AI multi-person writing system based on real-time cooperation

Abstract

The invention relates to the technical field of computer software, and discloses an AI (artificial intelligence) multi-person writing system based on real-time collaboration, which comprises a text input core module, an intention analysis core module, a version management core module and the like, a semantic anchoring module capable of being additionally arranged and a visual terminal. The system divides a collaboration area and an independent editing area through a text input module, and an intention analysis module triggers AI collaboration creation; the state tracking module collects text feature data and a real-time operation sequence, and the storage unit records version difference data and logic dependency weights; the version management module judges version conflicts and semantic association levels, and the branch generation module generates a dynamic branch strategy according to the version conflicts and the semantic association levels; the conflict judgment module is combined with logic dependence to construct a detection area to recognize conflicts, and the difference fusion module judges merging rationality based on semantic similarity; and the semantic anchoring module analyzes the semantic association strength and presents the semantic association strength through the visual terminal. According to the system, the real-time performance, the semantic coherence and the conflict resolution efficiency of multi-person collaborative writing are improved.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, specifically to an AI-based multi-person writing system based on real-time collaboration. Background Technology

[0002] In the digital age, collaborative writing scenarios are becoming increasingly common, such as team copywriting, collaborative academic paper writing, and multi-person novel writing. Traditional collaborative writing methods mainly rely on document sharing platforms or instant messaging tools, which have many problems that urgently need to be addressed.

[0003] From a real-time collaboration perspective, traditional systems struggle to achieve true real-time collaboration. Users' editing operations often require manual refreshes or waiting for periodic system synchronization, leading to information lag during collaboration and an inability to promptly perceive other participants' real-time modifications. This significantly impacts writing efficiency and the continuity of thought. For example, in urgent project document collaboration, the lack of real-time capability can result in delayed brainstorming among team members, hindering project progress.

[0004] In terms of version management, traditional systems have relatively simple version control mechanisms. When multiple people modify the same document simultaneously, version conflicts are prone to occur, and it is difficult to quickly and accurately determine the differences and semantic relationships between versions. Users need to spend a lot of time manually merging content from different versions, which not only increases workload but may also lead to content loss or logical confusion due to human error. Taking the collaborative writing of academic papers as an example, different chapters may have a large number of cross-references. Simple version control cannot effectively handle these complex relationships, making paper integration extremely difficult.

[0005] Ensuring semantic coherence is another challenge in traditional collaborative writing. When multiple people are writing, differences in writing style, thought process, and understanding of the topic among participants can easily lead to semantic incoherence, logical breaks, or redundant repetitions. Traditional systems lack effective semantic analysis and detection methods, making it impossible to identify these problems in real time and provide corresponding correction suggestions. They rely on manual reading of the entire text for adjustments, which is inefficient and prone to oversight. For example, in collaborative novel writing, inconsistencies in character personalities or abrupt plot developments may occur between different chapters; traditional systems cannot promptly detect and alert to these semantic coherence issues.

[0006] The conflict resolution mechanism is also inadequate. When multiple users edit the same area simultaneously, traditional systems typically only provide a conflict warning without intelligently assessing the severity and semantic relevance of the conflict, lacking effective conflict resolution strategies. Users must negotiate conflict resolution methods themselves, which is not only time-consuming and labor-intensive but can also negatively impact the collaborative atmosphere due to disagreements. In multi-person report writing within enterprises, different departments may interpret the same data differently, and traditional systems cannot provide scientific conflict assessment and integration solutions, leading to inconsistencies in the report content.

[0007] Furthermore, traditional collaborative writing systems have shortcomings in user operation tracking and visualization. They cannot accurately track users' editing focus and operation trajectory in real time, and it is difficult to intuitively display the semantic relationships of text structure and version evolution paths, which is detrimental to communication and collaboration among team members. For example, in the collaborative writing process of project proposals, team members cannot quickly understand the editing focus and thought process evolution of other members, affecting the level of synergy in collaboration.

[0008] With the development of artificial intelligence technology, applying AI to collaborative writing systems has become an important direction for solving the aforementioned problems. While some existing AI-assisted writing systems have improved writing efficiency to some extent, they still have shortcomings in real-time collaboration, version management, semantic analysis, and conflict resolution, failing to meet the needs of collaborative writing in complex scenarios. Therefore, there is an urgent need for an AI-based collaborative writing system based on real-time collaboration to address the problems of poor real-time performance, chaotic version management, difficulty in ensuring semantic coherence, and inefficient conflict resolution in traditional collaborative writing, thereby improving the efficiency and quality of collaborative writing. Summary of the Invention

[0009] The purpose of this invention is to provide an AI-based multi-person writing system based on real-time collaboration to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: an AI-based multi-person writing system based on real-time collaboration, the system comprising:

[0011] The module includes a text input module, an intent parsing module, a version management module, a storage unit, a branch generation module, a state tracking module, a difference fusion module, and a conflict determination module.

[0012] The text input module receives real-time text input commands from multiple users and divides the collaborative area into a collaborative area and an independent editing area; the intent parsing module controls the AI ​​model to initiate collaborative creation after receiving user commands; during the initial collaboration, the state tracking module collects text feature data input by each user and transmits it to the version management module; the storage unit stores version difference data of historical collaborative texts and transmits it to the version management module; the version management module determines the version conflict level and semantic association level of the current collaborative text, and generates a version evolution path and semantic association graph; the branch generation module generates a dynamic branching strategy for text merging based on the semantic association level and version conflict level.

[0013] When the AI ​​model performs collaborative creation based on the dynamic branching strategy, the state tracking module is also used to collect the real-time operation sequence of each user's editing behavior and transmit it to the conflict determination module; the storage unit is also used to store the logical dependencies between text paragraphs and the weight data corresponding to the dependencies and transmit it to the conflict determination module; the conflict determination module is used to determine whether the user's operation triggers a version conflict. If a conflict correction signal is generated, the real-time text paragraph corresponding to the conflict operation is transmitted to the difference fusion module; the difference fusion module is used to determine whether the text merging process deviates from the semantic consistency path based on the real-time operation sequence. If it deviates, a logical reconstruction instruction is sent to the intent parsing module.

[0014] Preferably, the system further includes a semantic anchoring module and a visualization terminal. The state tracking module is also used to collect real-time positioning data of the user's editing focus and transmit it to the semantic anchoring module. The semantic anchoring module is used to analyze the semantic association strength of the text structure in real time. If the analysis generates a logical break signal or a redundant repetition signal, the paragraph mark and the corresponding semantic association data are transmitted to the visualization terminal. The visualization terminal is used for the semantic association visualization processing of the text structure.

[0015] Preferably, the text feature data includes the edit density value of the user input content and the intent dispersion;

[0016] Version difference data includes the number of version switches for historical collaborative texts and the semantic transition amount corresponding to each switch.

[0017] Preferably, the determination process of the version management module is as follows:

[0018] Get the number of version switching times for each user's editing area, then get the semantic transition amount corresponding to each switch, and normalize the semantic transition amount corresponding to each switch to obtain the average semantic transition coefficient.

[0019] The version evolution intensity of each editing area is calculated based on the number of version switching and the average semantic transition coefficient.

[0020] The version evolution intensity is compared with the preset evolution threshold to determine whether the current collaborative text version conflict level is a level 1 conflict, level 2 conflict, or level 3 conflict.

[0021] Simultaneously acquire the editing density value and intent dispersion of each user's input, and calculate the semantic association coefficient of the current collaborative text;

[0022] The semantic association coefficient is compared with the preset association threshold to determine whether the semantic association level of the current collaborative text is first-level association, second-level association, or third-level association.

