Intelligent collaborative editing, publishing, auditing and proofreading cloud system

Abstract

The invention relates to the technical field of editing, publishing, auditing and proofreading, and discloses an intelligent collaborative editing, publishing, auditing and proofreading cloud system which comprises a distributed cloud storage module, an AI proofreading engine, a collaborative interaction module and a terminal adapter interface. According to the intelligent collaborative editing, publishing, auditing and proofreading cloud system, an AI proofreading engine carries out preprocessing such as word segmentation and part-of-speech tagging on an input text through a BERT-base and BiLSTM hybrid network in combination with a professional word library unit and a sensitive word detection unit, then sentence vectors and sequence features are spliced, grammar and logic errors are accurately recognized, correction suggestions are generated, manual intervention is reduced, and the efficiency is improved. Massive UGC contents can be dealt with; according to the cooperative interaction module, by means of WebSocket long connection and an OT algorithm, editing operation is disassembled into atomic operation, a version number is attached, a block containing a timestamp and a user ID is recorded in cooperation with an Ethereum private chain, real-time synchronization and operation tracing are achieved, multi-person cooperation conflicts are avoided, responsibility is defined, and meanwhile an AI engine improves adaptability through multi-field text fine tuning, and the problem that a model is single is solved.

Description

Technical Field

[0001] This invention relates to the field of editing, publishing, reviewing, and proofreading technology, specifically to an intelligent collaborative editing, publishing, reviewing, and proofreading cloud system. Background Technology

[0002] Editing, publishing, reviewing, and proofreading are crucial steps in the publishing process to ensure content quality. This involves systematically checking and correcting the text content, formatting specifications, logical structure, and copyright information of publications using manual or technical means. It covers tasks such as identifying typos, verifying grammar, judging logical coherence, checking the standardization of professional terminology, and controlling the consistency of content across multiple versions, aiming to ensure that publications meet publishing standards and dissemination requirements.

[0003] Currently, there are two prominent technical problems in editing, publishing, reviewing, and proofreading: First, manual reviewing and proofreading relies on the professional experience and attention of editors, which is not only inefficient but also prone to errors due to fatigue when processing large amounts of text. Especially in the context of the explosion of UGC content, manual review is difficult to cope with the pressure of reviewing massive amounts of information. Second, even with intelligent reviewing and proofreading technology, it often relies on a single model and lacks adaptive optimization for various types of text in the publishing field. It has insufficient accuracy in identifying complex grammatical and logical errors and non-standard use of professional terminology. At the same time, in multi-person collaborative scenarios, there is a lack of effective operation synchronization and version tracking mechanisms, which leads to frequent version conflicts and difficulty in defining responsibility during the proofreading process. Therefore, an intelligent collaborative editing, publishing, reviewing, and proofreading cloud system is proposed. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent collaborative editing, publishing, review, and proofreading cloud system. It features an AI proofreading engine that enables high-precision error detection and correction suggestion generation, and a collaborative interaction module that supports real-time collaborative editing by multiple users. This system solves the problems of low efficiency in manual review, insufficient adaptability of intelligent proofreading, and the tendency for conflicts and difficulties in defining responsibilities among multiple users in existing technologies.

[0006] (II) Technical Solution

[0007] To achieve the goal of high-precision error detection and correction suggestion generation by the AI ​​proofreading engine, and to support real-time collaborative editing by multiple users in conjunction with the collaborative interaction module, this invention provides the following technical solution: an intelligent collaborative editing, publishing, reviewing and proofreading cloud system, including a distributed cloud storage module, an AI proofreading engine, a collaborative interaction module and a terminal adaptation interface;

[0008] The distributed cloud storage module consists of three or more server nodes, each node is configured with a 2TB or more SSD storage unit, data is sharded and stored between nodes through a hash algorithm, and RAID5 technology is used for redundancy backup.

[0009] The AI ​​proofreading engine includes a pre-trained BERT-base model and a BiLSTM hybrid network. The BERT layer adopts a 12-layer Transformer structure, the BiLSTM layer sets 64-dimensional hidden states, the output layer adopts the Softmax activation function, and the model parameters are serialized and stored on the cloud server through the TensorFlow framework.

