attention weight is calculated (normalized by softmax through dot product similarity), resulting in a compensated value of 0.21×0.5 + 0.35×0.25 + 0.42×0.25 = 0.2975 for the 10th dimension. In the actual implementation, the weight is relatively high, so it is finally set to 0.28. The final masked vector V_clean = [0.28, -0.10, 0.48, 0.22, …, 0.01].

[0132] V. Cloud-based Restricted Search and Collaborative Reasoning

[0133] The semantic association graph topology of the middle layer is parsed in the cloud, and node identifiers A and B are extracted. Using these identifiers as search keys, a same-domain search is performed in the cloud knowledge base (including publicly available academic literature databases), with a matching threshold of 0.75. Node A retrieves 5 cloud-based literature fragments, with a highest matching score of 0.92; node B retrieves 3, with a highest matching score of 0.81. The system constructs a semantic association strength matrix between nodes, setting an association threshold of 0.5: the edge weight of node A with the original text (0.85) is greater than 0.5, and it is marked as a strongly associated node; the edge weight of node B (0.78) is also greater than 0.5, and it is also a strongly associated node. All search results from strongly associated nodes are retained. Subsequently, the cross-document semantic coherence score of each cloud-based literature fragment and the original text is calculated using a coherence evaluation model fine-tuned on the academic corpus, with a threshold set to 0.6. After calculation, 6 out of 8 fragments scored higher than 0.6, while the other 2, with scores of 0.52 and 0.48 respectively, were removed. The final limited search results were 6 highly relevant cloud-based document fragments.

[0134] The cloud-based large language model (using LLaMA-2-7B, with privacy-aware gating added to the embedding layer) receives the masked vector V_clean and the restricted retrieval results. The model fuses the features of the retrieval results into V_clean through a cross-attention mechanism. After 12 layers of Transformer decoding, a refined feature vector V_refined = [0.33, -0.05, 0.62,…, 0.71] (128 dimensions) is generated.

[0135] VI. Noise in Injection Results

[0136] The cloud generates a pseudo-random noise sequence based on the dimensional distribution of the noise mask identifier. The noise intensity is inversely proportional to the differential privacy budget; taking a scaling factor k=0.5, the cloud noise standard deviation σ_cloud = 0.5 / ε. The original ε=1.0, σ=0.5 for the 10th dimension generates a noise value of +0.12; the original ε=0.5, σ=1.0 for the 25th dimension generates a noise value of -0.24; and the original ε=0.1, σ=5.0 for the 40th dimension generates a noise value of +2.16. These noise values ​​are superimposed on the corresponding dimensions of V_refined to obtain the privacy-preserving result vector V_result = [0.33+0.12, -0.05, 0.62-0.24, …, 0.71+2.16] = [0.45, -0.05, 0.38, …, 2.87]. The cloud returns this vector to the local computing device.

[0137] VII. Local Noise Cancellation and Decoding

[0138] The local computing device stores the original noise mask identifier and differential privacy budget. For each noisy dimension in V_result, the pseudo-random noise injected from the cloud (calculated using the same generation formula to obtain the same noise value) is subtracted. 10th dimension: 0.45 - 0.12 = 0.33; 25th dimension: -0.05 - (-0.24) = 0.19; 40th dimension: 2.87 - 2.16 = 0.71. The remaining dimensions remain unchanged, resulting in the polished intermediate vector V_mid = [0.33, -0.05, 0.19, …, 0.71]. The system calls a multi-layer Transformer decoder paired with the encoder to map V_mid back to natural language text, obtaining the polished candidate text: "This study uses a Long Short-Term Memory network to predict time series data. The Adam optimizer was used in the experiment, with a learning rate set to 0.001. The test set accuracy reached 92.3%."