[0023] Preferably, the merging complexity corresponding to the third-level conflict is higher than that of the second-level conflict, and the merging complexity corresponding to the second-level conflict is higher than that of the first-level conflict.

[0024] The semantic coherence requirement for third-level associations is higher than that for second-level associations, and the semantic coherence requirement for second-level associations is higher than that for first-level associations.

[0025] Preferably, the branch generation module operates as follows:

[0026] The system prioritizes the semantic relevance level of the current text as the first priority factor, the version conflict level as the second priority factor, and the logical distance between the user's editing position and the main text stream as the third priority factor.

[0027] The system starts merging from the current main text stream. If there are multiple editing branches with the same semantic association level near the main text stream, the branch with the lower version conflict level is selected for merging first.

[0028] If the semantic association level and version conflict level of the edit branches are the same, then the branch that is logically closest to the main text stream is selected for merging;

[0029] When the system reaches the logical fork point:

[0030] If the branching point connects to only a single editing branch, the system directly merges the content of that branch;

[0031] If a branch point connects two or more editing branches, the branch with the higher semantic association level will be merged first; if multiple branches connected by a branch point have the same semantic association level, the branch with the lower version conflict level will be merged first; if both the semantic association level and the version conflict level of the branches connected by a branch point are the same, the branches will be randomly selected for merging.

[0032] If the branch point is not connected to other branches and there is unmerged content, the unmerged area with the closest logical distance will be processed first; when the system reaches any branch point, if all other areas have been merged, the initial version archive will be returned.

[0033] Generate a dynamic branching strategy for text merging according to the above rules;

[0034] When the system performs fusion according to the dynamic branching strategy, it adopts a differentiated semantic compensation mechanism in regions corresponding to different semantic association levels or level 3 conflicts.

[0035] Preferably, the weight data consists of the core anchor point, benchmark influence factor, and attenuation coefficient of the logical dependency relationship.

[0036] Preferably, the determination process of the conflict determination module is as follows:

[0037] Calculate the edit offset between the user's real-time action sequence and the logically dependent core anchor point;

[0038] Obtain the weight data corresponding to the logical dependencies between text paragraphs, extract the core anchor point, benchmark influence factor and attenuation coefficient, and obtain the correction radius of the logical dependency by multiplying the benchmark influence factor by the attenuation coefficient.

[0039] A conflict detection area for logical dependencies is constructed with the core anchor point as the center and the correction radius as the coverage area.

[0040] When the edit offset is greater than the correction radius, the user operation is outside the conflict detection area, and the current collaborative state is maintained.

[0041] When the edit offset is less than or equal to the correction radius, the user operation enters the conflict detection area, generating a conflict correction signal.

[0042] Preferably, the judgment process of the difference fusion module is as follows:

[0043] Set a standard semantic similarity threshold and maximum offset tolerance for text merging;

[0044] If the semantic vector alignment results are within the maximum offset tolerance and the contextual coherence is within the standard threshold range, the current merging strategy is maintained.

[0045] If the semantic vector comparison result exceeds the maximum offset tolerance or the contextual coherence deviates from the standard threshold range, a logical reconstruction instruction is generated.

[0046] Preferably, the parsing process of the semantic anchoring module is as follows:

[0047] Capture the hash value of the user's editing behavior;

[0048] Hash values ​​are mapped to version evolution graphs, and a sliding window mechanism is used to eliminate transient noise. A dynamic filtering threshold is set based on the correlation between operation frequency and version stability to extract effective editing feature sequences. A bidirectional LSTM network is applied to identify operation patterns and construct user intent evolution paths. Key operation nodes are weighted through an attention mechanism to match the semantic evolution features of the text structure and filter out invalid editing interference.

[0049] Synchronously acquire version difference hash values ​​of collaborative texts, set difference gradient range to identify abrupt change areas, use differential coding technology to enhance the display of version changes, combine timestamp sorting algorithm to extract version evolution trajectory, and identify potential logical conflicts;

[0050] If the operation trajectory hash value and the version difference hash value both show abnormal offsets, a logical break signal is generated and the corresponding segment is marked.

[0051] If neither of the two sets of data is abnormal, maintain the normal collaboration process;

[0052] If any of the data has an abnormal offset, a redundant repeat signal is generated and the corresponding segment is marked.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] In terms of real-time collaboration and intent parsing, the text input module can accurately divide the collaborative area into independent editing areas. Combined with the intent parsing module, it controls the AI ​​model to initiate collaborative creation, enabling the system to quickly capture and respond to the real-time text input commands of multiple users, achieving true real-time collaboration. For example, when multiple people are co-writing a news article, different reporters can simultaneously write different parts in their respective independent editing areas, while also sharing key information in real time in the collaborative area. The AI ​​model adjusts the writing direction in a timely manner based on user instructions, ensuring that team members' thinking is synchronized, greatly improving writing efficiency.

[0055] In version management and conflict level determination, the state tracking module collects text feature data input by users, the storage unit stores version difference data of historical collaborative texts, and the version management module accurately determines the version conflict level and semantic association level by calculating version evolution strength, semantic association coefficient, etc. This enables the system to clearly understand the differences and semantic relationships between versions, providing a scientific basis for subsequent branch generation and conflict resolution. Taking the collaborative writing of textbooks by multiple people as an example, the version switching and semantic leaps of different chapters can be accurately recorded and analyzed by the system, making it convenient for editors to quickly grasp the evolution process of the content and promptly discover potential version conflicts and semantic problems.

[0056] Regarding dynamic branching strategy generation, the branching generation module prioritizes semantic association level, version conflict level, and the logical distance between the user's editing position and the main text flow to generate dynamic branching strategies for text merging. This strategy can intelligently select the merging order based on the actual situation, employing differentiated semantic compensation mechanisms in areas corresponding to different semantic association levels or level three conflicts to ensure the rationality and semantic coherence of text merging. In multi-person novel writing, when multiple branches have different semantic associations and version conflicts, the system can prioritize merging branches with high semantic association and low version conflict with the main plot, avoiding plot confusion and making the story development smoother.

[0057] Regarding the conflict detection and difference fusion mechanism, the conflict detection module accurately determines whether a user's operation triggers a version conflict by calculating the editing offset and constructing a conflict detection region based on logical dependencies. The difference fusion module sets a standard semantic similarity threshold and a maximum offset tolerance to determine whether text merging deviates from the semantic consistency path. This mechanism can promptly detect and resolve conflicts during the collaboration process, ensuring the logical coherence of the text content. For example, in collaborative editing of academic papers, when different authors' modifications to the same argument conflict, the system can quickly determine the nature and severity of the conflict and perform reasonable logical reconstruction through the difference fusion module to ensure the academic rigor of the paper.

[0058] In terms of semantic anchoring and visualization, the semantic anchoring module collects real-time positioning data of the user's editing focus, analyzes the semantic correlation strength of the text structure, generates logical break signals or redundant repetition signals, and performs semantic correlation visualization processing through a visualization terminal. This allows team members to intuitively understand the semantic relationships of the text structure, promptly identify logical breaks or redundant repetitions, and facilitate targeted modifications and adjustments. In multi-person report writing within enterprises, the visualization terminal allows team members to clearly see the semantic relationships between different parts of the content, quickly locate areas that need optimization, and improve the quality and professionalism of the report.