[0010] The collaborative interaction module has a built-in WebSocket protocol processing unit and a blockchain node. The WebSocket unit establishes a long connection using port 8080, and the blockchain node adopts the Ethereum private chain architecture. Each editing operation generates a block containing a timestamp, user ID, and operation content, and the block size is limited to 2MB.

[0011] The terminal adaptation interface supports both HTTP and HTTPS protocols and is compatible with Android 8.0 and above, iOS 12.0 and above, and Windows 10 and above operating systems. The interface data format adopts Protobuf serialization.

[0012] Preferably, the model training process of the AI ​​proofreading engine includes training dataset construction, pre-training stage, and fine-tuning stage;

[0013] The training dataset was constructed by collecting 1 million articles from the publishing field, manually annotating them with 12 types of grammatical errors and 8 types of logical errors to form an annotated corpus, which was then divided into training, validation, and test sets in an 8:1:1 ratio.

[0014] During the pre-training phase, a learning rate of 5e-5 and a batch size of 32 are used. The training set is iterated for 10 rounds using the Adam optimizer. After each round, the F1 score is calculated on the validation set. Pre-training is stopped when the F1 score improves by less than 0.01 for 3 consecutive rounds.

[0015] During the fine-tuning phase, 50,000 domain-specific texts were introduced, including those related to science, literature, and education. The learning rate was adjusted to 1e-5, and the process was repeated 5 times. A gradient accumulation strategy was used to process small batches of data, and the final model achieved an accuracy of no less than 98.5% on the test set.

[0016] Preferably, the collaborative interaction module includes a real-time differential synchronization unit and an annotation layer rendering engine;

[0017] The real-time differential synchronization unit uses the OT algorithm to decompose the editing operation into three atomic operations: insertion, deletion, and replacement. Each operation is accompanied by a version number, which is in the format of YYYYMMDDHHMMSS plus the last 4 digits of the user ID.

[0018] The annotation layer rendering engine uses Canvas technology to implement annotation overlay. Annotation data includes coordinate information, annotation type, and color value. The coordinate information is accurate to the pixel level. Annotation types include text, highlight, and doodle. The color value is in RGB format and is stored in association with the original data through a timestamp.

[0019] Preferably, the distributed cloud storage module further includes a document parsing unit and a version management unit;

[0020] The document parsing unit supports PDF, DOCX, and EPUB format conversion. During the conversion process, OCR technology is used to process scanned documents, with a recognition accuracy of no less than 99.2%. After text extraction, an intermediate file in XML format is generated.

[0021] The version management unit adopts an incremental storage mechanism, which only saves the data that differs from the previous version. Each version is associated with a unique SHA-256 checksum, and the checksum is synchronized to the blockchain node in real time through the API interface.

[0022] Preferably, the operation of the AI ​​proofreading engine includes text preprocessing, error detection, and generation of correction suggestions;

[0023] Text preprocessing involves word segmentation, part-of-speech tagging, and sentence truncation of the input document. Word segmentation uses the Jieba word segmentation library and loads a custom dictionary for the publishing industry. Part-of-speech tagging uses the ICTCLAS tagging system. Sentence truncation limits a single sentence to a maximum of 512 characters.

[0024] Error detection involves concatenating the sentence vector output by the BERT layer with the sequence features extracted by the BiLSTM layer, mapping it to a 20-dimensional error type space through a fully connected layer, and triggering a label when the probability of a certain type of error exceeds 0.85.

[0025] The correction suggestion generation involves calling a pre-trained T5-small model to generate three candidate correction texts for the marked error location. The candidate texts are then compared with the original text using an edit distance algorithm, and the two with the smallest differences are selected as the output.

[0026] Preferably, it also includes an access control module and a log auditing unit;

[0027] The access control module adopts the RBAC model and sets up three levels of roles: editor, reviewer, and administrator. The editor role only has the right to comment (access code 001), the reviewer role has the right to approve (access code 010), and the administrator has the right to configure (access code 100). Access verification is implemented through AES-256 encrypted tokens.

[0028] The log auditing unit generates an operation log every 5 minutes. The log content includes IP address, operation type, and data ID. The IP address is encrypted with MD5. The log file is automatically uploaded to the cloud storage cold backup node.