[0139] VIII. Verification of Academic Consistency

[0140] The local computing device maps the polished candidate text back to the previously constructed semantic association graph. Consistency in academic terminology is checked: "LSTM network" in the original text is rewritten as "Long Short-Term Memory network" in the candidate text. The system confirms that the two are synonymous terms through the literature fragment of node A in the semantic association graph, and the verification passes. Citation conflicts are checked: the candidate text does not introduce any new citations and does not delete logical relationships in the original content, and the verification passes. Privacy elements (author names, affiliations, and grant numbers) have been removed as required. After one verification, the system outputs the final polished paper text: "This study uses Long Short-Term Memory networks to predict time series data. The Adam optimizer was used in the experiment, with a learning rate set to 0.001. The test set accuracy reached 92.3%."

[0141] Throughout the entire processing, the cloud server never accessed the author's name, institution name, grant number, or unpublished experimental data in the original plaintext, yet the polishing quality was professionally improved and privacy was fully protected.

[0142] Reference Figure 2 This application provides an academic paper editing system and a method for editing academic papers. Figure 2 This is a flowchart illustrating the first embodiment of the academic paper polishing method of this application. The academic paper polishing system includes: a local computing device and a cloud-based large language model server. The cloud-based large language model server includes a large language model with a privacy-aware gating mechanism in its embedded layer. The academic paper polishing method includes steps S10 to S70:

[0143] Step S10: The local computing device acquires the original text of the academic paper to be polished, maps the original text to a semantic subspace and a privacy subspace, generates a first vector in the semantic subspace, allocates a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and injects noise into the first vector to obtain a privacy-protected vector.

[0144] In step S20, the local computing device retrieves academic semantic elements extracted from the semantic subspace from the local knowledge base, constructs a semantic association graph between the retrieval results and the original text, and encodes the semantic association graph to obtain a second vector.

[0145] Step S30: The local computing device fuses the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and uploads it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding.

[0146] In step S40, the cloud-based large language model server receives the privacy-protected query vector, parses the noise mask identifier through the privacy-aware gating mechanism to locate and mask the noise dimension in the perturbation semantic embedding, and obtains the masked vector.

[0147] In step S50, the cloud-based large language model server performs a restricted search in the cloud knowledge base based on the semantic association graph topology, and performs collaborative reasoning with the restricted search results and the masked vector to generate a refined feature vector. Then, it injects result obfuscation noise that matches the noise mask identifier into the refined feature vector to obtain a privacy-protected result vector and returns it to the local computing device.

[0148] In step S60, the local computing device receives the privacy protection result vector, eliminates the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtains a refined intermediate vector.

[0149] In step S70, the local computing device decodes the intermediate vector for polishing to obtain polishing candidate text, performs academic consistency verification on the polishing candidate text based on the semantic association graph, and outputs the polished paper text.

[0150] The academic paper polishing method provided in this application and the aforementioned academic paper polishing system belong to the same inventive concept and are specific embodiments thereof. The steps described in the academic paper polishing method completely correspond to the functional modules executed by the local computing device and the cloud-based large language model server in the academic paper polishing system. Both employ the same technical means, solve the same technical problems, and achieve the same technical effects. The specific implementation methods and data interaction processes are also consistent. Therefore, the technical features in the method embodiment correspond one-to-one with the system embodiment; for details, please refer to the detailed description of the aforementioned system.

[0151] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the academic paper polishing methods of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0152] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0153] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An academic paper polishing system, characterized in that, The academic paper editing system includes: a local computing device and a cloud-based large language model server, wherein the cloud-based large language model server contains a large language model with a privacy-aware gating mechanism in the embedded layer; The local computing device is used to acquire the original text of the academic paper to be polished, map the original text to a semantic subspace and a privacy subspace, generate a first vector in the semantic subspace, allocate a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and inject noise into the first vector to obtain a privacy protection vector. The local computing device is also used to retrieve academic semantic elements extracted from the semantic subspace from the local knowledge base, construct a semantic association graph between the retrieval results and the original text, and encode the semantic association graph to obtain a second vector; The local computing device is further configured to fuse the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and upload it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding. The cloud-based large language model server is used to receive the privacy-protected query vector, and parse the noise mask identifier through the privacy-aware gating mechanism to locate and mask the noise dimension in the perturbation semantic embedding to obtain the masked vector. The cloud-based large language model server is also used to perform restricted retrieval in the cloud knowledge base based on the semantic association graph topology, and to perform collaborative reasoning with the restricted retrieval results and the masked vector to generate a polished feature vector. The polished feature vector is then injected with result obfuscation noise that matches the noise mask identifier to obtain a privacy-protected result vector and return it to the local computing device. The local computing device is further configured to receive the privacy-preserving result vector, eliminate the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtain a refined intermediate vector; The local computing device is also used to decode the polishing intermediate vector to obtain polishing candidate text, and after performing academic consistency verification on the polishing candidate text based on the semantic association graph, output the polished paper text.