[0059] In terms of operation sequence tracking and data processing, the state tracking module not only collects text feature data but also captures real-time operation sequences of user editing behavior. The semantic anchoring module constructs the user intent evolution path and identifies potential logical conflicts through a series of processes, such as capturing operation trajectory hash values ​​and mapping them to version evolution graphs. This precise tracking and analysis of user operations enables the system to gain a deep understanding of users' writing intentions and thought processes, laying the foundation for providing more intelligent collaborative support. In multi-person creative writing workshops, the system can analyze participants' writing styles and thought process changes based on their operation sequences, providing instructors with targeted guidance and suggestions to help participants improve their writing skills. Attached Figure Description

[0060] Figure 1 This is a schematic diagram illustrating the working principle of the AI ​​multi-person writing system based on real-time collaboration described in this invention.

[0061] Figure 2 Design diagram for the semantic anchoring and visualization processing module;

[0062] Figure 3 A design diagram for the version management module's decision-making process;

[0063] Figure 4 Design diagram of the conflict determination module's decision-making process;

[0064] Figure 5 Design diagram of the difference fusion module's judgment process;

[0065] Figure 6 A design diagram of the semantic anchoring module parsing process. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some 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.

[0067] Please see Figures 1-6 The present invention relates to an AI-based multi-person writing system based on real-time collaboration, the system comprising:

[0068] The system includes a text input module, an intent parsing module, a version management module, a storage unit, a branch generation module, a state tracking module, a difference fusion module, and a conflict determination module. These modules work together to enable real-time collaborative writing and conflict resolution among multiple users.

[0069] During system operation, the text input module first receives real-time text input commands from multiple users and divides the system into collaborative and independent editing areas based on preset rules. Text within the collaborative area can be edited by all users, while text within the independent editing area can only be modified by the corresponding user. The intent parsing module controls the AI ​​model to initiate the collaborative creation process upon receiving user commands; for example, users can trigger the AI ​​to generate relevant paragraphs or optimize existing content through commands.

[0070] During the initial collaboration phase, the state tracking module collects text feature data from each user's input, including edit density (the number of characters or paragraphs entered per unit time) and intent dispersion (the degree of difference in topic and style between different users' input content), and transmits this data to the version management module. The storage unit simultaneously stores version difference data for historical collaborative text, specifically the number of version switches and the semantic transition amount corresponding to each switch (the magnitude of semantic change in the text before and after the version switch), and transmits this data to the version management module.

[0071] The version management module determines the version conflict level and semantic association level of the current collaborative text. First, it obtains the number of version switches in each user's editing area and the semantic transition amount of each switch. The average semantic transition coefficient is calculated through normalization, and the version evolution strength is calculated based on the number of version switches. The version evolution strength is compared with a preset evolution threshold. If the evolution strength is higher than the level three threshold, it is determined to be a level three conflict, corresponding to the highest merging complexity; between the level two and level three thresholds, it is a level two conflict; and below the level one threshold, it is a level one conflict, with the lowest merging complexity. Simultaneously, the semantic association coefficient is calculated using each user's editing density value and intent dispersion. After comparing this with a preset association threshold, the semantic association level is divided into three levels (highest coherence requirement), two levels, and one level (lowest coherence requirement), generating a version evolution path and a semantic association graph.

[0072] The branch generation module generates a dynamic branching strategy for text merging based on semantic association level and version conflict level. The system prioritizes semantic association level as the first priority, version conflict level as the second priority, and logical distance between the user's editing position and the main text stream as the third priority. Merging begins in the main text stream. If multiple editing branches with the same semantic association level exist nearby, the branch with the lower version conflict level is prioritized for merging; if both the level and conflict level are the same, the branch with the closest logical distance is selected. Upon reaching a logical fork point, if only a single branch is connected, merging is performed directly; if multiple branches are connected, the branch with the higher semantic association level is prioritized for merging; if the levels are the same, the branch with the lower conflict level is prioritized; if all are the same, a random selection is made. If the fork point does not connect to other branches and contains unmerged content, the area with the closest logical distance is processed first, and the initial version is returned to the archive after merging. Differentiated semantic compensation mechanisms are used for different semantic association levels or level 3 conflict areas, such as adding AI semantic verification steps for high-conflict areas.

[0073] When the AI ​​model executes collaborative creation based on a dynamic branching strategy, the state tracking module collects the operation sequence of each user's editing behavior in real time (such as the time order and content of input, deletion, and modification operations), and transmits it to the conflict determination module. The storage unit stores the logical dependencies between text paragraphs and their corresponding weight data, including core anchors (key semantic nodes of a paragraph), baseline influence factors (the basic strength of logical dependencies), and attenuation coefficients (the degree to which dependencies weaken with increasing distance). The conflict determination module calculates the editing offset between the user's real-time operation sequence and the logical dependency core anchors, multiplies the baseline influence factor by the attenuation coefficient to obtain the correction radius, and constructs a conflict detection area centered on the core anchors. If the editing offset is greater than the correction radius, the user's operation is outside the area, maintaining the collaborative state; if it is less than or equal to the correction radius, the user enters the area, generates a conflict correction signal, and transmits the text paragraph corresponding to the conflicting operation to the difference fusion module.

[0074] The difference fusion module determines whether text merging deviates from the semantic consistency path based on the real-time operation sequence. It sets a standard semantic similarity threshold and a maximum deviation tolerance, compares semantic vectors, and checks contextual coherence. If the comparison result is within the tolerance and coherence is within the standard threshold range, the current merging strategy is maintained; if it exceeds the tolerance or coherence deviates from the range, a logical reconstruction instruction is generated and sent to the intent parsing module to trigger the AI ​​model to re-parse the user's intent and adjust the creation direction.

[0075] The present invention will be further described below with reference to Examples 1 to 5:

[0076] Example 1:

[0077] This embodiment relates to the system's extended modules and enhanced state tracking capabilities. The system also includes a semantic anchoring module and a visualization terminal, which work in conjunction with the state tracking module to achieve real-time parsing and visualization of semantic relationships within the text structure, assisting users in identifying logical breaks or redundant repetitions.

[0078] Building upon existing data collection of text feature data (editing density value, intent dispersion) and editing behavior sequences, the state tracking module further collects real-time positioning data of the user's editing focus. This data specifically includes the coordinates of the currently edited paragraph, the timestamp sequence of the cursor movement trajectory, and the start and end positions of the text selection area. For example, when a user continuously modifies the third paragraph of a document, the state tracking module records the paragraph's unique identifier, the start time of the user's operation, the cursor's movement path within the paragraph, and the specific location range of the modified content, transmitting this data in structured form to the semantic anchoring module in real time.

[0079] The core function of the semantic anchoring module is to analyze the semantic association strength of text structure in real time. Its processing flow consists of three stages: operation trajectory analysis, version difference analysis, and anomaly signal generation. In the operation trajectory analysis stage, the module captures the hash value of the user's editing behavior. This hash value is a unique identifier generated based on the time sequence of the user's operations, the operation type (such as insertion, deletion, and modification), and the operation content, used to characterize the user's editing behavior pattern. For example, if a user inserts keywords into the same paragraph three times consecutively, a specific hash value will be generated, reflecting their intention to strengthen the theme of that paragraph.

[0080] The semantic anchoring module maps operation trajectory hash values ​​to a version evolution graph. The version evolution graph is a visual model built upon version switching records of historical collaborative text. The horizontal axis represents time, and the vertical axis represents the logical hierarchy of text paragraphs. Each node represents a specific paragraph of a version, and the lines between nodes represent version iteration relationships. Through this mapping, real-time user operations can be associated with historical version changes, allowing analysis of the current operation's position and impact within the version evolution process.