[0029] Preferably, the blockchain node of the collaborative interaction module interacts with the WebSocket protocol processing unit of the module through a JSON-RPC interface, and its block structure includes a block header and a block body;

[0030] The block header contains the previous block hash, Merkle root, and timestamp. The previous block hash and Merkle root are both 32 bytes, and the timestamp is in UTC format.

[0031] The block contains 50-100 operation transactions. Each transaction contains the operator's public key, the hash of the operation content, and a digital signature. The digital signature uses the ECDSA algorithm.

[0032] The consensus mechanism adopts Proof-of-Agent (PoA), with three or more trusted nodes taking turns generating blocks. The trusted nodes are the audit administrator accounts, and the block generation interval is 10 seconds.

[0033] Preferably, the AI ​​proofreading engine also includes a professional lexicon unit and a sensitive word detection unit;

[0034] The professional terminology unit stores 800,000 professional terms in the publishing industry, and constructs a prefix tree index according to subject classification. The terminology is updated through an incremental synchronization mechanism, and is reconciled with the central terminology server at 3:00 AM every day.

[0035] The sensitive word detection unit uses the AC automaton algorithm, has a built-in database of 100,000 sensitive words, and takes no more than 100ms per thousand words to detect. When a word is detected, a red highlight is triggered and the submission operation is blocked.

[0036] Preferably, the formula recognition module of the document parsing unit is the core sub-module of the unit. It uses the YOLOv5 algorithm to detect formula areas in the document and combines the Mathpix API to perform formula-to-LaTeX conversion, with a recognition accuracy of not less than 95%.

[0037] The table extraction module of the document parsing unit uses a combination of OpenCV edge detection and Tesseract-OCR to convert the table structure into HTML format while preserving attributes such as cell merging and border styles.

[0038] Preferably, the terminal adapter interface further includes an offline caching unit and a resume interrupted download module;

[0039] The offline caching unit supports local caching of the three most recently edited documents on the client side. The cached data is stored in an SQLite database, and incremental synchronization is achieved by comparing version numbers when the network is connected again.

[0040] The breakpoint resume module uses the HTTP Range request header to implement large file chunked uploads, with each chunk being 5MB in size. The upload progress is fed back to the client in real time via WebSocket.

[0041] (III) Beneficial Effects

[0042] Compared with existing technologies, this invention provides an intelligent collaborative editing, publishing, reviewing, and proofreading cloud system, which has the following beneficial effects:

[0043] 1. This intelligent collaborative editing, publishing, reviewing, and proofreading cloud system uses an AI proofreading engine that employs a hybrid network of BERT-base and BiLSTM, combined with a professional lexicon unit and a sensitive word detection unit. It first performs preprocessing on the input text, such as word segmentation and part-of-speech tagging, and then concatenates sentence vectors and sequence features to accurately identify grammatical and logical errors and generate correction suggestions. This reduces human intervention, addresses massive amounts of UGC content, and solves the problems of low efficiency and easy omissions in manual processes.

[0044] 2. This intelligent collaborative editing, publishing, reviewing, and proofreading cloud system utilizes WebSocket long connections and OT algorithms in its collaborative interaction module to break down editing operations into atomic operations with version numbers. Combined with Ethereum private chain records blocks containing timestamps and user IDs, it achieves real-time synchronization and operation traceability, avoids conflicts in multi-person collaboration, clarifies responsibilities, and improves the adaptability of the AI ​​engine through multi-domain text fine-tuning, solving the problem of single model. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the cloud system architecture for editing, publishing, reviewing, and proofreading in this invention. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1A smart collaborative editing, publishing, reviewing, and proofreading cloud system, including a distributed cloud storage module, an AI proofreading engine, a collaborative interaction module, and a terminal adaptation interface;

[0048] The distributed cloud storage module consists of three or more server nodes, each node is configured with a 2TB or more SSD storage unit, data is sharded and stored between nodes through a hash algorithm, and RAID5 technology is used for redundancy backup.

[0049] The AI ​​proofreading engine includes a pre-trained BERT-base model and a BiLSTM hybrid network. The BERT layer adopts a 12-layer Transformer structure, the BiLSTM layer sets 64-dimensional hidden states, the output layer adopts the Softmax activation function, and the model parameters are serialized and stored on the cloud server through the TensorFlow framework.