2. The academic paper polishing system as described in claim 1, characterized in that, The local computing device is also used to perform word segmentation and part-of-speech tagging on the original text, and to identify at least one type of academic semantic element in academic terms, research method descriptions, and experimental data; The local computing device is also used to encode the academic semantic elements into semantic vectors based on a word embedding model, construct a semantic subspace using the spanned space of the semantic vectors, and generate a first vector in the semantic subspace. The local computing device is also used to synchronously identify at least one type of privacy-sensitive element in the original text, including author identity information, institutional affiliation, funding project number, and unpublished experimental data, and to assess the semantic exposure risk level of various privacy-sensitive elements based on a preset privacy rule base. The local computing device is further configured to encode the privacy-sensitive elements into privacy vectors, construct a privacy subspace using the spanned space of the privacy vectors, and allocate a differential privacy budget in the privacy subspace based on the semantic exposure risk level.

3. The academic paper polishing system as described in claim 1, characterized in that, The risk levels of the privacy-sensitive elements include high-risk, medium-risk, and low-risk privacy-sensitive elements. The allocation of differential privacy budgets includes a first differential privacy budget, a second differential privacy budget, and a third differential privacy budget. The local computing device is further configured to allocate the high-risk privacy-sensitive elements to the first differential privacy budget, allocate the medium-risk privacy-sensitive elements to the second differential privacy budget, and allocate the low-risk privacy-sensitive elements to the third differential privacy budget, wherein the first differential privacy budget is greater than the second differential privacy budget, and the second differential privacy budget is greater than the third differential privacy budget. The local computing device is further configured to inject structured noise of corresponding intensity into the first vector according to the differential privacy budget to obtain a privacy-preserving vector, wherein the structured noise is Gaussian noise or Laplace noise, and the noise intensity of the privacy-preserving vector is positively correlated with the differential privacy budget.

4. The academic paper polishing system as described in claim 1, characterized in that, The local computing device is also used to retrieve relevant literature fragments from the local knowledge base based on the academic semantic elements extracted from the semantic subspace, and to calculate the semantic similarity between the relevant literature fragments and the original text; The local computing device is also used to construct a semantic association graph by taking relevant document fragments with semantic similarity exceeding a preset threshold as nodes and taking co-occurrence citation relationships or semantic dependency relationships as edges. The node weight reflects the association strength between the document fragment and the original text, and the edge weight reflects the logical association strength between the document fragments.

5. The academic paper polishing system as described in claim 1, characterized in that, The cloud-based large language model server is also used to parse the noise mask identifier through the privacy-aware gating mechanism to extract the noise dimension label sequence. The noise dimension label sequence records the position index of the noise dimension injected with structured noise in the perturbation semantic embedding and the corresponding noise type identifier. The cloud-based large language model server is also used to call the corresponding denoising decoder based on the noise type identifier, perform typed noise stripping on the noise dimension, and obtain the denoised dimension value. The denoising decoder includes at least one of a Bayesian posterior estimator for Gaussian noise, a sparse reconstructor for Laplace noise, and an adaptive gated fusion for mixed noise. The cloud-based large language model server is also used to perform attention weighted interpolation on the denoised dimension value and the adjacent unperturbed dimension to compensate for the semantic loss introduced by noise stripping and obtain the masked vector.