[0081] To eliminate transient noise (such as quick undo operations after user mis-input), the module employs a sliding window mechanism to filter operation trajectory data. The sliding window's time range can be set to 5-10 seconds, retaining only stable operation sequences within the window. Simultaneously, a dynamic filtering threshold is set based on the correlation between operation frequency and version stability: when the user's operation frequency is less than 5 times per minute, it is considered a low-activity state, and the threshold is automatically increased to filter out more invalid operations; when the operation frequency is more than 10 times per minute, it is considered a high-activity state, and the threshold is decreased to retain more detailed data. Through this mechanism, effective editing feature sequences that truly reflect the user's intent are extracted.

[0082] The semantic anchoring module uses a bidirectional LSTM network to analyze effective editing feature sequences. The bidirectional LSTM network can capture the dependencies between sequences of data and, through weight updates of multiple layers of neurons, identify user operation patterns, such as continuous editing, paragraph reorganization, or topic switching. For example, if a user frequently performs copy-paste operations between multiple paragraphs, the network can recognize that they are reorganizing content, thereby constructing a path of user intent evolution and predicting their possible subsequent editing directions.

[0083] To highlight the impact of key operations on semantic evolution, the module uses an attention mechanism to weight key nodes in the operation sequence. Key nodes include modifications to core paragraphs (such as the abstract and conclusion), adjustments to logical connectors (such as the addition or deletion of "therefore" or "however"), and the addition or deletion of keywords. Through weighting, these key nodes have a higher weight in semantic analysis than ordinary operation nodes, thus more accurately matching the semantic evolution features of the text structure and filtering out invalid editing interference with little impact on the overall semantics, such as punctuation modifications or adjustments to individual words.

[0084] During the version difference analysis phase, the semantic anchoring module synchronously acquires the version difference hash value of the collaborative text. This hash value, generated by comparing the text content of the current version with the previous version, can characterize the overall degree of difference between versions. The module sets a difference gradient range (e.g., a numerical range of 0-100, with larger values ​​indicating greater differences). When the difference hash value exceeds a preset threshold (e.g., 40), it is determined that there is a sudden abrupt change in the text content, such as the sudden insertion of chapters unrelated to the topic or the deletion of key arguments.

[0085] To clearly present version changes, the module employs differential encoding technology to enhance the display of version differences. Differential encoding only records the differences between adjacent versions, rather than repeatedly storing the entire text. For example, it only marks the newly added paragraph "Section 5.2" and the deleted paragraph "Content at the end of Section 3.1". Combined with a timestamp sorting algorithm, the module arranges the differences in chronological order of operation, extracting the version evolution trajectory and intuitively displaying the semantic change paths between versions. This facilitates the identification of potential logical conflicts, such as the deletion of the "technical advantages" paragraph in a certain version, followed by subsequent versions failing to adjust the related conclusions, leading to semantic contradictions.

[0086] In the anomaly signal generation stage, the semantic anchoring module jointly analyzes the operation trajectory hash value and the version difference hash value. If both show abnormal offsets (e.g., the operation trajectory shows that the user frequently modifies multiple unrelated paragraphs, and the version difference hash value increases significantly), it is determined that there is a logical break in the text structure, generating a logical break signal and marking the corresponding paragraph. For example, if a user frequently modifies data in the "Experimental Methods" section but does not simultaneously adjust the "Experimental Conclusions" section, resulting in logical inconsistencies between the two parts, the module will mark these two paragraphs.

[0087] If both the operation trajectory hash value and the version difference hash value are normal, the normal collaboration process continues without triggering any warnings. If either of these data shows an abnormal shift (e.g., the operation trajectory shows the user repeatedly modifying the same paragraph but the version difference hash value does not change significantly), it is determined that there is a redundancy / duplication problem, a redundancy / duplication signal is generated, and the corresponding paragraph is marked. For example, if a user repeatedly inserts similar content into the same paragraph, resulting in lengthy text, but the semantic differences between versions do not change significantly, the module will mark that paragraph.

[0088] The paragraph markers and corresponding semantic association data (such as the specific type of logical break and the degree of redundancy) generated by the semantic anchoring module are transmitted to the visualization terminal in real time. The visualization terminal visualizes the semantic associations of the text structure, specifically as follows: different colored lines are used to mark the strength of semantic associations between paragraphs. For example, green lines represent strong associations (such as causal relationships), yellow lines represent medium associations (such as parallel relationships), and red lines represent weak associations (such as coincidental associations). For paragraphs with logical breaks or redundancy, a warning panel is generated on the right side of the document editing interface, displaying the specific anomaly type and paragraph location, and marked with a red triangle icon or a yellow exclamation mark icon.

[0089] Users can visually view the results during the editing process. For example, when a paragraph is marked with a red triangle, they can click to view detailed information and understand the logical breakpoints between this paragraph and other paragraphs. For paragraphs marked with yellow exclamation marks, redundant content can be quickly located and removed. Through this visualization method, users can more efficiently discover semantic problems in the text structure, adjust editing strategies in a timely manner, and improve the content quality and logical coherence of collaborative writing.

[0090] Example 2:

[0091] This embodiment details the specific definitions and collection methods of text feature data and version difference data. As the basic data for system operation, these two data provide key inputs for the conflict level determination and semantic association level determination of the version management module.

[0092] Text feature data includes edit density and intent dispersion, used to characterize the activity level of user editing behavior and the consistency of creative intent. Edit density is calculated by statistically analyzing the amount of text input and modification frequency per unit of time. Specifically, the system monitors user keyboard input events, text selection operations, and paragraph adjustment behaviors in real time: when a user inputs text in a collaborative or independent editing area, the system counts the number of characters or paragraphs; for deletion and modification operations, the system records the text length affected by each operation and the number of operations. For example, if user A inputs 800 characters, deletes 200 characters, and adjusts the format of 3 paragraphs within 15 minutes, their edit density value can be derived by combining the number of characters input, deleted characters, and the number of operations, reflecting a high level of editing activity during that time period.

[0093] Intent dispersion analyzes the differences in theme, sentiment, and style among different user input content using natural language processing (NLP) technology. The system first segments the text input by each user, extracting thematic keywords (such as technical terms and core concept vocabulary), and then uses a sentiment analysis model to determine the text's sentiment orientation (e.g., positive, neutral, negative). Simultaneously, it analyzes stylistic features such as sentence structure and word choice preferences, determining whether the text belongs to formal written language, colloquial expression, or professional technical style. Then, it calculates the keyword overlap rate, sentiment orientation difference, and style feature distance between different user content: the lower the keyword overlap rate, the greater the sentiment orientation difference, and the greater the style feature distance, the higher the intent dispersion, indicating a more significant difference in creative intent between users. For example, User B's input content is themed "technical principles," using rigorous professional terminology with a neutral sentiment orientation; User C's input content revolves around "application scenarios," using colloquial language with a positive promotional tendency. The keyword overlap rate between the two is only 20%, the sentiment orientation difference is 0.7, and the style feature distance is 0.8. The overall determination shows a high intent dispersion, indicating a significant divergence in their creative directions.

[0094] Version difference data represents the number of version switches for historical collaborative text and the semantic transitions corresponding to each switch, used to record the evolution process and semantic changes of collaborative text. The number of version switches refers to the cumulative number of times collaborative text has been updated from one version to another. Each version switch is triggered by a user's active update submission or by the system's automatic saving. For example, in a multi-user collaboration process, each time a user completes a round of modifications and clicks the "Save" button, the system generates a new version, and the version switch count increments by 1. The storage unit records the timestamp, triggering user, and version number of each version switch in real time, forming a version evolution timeline.