[0050] The collaborative interaction module has a built-in WebSocket protocol processing unit and a blockchain node. The WebSocket unit establishes a long connection using port 8080, and the blockchain node adopts the Ethereum private chain architecture. Each editing operation generates a block containing a timestamp, user ID, and operation content, and the block size is limited to 2MB.

[0051] The terminal adaptation interface supports both HTTP and HTTPS protocols and is compatible with Android 8.0 and above, iOS 12.0 and above, and Windows 10 and above operating systems. The interface data format adopts Protobuf serialization.

[0052] Example 1:

[0053] This embodiment details the complete implementation steps of an AI proofreading engine based on a hybrid network model, from dataset construction to actual operation.

[0054] First, a training dataset was constructed by collecting 1 million audiobook texts covering literature, popular science, children's stories, and other genres. Twelve types of grammatical errors were manually labeled, including misuse of "de / de / di" and mismatched punctuation, as well as eight types of logical errors, including plot temporal contradictions and scientific fact errors. The labeled corpus was then divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0055] For children's audiobook texts, additional annotations are made for specific error types such as "sentences are too long" and "words are not appropriate for the age group's cognitive level" to improve the model's adaptability to specific scenarios.

[0056] Next, model training of the AI proofreading engine is carried out. The BERT-base model is adopted, which contains 12 layers of Transformer structures and a hybrid network with BiLSTM. This network is set with a 64-dimensional hidden state, and the serialization storage of model parameters is completed through the TensorFlow framework;

[0057] In the pre-training stage, the learning rate is set to 5e-5, the batch size is 32, and it is iterated 10 rounds on the training set. The Adam optimizer is used. After each round, the F1 value is calculated on the validation set. When the improvement of the F1 value is less than 0.01 for three consecutive rounds, the pre-training is stopped;

[0058] Subsequently, 50,000 pieces of audiobook text in a specific domain are introduced for fine-tuning. The learning rate is adjusted to 1e-5, and it is iterated 5 rounds. The gradient accumulation strategy is adopted to process small batches of data. Finally, the accuracy of the model on the test set reaches more than 98.5%.

[0059] During actual operation, first, the input audiobook text is preprocessed. The Jieba word segmentation library is adopted. This word segmentation library loads an industry-customized dictionary containing terms such as "narrator", "sound effect", "anchor", etc. for word segmentation, and the word tagging is completed through the ICTCLAS annotation system, and the length of each single sentence is truncated to within 512 characters;

[0060] In the error detection stage, the sentence vector output by the BERT layer and the sequence features extracted by the BiLSTM layer are concatenated, and mapped to a 20-dimensional error type space through a fully connected layer. When the probability of a certain type of error exceeds 0.85, a mark is triggered;

[0061] For example, when it is detected that the use of "的" in "他兴奋的奔跑" is incorrect and the probability is 0.92, the pre-trained T5-small model is called to generate 3 candidate correction texts, and the 2 with the smallest difference are selected through the edit distance algorithm, such as "他兴奋地奔跑" and "他兴奋地跑着", as the output result.

[0062] Embodiment 2:

[0063] This embodiment details the specific operation method of the collaborative interaction module based on the private chain architecture from deployment to operation.

[0064] In the system deployment stage, blockchain nodes based on the Ethereum private chain architecture are deployed in the collaborative interaction module. 3 trusted nodes served by audit administrator accounts are configured. The PoA consensus mechanism is adopted, and the block generation interval is set to 10 seconds;

[0065] Each node conducts data interaction with the WebSocket protocol processing unit through the JSON-RPC interface. This processing unit establishes a long connection using port 8080 to ensure the real-time chain-up of editing operations.

[0066] When users upload audiobook UGC content, editors can make annotations through the system terminal. At this time, the real-time differential synchronization unit of the collaborative interaction module decomposes the operation into three atomic operations: insertion, deletion, and replacement. Each operation is accompanied by a version number in the format "YYYYMMDDHHMMSS + the last 4 digits of the user ID".

[0067] For example, when an editor performs the operation "delete 'error sound effect description' and insert 'suggest adding bird chirping sound effect'" on a piece of text, the system automatically generates an atomic operation record containing the operation type, location coordinates, and content hash.

[0068] Blockchain nodes package 50-100 transactions into a block, with the block size strictly controlled within 2MB.