6. The academic paper polishing system as described in claim 1, characterized in that, The cloud-based large language model server is also used to extract node identifiers and edge weights by parsing the topology of the semantic association graph; The cloud-based large language model server is also used to perform a same-domain search in the cloud knowledge base using the node identifier as the search key, and return cloud document fragments with a matching degree higher than the threshold of the node identifier. The cloud-based large language model server is also used to sort and filter the cloud-based document fragments according to the edge weights, retaining cloud-based document fragments that have a direct semantic relationship with the original text as limited search results.

7. The academic paper polishing system as described in claim 6, characterized in that, The cloud-based large language model server is also used to construct a semantic association strength matrix between nodes based on the edge weights, and to mark the nodes connected by edges whose edge weights are lower than a preset association threshold as weakly associated nodes, and the nodes connected by edges whose edge weights are higher than the preset association threshold as strongly associated nodes. The cloud-based large language model server is also used to perform initial screening of the cloud-based document fragments, remove cloud-based document fragments originating from the weakly related nodes, and retain a set of candidate cloud-based document fragments originating from the strongly related nodes. The cloud-based large language model server is also used to calculate the cross-document semantic coherence score between each cloud document fragment in the candidate cloud document fragment set and the original text, and retain the cloud document fragments with cross-document semantic coherence scores higher than the coherence threshold as restricted search results.

8. The academic paper polishing system as described in claim 1, characterized in that, The cloud-based large language model server is also used to generate a pseudo-random noise sequence that matches the dimensional distribution of the noise mask identifier, wherein the noise intensity of the pseudo-random noise sequence is inversely proportional to the differential privacy budget. The cloud-based large language model server is also used to superimpose the pseudo-random noise sequence onto the corresponding dimension of the polished feature vector to generate a privacy-preserving result vector.

9. The academic paper polishing system as described in claim 1, characterized in that, The local computing device is also used to map the polished candidate text to the semantic association graph and detect the consistency of the academic terms in the polished candidate text with the terminology used in the document fragments at the nodes in the semantic association graph; The local computing device is also used to, if term drift or citation conflict is detected, re-retrieve alternative document fragments based on the edge weight paths of the semantic association graph, and trigger the cloud-based large language model server to perform local re-inference until the verification is passed.

10. A method for polishing academic papers, characterized in that, The academic paper polishing method is applied to an academic paper polishing system, which includes: a local computing device and a cloud-based large language model server, wherein the cloud-based large language model server contains a large language model with a privacy-aware gating mechanism in its embedding layer; the method includes: The local computing device acquires the original text of the academic paper to be polished, maps the original text to a semantic subspace and a privacy subspace, generates a first vector in the semantic subspace, allocates a differential privacy budget based on the risk level of privacy-sensitive elements in the privacy subspace, and injects noise into the first vector to obtain a privacy protection vector. The local computing device retrieves academic semantic elements extracted from the semantic subspace from the local knowledge base, constructs a semantic association graph between the retrieval results and the original text, and encodes the semantic association graph to obtain a second vector; The local computing device fuses the privacy-preserving vector with the second vector into a privacy-preserving query vector with a hierarchical coding structure and uploads it to the cloud-based large language model server. The hierarchical coding structure includes an upper-layer noise mask identifier, a middle-layer semantic association graph topology, and a lower-layer perturbation semantic embedding. The cloud-based large language model server receives the privacy-protected query vector and uses the privacy-aware gating mechanism to parse the noise mask identifier to locate and mask the noise dimension in the perturbation semantic embedding, thereby obtaining the masked vector. The cloud-based large language model server performs restricted retrieval in the cloud knowledge base based on the semantic association graph topology, and performs collaborative reasoning with the restricted retrieval results and the masked vector to generate a refined feature vector. The refined feature vector is then injected with result obfuscation noise that matches the noise mask identifier to obtain a privacy-protected result vector and return it to the local computing device. The local computing device receives the privacy-preserving result vector, eliminates the result obfuscation noise based on the correspondence between the noise mask identifier and the differential privacy budget, and obtains a refined intermediate vector; The local computing device decodes the intermediate vector for polishing to obtain polishing candidate text, and performs academic consistency verification on the polishing candidate text based on the semantic association graph, and then outputs the polished paper text.

Images