[0095] The semantic leap measure calculates the degree of semantic difference between adjacent versions using a semantic analysis model. The system employs a pre-trained language model (such as BERT) to vectorize the text of both the old and new versions, converting the text into semantic vectors in a high-dimensional space. Then, the cosine similarity between the two semantic vectors is calculated. The closer the similarity value is to 1, the smaller the semantic difference between versions; the closer it is to 0, the greater the difference. The semantic leap measure is defined as 1 minus the cosine similarity value. For example, if the cosine similarity between the semantic vectors of version V2 and version V1 is 0.7, then the semantic leap measure is 0.3, indicating that the semantic change of the new version relative to the old version is 30%. For texts containing multiple paragraphs, the system can further calculate the semantic leap measure of each paragraph and then calculate a weighted average to obtain the overall semantic leap measure. The weights can be set according to the logical importance of the paragraphs (such as abstract, body text, and conclusion). For example, the weight of the abstract paragraph is 0.3, the weight of the body text paragraph is 0.5, and the weight of the conclusion paragraph is 0.2.

[0096] During data acquisition, the state tracking module and storage unit work closely together. The state tracking module monitors user editing behavior in real time. When a version switching event is detected (such as a user submitting an update), it triggers the semantic transition calculation process: First, it extracts the text content of the current version (new version) and the previous version (old version), and calls the semantic analysis model to generate semantic vectors for both versions; then, it calculates the cosine similarity and semantic transition, and transmits the results to the storage unit. The storage unit stores version difference data in a structured table format. The table fields include version number, switching time, triggering user, number of version switches, semantic transition, and text content hash value (used for quick comparison of version content). For example, a record in the table shows: Version V3, switching time May 24, 2025, 14:30, triggering user Zhang San, number of version switches 3, semantic transition 0.25, hash value abc123, indicating that this version is the third switch, and the semantic change compared to V2 is 25%.

[0097] The collection frequency of edit density and intent dispersion is synchronized with user editing behavior; that is, the system updates relevant data in real time whenever a user inputs or modifies text. Version difference data collection is triggered with each version switch, ensuring a complete record of the version evolution trajectory. This data provides a basis for quantitative analysis in the version management module: edit density reflects user participation in collaboration, with high-density editing areas potentially indicating more real-time modification conflicts; intent dispersion suggests consistency in creative direction among users, with high dispersion potentially leading to a decrease in semantic relevance; the number of version switches and the amount of semantic transitions measure the evolutionary activity and stability of the text, with high numbers or high transition amounts potentially indicating an increased level of version conflict.

[0098] By accurately defining and collecting text feature data and version difference data in real time, the system can dynamically perceive editing behavior patterns and text semantic changes during multi-person collaborative writing. This provides reliable data support for subsequent version conflict determination, semantic association analysis, and dynamic branching strategy generation, thereby achieving more efficient collaborative conflict management and ensuring semantic coherence. For example, when the editing density value of a certain editing area suddenly increases and the intention dispersion increases, the system can issue an early warning that semantic divergence may occur in that area, triggering more frequent conflict detection. When the number of version switching increases and the semantic jump exceeds a preset threshold, the system can determine that the text has entered a high-conflict evolution stage, automatically adjust the branching strategy, and prioritize branches with high semantic relevance to maintain the stability of the text structure.

[0099] Example 3:

[0100] This embodiment further elaborates on the determination process of the version management module. This module provides the core basis for the system to generate dynamic branching strategies through quantitative analysis of version conflict levels and semantic association levels. The following explanation covers three stages: data acquisition, conflict level determination, and semantic association level determination.

[0101] I. Data Acquisition and Preprocessing

[0102] The version management module first obtains two types of basic data from the status tracking module and the storage unit:

[0103] Version switching related data: including the number of version switches for each user's editing area (e.g., user A's version switching count in the collaboration area is...). User B switched versions the following number of times in the independent editing area: ), and the semantic transition amount corresponding to each switch (such as the first ), The semantic transition amount of each switch is The range of values ​​is The larger the value, the greater the semantic difference between the previous and subsequent versions.

[0104] Editing behavior characteristic data: including the editing density value of each user's input content (e.g., the editing density value of user C is...). (unit: characters per minute) and intent dispersion (e.g., the intent dispersion between user D and user E is...). The range of values ​​is The larger the value, the greater the difference in creative intent.

[0105] After acquiring the data, the module needs to normalize the semantic transition values ​​to eliminate the influence of data of different magnitudes. The normalization formula is:

[0106]

[0107] in, These are the normalized semantic transition coefficients. and These are the minimum and maximum values ​​of all semantic transitions within the corresponding edit region, respectively. For example, if the set of semantic transitions for a certain edit region is... ,but , The normalization coefficient for the third switch is This indicates that this switch is the version update with the most significant semantic changes in this region.

[0108] II. Version Conflict Level Determination

[0109] Version conflict level determination is based on the "version evolution strength" indicator, which comprehensively considers version switching frequency and semantic change magnitude. The calculation formula is as follows:

[0110]

[0111] in, This represents the version evolution strength; a higher value indicates a higher risk of version conflicts. This represents the total number of version switches in the editing area. This is the sum of all normalized semantic transition coefficients within the region; and Preset weighting coefficients ( This setting is used to adjust the relative importance of handover frequency and semantic changes. For example, if the system is more concerned with the impact of handover frequency on conflicts, it can be set to... , The module will calculate the version evolution strength. With the three-level preset evolution threshold ( Compare and determine the conflict level:

[0112] Level 1 conflict: when This indicates a low version switching frequency and minimal semantic changes, resulting in the lowest merging complexity. For example, if an editing area only switches versions twice within one hour, and each semantic jump is less than 0.2, the normalized cumulative value is 0.3. ,but It was determined to be a Level 1 conflict.

[0113] Second-level conflict: when At this time, when the version switching frequency or semantic change magnitude is moderate, merging requires medium-complexity processing. For example, if another region switches 5 times, the cumulative semantic transition value is 1.8. ,like , Therefore, But not exceeding It was determined to be a Level 2 conflict.

[0114] Level 3 conflict: When At this stage, frequent version switching and significant semantic changes result in the highest merging complexity. For example, if a certain region switches 10 times, the cumulative semantic transition value is 4.5. The conflict has been classified as a Level 3 conflict, requiring priority handling and a more stringent merging strategy.

[0115] III. Semantic Relevance Level Determination

[0116] The semantic association level is determined based on the "semantic association coefficient" indicator, which comprehensively reflects the synergy of user editing behavior and the consistency of creative intent. The calculation formula is as follows:

[0117]

[0118] in, This is the semantic relevance coefficient; a larger value indicates a higher requirement for semantic coherence. The average edit density value for all users, reflecting collaborative activity (unit: characters / minute); The average value of the intent dispersion among all users (range) ), Then it means Figure 1 To the point of being compatible; and Preset weighting coefficients ( ), used to regulate activity level and intention Figure 1 The relative importance of consistency. For example, if the system pays more attention to consistency... Figure 1 The impact of consistency on semantic coherence can be set. , The module will use semantic association coefficients. Associated threshold with Level 3 ( Compare and determine the association level:

[0119] First-level association: when At this time, user editing activity is low and intent dispersion is high, resulting in the lowest demand for semantic coherence. For example, if the average editing density is 200 characters / minute and the average intent dispersion is 0.6, ,like Therefore, It is determined to be a first-level association, allowing for a certain degree of semantic looseness in the text.