[0069] The block header contains the hash of the previous block (32 bytes), the Merkle root (32 bytes), and a UTC-formatted timestamp. The block body records the public key of the operator for each transaction, the hash of the operation content (SHA-256 encrypted), and an ECDSA digital signature.

[0070] After the block is generated, it is synchronized to all trusted nodes, enabling full traceability of the operation.

[0071] If multiple people are editing the same text at the same time, the system uses the OT algorithm to sort and merge the operations, and executes the operations in the order of version number to avoid conflicts;

[0072] For example, when editor A and editor B modify the content of paragraph 3 at the same time, the system will prioritize the operation with the earlier version number, and then perform subsequent operations based on that result to ensure the consistency of the editing process;

[0073] The version management unit adopts an incremental storage mechanism, which only saves the data that differs from the previous version. Each version is associated with a unique SHA-256 checksum and is synchronized to the blockchain node in real time through the API interface. It supports querying historical versions by timestamp and user ID.

[0074] Example 3:

[0075] This embodiment details the specific implementation process of the distributed cloud storage module.

[0076] The distributed cloud storage module is deployed with 5 server nodes, each node is configured with a 2TB SSD storage unit, and data is sharded and stored between nodes through a hash algorithm, and RAID5 technology is used for redundancy backup.

[0077] For audio files of audiobooks, including speech synthesis files and user voice replication materials, they are divided into segments of 10 minutes each, supporting MP3 and WAV formats, and using adaptive bitrate encoding of 256-320kbps to balance storage efficiency and sound quality.

[0078] For AR model files, such as 3D animated models in popular science audiobooks, the files are divided into segments of 50,000 triangles each and stored in glTF format to meet the fast loading requirements of VR devices.

[0079] When the document parsing unit processes multimodal content, it uses OCR technology to extract text from scanned documents, with a recognition accuracy of no less than 99.2%, and generates an intermediate file in XML format after extraction.

[0080] For text containing formulas, such as scientific formula explanations in popular science audiobooks, the YOLOv5 algorithm is used to detect formula areas, and the Mathpix API is called to convert it into LaTeX format, ensuring that the recognition accuracy reaches more than 95%.

[0081] For table content, such as character voice information tables, the OpenCV edge detection combined with Tesseract-OCR is used to convert it into HTML format, fully preserving attributes such as cell merging and border styles.

[0082] The system establishes associations between text, audio, and AR models through metadata. For example, "the audio segment corresponding to Chapter 2 of the text is 00:05:10-00:08:30, and the associated AR model ID is 'dino_001'". The metadata is stored in the index nodes of distributed cloud storage and supports retrieval by "content keywords + type". A cold backup mechanism is triggered at 3:00 AM every day to synchronize data to a remote node. Backup files are retained for 30 days and can be quickly restored by version number and content type, ensuring data security and continuity.

[0083] Example 4:

[0084] This embodiment details the specific implementation methods of the access control module and the log auditing unit.

[0085] The access control module adopts the RBAC model and sets up three levels of roles: editor, reviewer, and administrator. The editor role only has the right to comment and the permission code is 001, the reviewer role has the right to approve and the permission code is 010, and the administrator has the right to configure and the permission code is 100.

[0086] When a user logs in, the system verifies permissions using an AES-256 encrypted token. The token is valid for 2 hours, after which the user must re-verify their identity.

[0087] The AI ​​proofreading engine's professional thesaurus unit stores 800,000 academic terms, which are categorized by disciplines such as computer science, medicine, and literature and built into a prefix tree index. It is reconciled and updated with the central thesaurus server every day at 3 a.m. to ensure the timeliness of the terminology database.

[0088] The sensitive word detection unit uses the AC automaton algorithm and has 100,000 built-in words related to academic misconduct. The detection time is controlled within 100ms / thousand words. When words such as "data forgery" and "plagiarism" are detected, a red highlight mark is immediately triggered and the submission operation is blocked. At the same time, an alarm message is sent to the administrator terminal.

[0089] The log auditing unit generates an operation log every 5 minutes, which includes the IP address encrypted by MD5, the operation type, such as "comment", "approved", "format modification", etc., and the data ID. The log file is automatically uploaded to the cloud storage cold backup node.