[0120] Secondary association: when At that time, activity level or intention Figure 1 Moderate consistency is required, ensuring a moderate level of semantic coherence. For example, increasing the average editing density to 300 characters / minute reduces the average intent dispersion to 0.3. ,like If the association is not found, it is considered a secondary association, requiring the AI ​​model to perform basic semantic verification during the merging process.

[0121] Third-level association: When At this time, user activity is high and intent is highly consistent, with the highest demand for semantic coherence. For example, the average editing density reaches 400 characters / minute, and the average intent dispersion is only 0.1. If it is determined to be a level 3 association, the AI ​​model needs to perform deep semantic fusion and forcibly check the logical connection of the context.

[0122] IV. Application and Feedback of Judgment Results

[0123] After determining the conflict level and association level, the version management module generates two core outputs:

[0124] Version evolution path: Based on the timeline, the conflict level and semantic leap of each version are marked to form a visual version iteration map. For example, red nodes in the map represent level 3 conflict versions, blue nodes represent level 1 conflict versions, and the size of the nodes reflects the magnitude of the semantic leap.

[0125] Semantic Relationship Map: Using text paragraphs as nodes, lines are used to represent the semantic relationship strength between paragraphs (e.g., thick solid lines for third-level relationships and dashed lines for first-level relationships), and the core anchor points of logical dependencies (e.g., key argument paragraphs) are marked.

[0126] The above results are transmitted to the branch generation module in real time as the basis for generating dynamic branching strategies. For example, if the semantic association level of the current text is level three and the version conflict level is level two, the branch generation module will prioritize merging branches with high semantic association and adopt a medium-complexity fusion strategy for level two conflict areas (such as adding an AI automatic semantic verification step). At the same time, the judgment results will also be fed back to the state tracking module to adjust its data collection frequency (such as increasing the operation sequence collection density for high conflict areas) and conflict detection accuracy (such as reducing the radius of the conflict detection area for logical dependencies).

[0127] V. Key Details

[0128] Threshold setting mechanism: Preset evolution threshold and associated threshold It can be dynamically adjusted according to the text type. For example, in collaborative academic paper scenarios, where high semantic coherence is required, it can be... Set it to a higher value (e.g., 150); however, in creative writing scenarios, it can be appropriately reduced. To allow for more divergent content.

[0129] Real-time data assurance: The version management module's decision-making cycle is synchronized with the version switching frequency, meaning a recalculation is triggered after each version switch to ensure the system responds promptly to changes in collaboration status. For example, when multiple people are editing simultaneously, leading to frequent version switching, the module can complete multiple decisions per second, avoiding delays in conflict handling.

[0130] Abnormal data handling: If a user's edit density value or intent dispersion data is missing (e.g., due to network latency causing data not to be transmitted), the module will use a moving average method to fill the missing value, for example, by replacing it with the average of the user's previous three valid data, to ensure the continuity of the judgment process.

[0131] Example 4:

[0132] This embodiment details the workflow and strategy generation rules of the branch generation module. This module dynamically generates text merging branch strategies, prioritizing semantic relevance level as the first priority, version conflict level as the second priority, and the logical distance between the user's editing position and the main text flow as the third priority. This ensures efficient fusion of collaborative content while maintaining semantic coherence. The following explanation is illustrated with specific collaborative scenario examples.

[0133] I. Priority Logic of Strategy Generation and Initial Merging Process

[0134] Assuming the main text flow of the current collaborative text is structured as "Introduction → Technical Principles → Application Scenarios," users A, B, and C have generated branch content in different editing areas. The system first prioritizes semantic relevance level to determine the semantic match between each branch and the main text flow. For example:

[0135] User A's branch: Add an "Algorithm Improvement" paragraph after the "Technical Principles" section, with a semantic relevance level of three (strong relevance, keywords highly overlapping with "Technical Principles").

[0136] User B's branch: Insert a "Market Competitor Analysis" paragraph after the "Application Scenarios" section, with a semantic relevance level of two (medium relevance, involving extensions of application scenarios but not direct dependencies).

[0137] User C's branch: Add a "Research Background" paragraph before "Introduction", with a semantic relevance level of three (strong relevance, belonging to the preceding logical content of the introduction).

[0138] The system begins merging from the "Introduction" section at the start of the main text stream. At this point, the nearby edit branches in the main text stream include user C's "Research Background" (located before the Introduction, with a logical distance of 1) and user A's "Algorithm Improvement" (located after the Technical Principles, with a logical distance of 2). Since both have a semantic association level of three (same priority), the system proceeds to the second priority check, comparing the version conflict levels:

[0139] User C's branch version switching count is 1, and the semantic jump is 0.1 (low conflict).

[0140] User A's branch version switching count is 3, with a semantic jump of 0.4 (medium conflict). Therefore, the system prioritizes merging User C's branch, inserting "Research Background" before the introduction to form the main text flow of "Research Background → Introduction → Technical Principles → Application Scenarios".

[0141] II. Branching rules at logical fork points

[0142] When the main text flow reaches the "Technical Principles" section, a logical branching point occurs: User A's "Algorithm Improvement" branch (third-level association, second-level conflict) and User D's newly submitted "Technical Limitations" branch (second-level association, first-level conflict) both connect to the end of this section. At this point, the system processes the data according to the following rules:

[0143] Priority semantic association level: User A's branch is a level 3 association, which is higher than User D's level 2 association. Therefore, the "algorithm improvement" paragraph is merged first.

[0144] If the levels are the same, the conflict level is compared: Suppose there is another branch E (three-level association, three-level conflict) that is parallel to user A's branch. Because the conflict level is higher than user A's two-level conflict, the system still prioritizes user A's branch.

[0145] If the levels and conflicts are the same, the logical distance will be used for selection: If branch F (third-level association, second-level conflict) has the same semantic level and conflict level as user A's branch, but is located after the "Application Scenarios" chapter (logical distance is 3), while user A's branch is located after "Technical Principles" (logical distance is 1), then user A's branch will be merged first.

[0146] After merging User A's branch, the main text stream is updated to "Research Background → Introduction → Technical Principles → Algorithm Improvement → Application Scenarios". At this point, User D's "Technical Limitations" branch, although having a low semantic relevance, still needs to be processed. The system determines its logical distance from the main text stream (the distance from "Technical Principles" to "Technical Limitations" is 1, due to the insertion of "Algorithm Improvement" in between). Since this branch is not covered by other higher-priority branches, after processing the current fork point, the system returns to merge this branch and inserts it after "Algorithm Improvement", forming the structure "Technical Principles → Algorithm Improvement → Technical Limitations → Application Scenarios".

[0147] III. Handling of Single Branches and Unintegrated Content

[0148] At the end of the "Application Scenarios" section, the system encounters a branch that connects only a single editing branch: User B's "Market Competitor Analysis" (secondary association, primary conflict). Since there are no other competing branches, the system directly merges this branch, expanding the main text flow to "Application Scenarios → Market Competitor Analysis".

[0149] If there is an independent editable area that is not connected to any branching point (such as the "Project Overview" paragraph added by user E before "Research Background", with a semantic relevance level of three and a conflict level of one), the system determines its logical distance from the main text flow (distance from "Research Background" is 1), prioritizes processing this unmerged area, inserts "Project Overview" before "Research Background", forming "Project Overview → Research Background → Introduction → Technical Principles → Algorithm Improvement → Technical Limitations → Application Scenarios → Market Competitor Analysis".