[0090] For example, the record "User ID:exp_001 performed the approval operation for paper ID:art_886 at 10:30 on 2024-06-15" supports searching by time range, operation type, user role, and other dimensions, providing traceable evidence for compliance review of the academic publishing process.

[0091] In summary, this intelligent collaborative editing, publishing, reviewing, and proofreading cloud system utilizes an AI proofreading engine that employs a hybrid network of BERT-base and BiLSTM, combined with a professional thesaurus unit and a sensitive word detection unit. It first performs preprocessing on the input text, such as word segmentation and part-of-speech tagging, then concatenates sentence vectors and sequence features to accurately identify grammatical and logical errors and generate correction suggestions. This reduces human intervention, addresses massive amounts of UGC content, and solves the problems of low efficiency and easy omissions associated with manual methods.

[0092] Furthermore, this intelligent collaborative editing, publishing, reviewing, and proofreading cloud system utilizes WebSocket long connections and OT algorithms in its collaborative interaction module to break down editing operations into atomic operations with version numbers. Combined with Ethereum private chain records blocks containing timestamps and user IDs, it achieves real-time synchronization and operation traceability, avoiding conflicts in multi-person collaboration and clarifying responsibilities. At the same time, the AI ​​engine has been fine-tuned through multi-domain text to improve adaptability and solve the problem of single model. It addresses the issues of low efficiency in manual review, insufficient adaptability of intelligent proofreading, and easy conflicts and difficulty in defining responsibilities in multi-person collaboration in existing technologies.

[0093] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cloud-based intelligent collaborative editing, publishing, reviewing, and proofreading system, characterized in that: This includes a distributed cloud storage module, an AI proofreading engine, a collaborative interaction module, and a terminal adaptation interface; The distributed cloud storage module consists of three or more server nodes, each node is configured with a 2TB or more SSD storage unit, data is sharded and stored between nodes through a hash algorithm, and RAID5 technology is used for redundancy backup. The AI ​​proofreading engine includes a pre-trained BERT-base model and a BiLSTM hybrid network. The BERT layer adopts a 12-layer Transformer structure, the BiLSTM layer sets 64-dimensional hidden states, the output layer adopts the Softmax activation function, and the model parameters are serialized and stored on the cloud server through the TensorFlow framework. The collaborative interaction module has a built-in WebSocket protocol processing unit and a blockchain node. The WebSocket unit establishes a long connection using port 8080, and the blockchain node adopts the Ethereum private chain architecture. Each editing operation generates a block containing a timestamp, user ID, and operation content, and the block size is limited to 2MB. The terminal adaptation interface supports both HTTP and HTTPS protocols and is compatible with Android 8.0 and above, iOS 12.0 and above, and Windows 10 and above operating systems. The interface data format adopts Protobuf serialization.

2. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 1, characterized in that, The model training process of the AI ​​proofreading engine includes training dataset construction, pre-training stage, and fine-tuning stage; The training dataset was constructed by collecting 1 million articles from the publishing field, manually annotating them with 12 types of grammatical errors and 8 types of logical errors to form an annotated corpus, which was then divided into training, validation, and test sets in an 8:1:1 ratio. During the pre-training phase, a learning rate of 5e-5 and a batch size of 32 are used. The training set is iterated for 10 rounds using the Adam optimizer. After each round, the F1 score is calculated on the validation set. Pre-training is stopped when the F1 score improves by less than 0.01 for 3 consecutive rounds. During the fine-tuning phase, 50,000 domain-specific texts were introduced, including those related to science, literature, and education. The learning rate was adjusted to 1e-5, and the process was repeated 5 times. A gradient accumulation strategy was used to process small batches of data, and the final model achieved an accuracy of no less than 98.5% on the test set.

3. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 1, characterized in that, The collaborative interaction module includes a real-time differential synchronization unit and an annotation layer rendering engine; The real-time differential synchronization unit uses the OT algorithm to decompose the editing operation into three atomic operations: insertion, deletion, and replacement. Each operation is accompanied by a version number, which is in the format of YYYYMMDDHHMMSS plus the last 4 digits of the user ID. The annotation layer rendering engine uses Canvas technology to implement annotation overlay. Annotation data includes coordinate information, annotation type, and color value. The coordinate information is accurate to the pixel level. Annotation types include text, highlight, and doodle. The color value is in RGB format and is stored in association with the original data through a timestamp.

4. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 1, characterized in that, The distributed cloud storage module also includes a document parsing unit and a version management unit; The document parsing unit supports PDF, DOCX, and EPUB format conversion. During the conversion process, OCR technology is used to process scanned documents, with a recognition accuracy of no less than 99.2%. After text extraction, an intermediate file in XML format is generated. The version management unit adopts an incremental storage mechanism, which only saves the data that differs from the previous version. Each version is associated with a unique SHA-256 checksum, and the checksum is synchronized to the blockchain node in real time through the API interface.

5. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 2, characterized in that, The operation of the AI ​​proofreading engine includes text preprocessing, error detection, and generation of correction suggestions; Text preprocessing involves word segmentation, part-of-speech tagging, and sentence truncation of the input document. Word segmentation uses the Jieba word segmentation library and loads a custom dictionary for the publishing industry. Part-of-speech tagging uses the ICTCLAS tagging system. Sentence truncation limits a single sentence to a maximum of 512 characters. Error detection involves concatenating the sentence vector output by the BERT layer with the sequence features extracted by the BiLSTM layer, mapping it to a 20-dimensional error type space through a fully connected layer, and triggering a label when the probability of a certain type of error exceeds 0.

85. The correction suggestion generation involves calling a pre-trained T5-small model to generate three candidate correction texts for the marked error locations. The candidate texts are then compared with the original text using an edit distance algorithm, and the two with the smallest differences are selected as the output.

6. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 1, characterized in that, It also includes an access control module and a log auditing unit; The access control module adopts the RBAC model and sets up three levels of roles: editor, reviewer, and administrator. The editor role only has the right to comment (access code 001), the reviewer role has the right to approve (access code 010), and the administrator has the right to configure (access code 100). Access verification is implemented through AES-256 encrypted tokens. The log auditing unit generates an operation log every 5 minutes. The log content includes IP address, operation type, and data ID. The IP address is encrypted with MD5. The log file is automatically uploaded to the cloud storage cold backup node.

7. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 3, characterized in that, The blockchain node of the collaborative interaction module interacts with the WebSocket protocol processing unit of the module through the JSON-RPC interface, and its block structure includes a block header and a block body. The block header contains the previous block hash, Merkle root, and timestamp. The previous block hash and Merkle root are both 32 bytes, and the timestamp is in UTC format. The block contains 50-100 operation transactions. Each transaction contains the operator's public key, the hash of the operation content, and a digital signature. The digital signature uses the ECDSA algorithm. The consensus mechanism adopts Proof-of-Agent (PoA), with three or more trusted nodes taking turns generating blocks. The trusted nodes are the audit administrator accounts, and the block generation interval is 10 seconds.

8. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 5, characterized in that, The AI ​​proofreading engine also includes a professional lexicon unit and a sensitive word detection unit; The professional terminology unit stores 800,000 professional terms in the publishing industry, and constructs a prefix tree index according to subject classification. The terminology is updated through an incremental synchronization mechanism, and is reconciled with the central terminology server at 3:00 AM every day. The sensitive word detection unit uses the AC automaton algorithm, has a built-in database of 100,000 sensitive words, and takes no more than 100ms per thousand words to detect. When a word is detected, a red highlight is triggered and the submission operation is blocked.

9. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 4, characterized in that, The formula recognition module of the document parsing unit is the core sub-module of the unit. It uses the YOLOv5 algorithm to detect formula areas in the document and combines the Mathpix API to convert formulas to LaTeX. The recognition accuracy is no less than 95%. The table extraction module of the document parsing unit uses a combination of OpenCV edge detection and Tesseract-OCR to convert the table structure into HTML format while preserving attributes such as cell merging and border styles.

10. The intelligent collaborative editing, publishing, reviewing, and proofreading cloud system according to claim 1, characterized in that, The terminal adaptation interface also includes an offline caching unit and a breakpoint resume module; The offline caching unit supports local caching of the three most recently edited documents on the client side. The cached data is stored in an SQLite database, and incremental synchronization is achieved by comparing version numbers when the network is connected again. The breakpoint resume module uses the HTTP Range request header to implement large file chunked uploads, with each chunk being 5MB in size. The upload progress is fed back to the client in real time via WebSocket.

Images