[0150] IV. Random Selection Mechanism and Version Archive

[0151] When the system reaches a branching point, if there are multiple branches with the same semantic association level, version conflict level, and equal logical distance (e.g., branches G and H are both second-level associations, first-level conflicts, and both connect to the "Technical Limitations" section, with a logical distance of 1), the system will randomly select one of the branches for merging to avoid process stagnation due to indecisiveness. For example, branch G might be randomly selected, and after merging, branch H might continue processing, or the system might return to processing after merging in other areas.

[0152] Once all branch points have been processed and there is no unmerged content, the system returns to the initial version archive, recording the final merge result and branch strategy path. For example, the archive will record the branch merging order of "User C → User A → User D → User B → User E" and the specific position of each branch in the main text stream, facilitating subsequent version backtracking or conflict tracing.

[0153] V. Application of Differentiated Semantic Compensation Mechanism

[0154] In regions corresponding to different semantic association levels or level 3 conflicts, the system employs a differentiated semantic compensation mechanism:

[0155] Level 3 related areas: If there is a strong logical dependency between the "Technical Principles → Algorithm Improvement" paragraphs, the system triggers the AI ​​model to perform deep semantic verification to ensure that the content of "Algorithm Improvement" is completely consistent with the formula derivations and theoretical foundations in "Technical Principles". For example, if "Technical Principles" mentions "the complexity of the traditional algorithm is O(n²)", the AI ​​model will check whether "Algorithm Improvement" mentions "the complexity is reduced to O(nlogn) after optimization" and automatically supplement the necessary derivation process.

[0156] Level 3 Conflict Area: Suppose an editing area is classified as a Level 3 conflict due to frequent version switching (10 times) and a large semantic jump (average 0.6). During the merging process, the system automatically inserts a manual review node, requiring at least one user to confirm the merged result of the conflicting paragraphs. For example, if user A and user F have conflicting statements about the "technical limitations" paragraph (one emphasizes "high hardware costs," while the other believes "costs are controllable"), the system will display both statements side by side, prompting the user to choose the final version or triggering AI to generate a compromise solution.

[0157] Low-association areas: such as “market competitor analysis” and “technical principles” are secondary associations. The system allows the AI ​​model to perform appropriate semantic expansion during fusion, such as automatically adding transitional sentences like “Compared with competitors, the advantages of this technology in algorithm efficiency are reflected in XX aspects” to enhance the logical connection between paragraphs.

[0158] VI. Exception Handling in Collaborative Scenarios

[0159] If a branch connected to a certain branch point is suddenly withdrawn by the user (e.g., user D deletes the "Technical Limitations" branch), the system will detect the branch disappearance in real time, skip the processing of that branch point, and continue to advance the main text flow. If the withdrawal operation causes a logical break in the main text flow (e.g., the withdrawn content is a key argument), the state tracking module will capture the abnormal operation trajectory hash value, trigger the semantic anchoring module to re-parse the association strength, and generate a logical break signal if necessary, prompting the user to fix it.

[0160] In addition, when the system detects a dynamic change in the semantic relevance level of a branch (such as when user B modifies the content of "market competitor analysis" to increase the overlap rate between its keywords and "application scenarios" from 40% to 70%, and upgrades the semantic relevance level from level two to level three), the branch generation module will re-evaluate the priority of the branch. If the current main text stream has not yet processed to the fork point, the fusion order will be adjusted to prioritize the processing of the upgraded high-relevance branch.

[0161] VII. Summary of Core Principles of Strategy Generation

[0162] The strategy generation process of the branch generation module always follows these principles:

[0163] Semantic priority: Ensure that the theme of the merged text does not deviate from the core logic, and always prioritize highly related branches over lowly related branches to avoid content defocusing due to blind merging.

[0164] Conflict downgrading: When semantic association levels are the same, low-conflict branches are processed first to reduce the cost of manual intervention during the merging process and improve collaboration efficiency.

[0165] Logical proximity: By assessing the closeness between branches and the main text flow through logical distance, proximity fusion can reduce the sense of semantic jump and enhance reading coherence.

[0166] Dynamic adaptation: Responds to changes in branch status in real time (such as adding, modifying, or deleting branches), recalculates priorities, and adjusts strategies to ensure the system remains flexible in real-time multi-user collaboration scenarios.

[0167] Example 5:

[0168] This embodiment details the specific working process of the conflict determination module and the difference fusion module. These two modules, through logical dependency analysis and semantic consistency verification, achieve real-time detection of user operation conflicts and dynamic adjustment of merging strategies, ensuring the logical coherence of collaborative writing among multiple users. The following explanation is based on a specific collaborative scenario.

[0169] I. Logical Dependency Analysis of the Conflict Detection Module

[0170] The core of the conflict detection module is to construct conflict detection regions based on the logical dependencies between text paragraphs. Its workflow is as follows: Assume a user is collaboratively editing a document on "Applications of Artificial Intelligence Algorithms," where paragraph P2 in the "Algorithm Principles" section (core argument: "Gradient descent solves the optimal solution by iteratively updating parameters") is the core anchor point of the logical dependency. The storage unit records the weight data of this paragraph: a baseline influence factor of 0.8 (indicating strong logical dependency) and a decay coefficient of 0.6 (indicating that the logical dependency weakens as the paragraph distance increases).

[0171] When user A enters "Newton's method does not require setting a learning rate and converges faster" in paragraph P5 (the third paragraph downstream of P2), the conflict determination module performs the following operations:

[0172] Calculate the editing offset: Determine the paragraph distance between the user's operation position and the core anchor point P2 as 3 (P3, P4, P5), and calculate the offset based on the attenuation coefficient: .

[0173] Constructing the collision detection region: The corrected radius is obtained by multiplying the baseline impact factor by the attenuation coefficient. Centered on P2, the coverage area has a corrected radius of 0.48, meaning the collision detection region includes P2 and paragraphs with a logical distance ≤ 0.48 from it. Since paragraph distances are integers, the actual detection range can be mapped to the paragraphs adjacent to P2 (such as P1, P2, and P3).

[0174] Conflict determination: User A's operation position is P5, and the edit offset of 1.8 is greater than the correction radius of 0.48. It is determined that the conflict is outside the conflict detection area. The current collaborative state will be maintained and no intervention is required.

[0175] If user B deletes the keyword "iterative update" in paragraph P3 (1 unit away from P2), the edit offset is... If the distance is less than the correction radius of 0.48 (assuming that the system uses rounding up or continuous value mapping for distance calculation, in actual scenarios the paragraph distance may be converted into a continuous logical distance value through an algorithm), it is determined to have entered the conflict detection area, a conflict correction signal is generated, and the real-time text paragraph of P3 is transmitted to the difference fusion module.

[0176] II. Semantic Consistency Judgment in the Difference Fusion Module

[0177] After receiving conflicting paragraph data, the difference fusion module performs semantic consistency verification. Taking user B deleting the keyword "iterative update" as an example, the module executes the following steps:

[0178] Set the verification threshold: the standard semantic similarity threshold is 0.8 (meaning that the paragraph and the context must be at least 80% consistent), and the maximum offset tolerance is 0.15 (allowing the semantic similarity to drop by 15% from the threshold, i.e., the minimum tolerance value is 0.65).

[0179] Semantic vector comparison: Generate semantic vectors of paragraph P3 before modification (containing "iterative update") and after modification (deleting the keyword) using a natural language processing model, and calculate the cosine similarity between the two and the context (P2 and P4).

[0180] Before modification: P3 has a similarity of 0.9 with P2 and a similarity of 0.85 with P4;

[0181] After modification: After removing "iterative update" from P3, the similarity with P2 decreased to 0.6 (because P2's core argument relies on this keyword), and the similarity with P4 decreased to 0.7.

[0182] Coherence Assessment: Contextual coherence is scored through logical connector detection (e.g., whether "therefore" in P4 corresponds to the content in P3) and grammatical consistency analysis (e.g., whether the subjects are consistent), with a maximum score of 100. Before the revision, the coherence score was 85 points. After the revision, due to logical breaks (P2 mentions "iterative updates," and after P3 is deleted, "therefore" in P4 loses its logical basis), the score dropped to 55 points.

[0183] Strategy decision: If the modified semantic similarity (0.6) is lower than the minimum value allowed by the maximum offset tolerance (0.65) and the coherence score (55 points) deviates from the standard threshold range (≥70 points), it is determined to be a deviation from the semantic consistency path, and a logical reconstruction instruction is generated and sent to the intent parsing module.

[0184] III. Reconstruction Response of the Intent Parsing Module

[0185] After receiving the logic reconstruction instruction, the intent parsing module triggers the AI ​​model to re-analyze the user intent and text semantics:

[0186] User intent identification: Based on user B's operation trajectory (continuously deleting keywords and quickly saving), it is determined that his / her possible intent is to simplify the expression, but he / she is unaware of the importance of the keyword to the P2-P4 logical chain.

[0187] Semantic Reconstruction Scheme Generation: The AI ​​model generates two reconstruction schemes:

[0188] Option 1: Restore the keyword "iterative update" and add supplementary explanation in P3: "Although the learning rate needs to be manually adjusted, gradient descent is still applicable to most scenarios."

[0189] Option 2: Modify the content of P2 and P4 so that the logic chain does not depend on the keyword "iterative update" (e.g., change P2 to "gradient descent method solves the optimal solution by adjusting parameters" and P4 to "Newton's method has a greater advantage in parameter adjustment efficiency").

[0190] Solution Selection and Execution: Based on the semantic association level of the current text (assuming a level 3 association, requiring strong logical coherence), the system prioritizes Solution 1, automatically recovers keywords and inserts supplementary explanations to ensure the integrity of the logical chain from P2 to P4. After reconstruction, the similarity between P3 and P2 recovers to 0.88, and the similarity between P3 and P4 recovers to 0.82. The coherence score recovers to 80 points, meeting the semantic consistency requirement.

[0191] IV. Examples of Handling Complex Conflict Scenarios

[0192] Suppose user C inserts "Genetic algorithms are suitable for combinatorial optimization problems" in paragraph P8 of the "Algorithm Applications" section, while paragraph P6 (the core anchor point, baseline impact factor 0.7, decay coefficient 0.5) states "Neural networks optimize weights through backpropagation," and the logical distance between the two is 2. Calculate the edit offset: Corrected radius: Since offset 1 > 0.35, the initial determination is that there is no conflict.

[0193] However, records P6 and P8 in the storage unit have an indirect logical dependency (both belong to the category of "algorithm type", and the core anchor point in the weight data contains a common parent node "algorithm classification"). The conflict determination module detects this hidden dependency through the semantic association graph, recalculates the logical distance to 1 (related through the parent node), and the offset becomes... If the value is less than the correction radius of 0.35, a conflict correction signal is generated. The difference fusion module found that the content of P8 has no direct semantic relationship with the "Neural Network" branch to which P6 belongs, with a semantic similarity of only 0.5, triggering logical reconstruction. It is recommended to move P8 to a separate "Evolutionary Algorithm" chapter and add a transition sentence at the end of the chapter where P6 is located: "In addition to neural networks, other algorithms such as genetic algorithms also play an important role in the field of combinatorial optimization."

[0194] V. Key Mechanisms of Module Collaboration

[0195] Real-time data closed loop: The status tracking module continuously collects operation sequences and editing focus data, the conflict determination module updates the conflict detection area every 500 milliseconds, and the difference fusion module completes semantic verification within 1 second after receiving a conflict signal to ensure the real-time nature of collaborative response.

[0196] Logical dependency dynamic update: When a user modifies the paragraph content, causing changes in logical dependencies (such as adding conjunctions like "therefore" or "for example"), the storage unit updates the weight data in real time, and the conflict determination module recalculates the correction radius and offset to avoid misjudgments caused by the static dependency model.

[0197] Parallel processing of multiple conflicts: If multiple conflicting segments exist simultaneously (e.g., user B modifies P3 and user D modifies P7), the system sorts them by conflict level (level 3 conflict takes priority) and logical distance (nearest segment takes priority) and processes them one by one to avoid resource contention.

[0198] VI. Fault Tolerance Handling for Abnormal Operations

[0199] If the conflict determination module experiences missing weight data due to data transmission delays, the system automatically uses default parameters (benchmark influence factor 0.5, attenuation coefficient 0.5) for temporary calculations and marks it as "estimated data," to be re-determined after complete data synchronization. If the difference fusion module is unable to generate reconstruction instructions in a timely manner due to excessive AI model load, the system prompts the user "The current operation may cause logical conflicts; do you want to save the changes and continue editing?" to avoid interruption of collaboration.

[0200] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0201] 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. An AI-powered multi-person writing system based on real-time collaboration, characterized in that: It includes a text input module, an intent parsing module, a version management module, a storage unit, a branch generation module, a state tracking module, a difference fusion module, and a conflict determination module; The text input module receives real-time text input commands from multiple users and divides the collaborative area into a collaborative area and an independent editing area; the intent parsing module controls the AI ​​model to initiate collaborative creation after receiving user commands; during the initial collaboration, the state tracking module collects text feature data input by each user and transmits it to the version management module; the storage unit stores version difference data of historical collaborative texts and transmits it to the version management module; the version management module determines the version conflict level and semantic association level of the current collaborative text, and generates a version evolution path and semantic association graph; the branch generation module generates a dynamic branching strategy for text merging based on the semantic association level and version conflict level. When the AI ​​model performs collaborative creation based on the dynamic branching strategy, the state tracking module is also used to collect the real-time operation sequence of each user's editing behavior and transmit it to the conflict determination module; the storage unit is also used to store the logical dependencies between text paragraphs and the weight data corresponding to the dependencies and transmit it to the conflict determination module. The conflict determination module is used to determine whether a user operation triggers a version conflict. If a conflict correction signal is generated, the real-time text paragraph corresponding to the conflict operation is transmitted to the difference fusion module. The difference fusion module is used to determine whether the text merging process deviates from the semantic consistency path based on the real-time operation sequence. If it does, it sends a logical reconstruction instruction to the intent parsing module.

2. The AI ​​multi-person writing system based on real-time collaboration according to claim 1, characterized in that, It also includes a semantic anchoring module and a visualization terminal. The state tracking module is also used to collect real-time positioning data of the user's editing focus and transmit it to the semantic anchoring module. The semantic anchoring module is used to analyze the semantic association strength of the text structure in real time. If the analysis generates a logical break signal or redundant repetition signal, the paragraph mark and the corresponding semantic association data are transmitted to the visualization terminal. The visualization terminal is used for the semantic association visualization processing of the text structure.

3. The AI ​​multi-person writing system based on real-time collaboration according to claim 1, characterized in that, Text feature data includes the edit density value of user input content and the dispersion of intent; Version difference data includes the number of version switches for historical collaborative texts and the semantic transition amount corresponding to each switch.

4. The AI ​​multi-person writing system based on real-time collaboration according to claim 3, characterized in that, The determination process of the version management module is as follows: Get the number of version switching times for each user's editing area, then get the semantic transition amount corresponding to each switch, and normalize the semantic transition amount corresponding to each switch to obtain the average semantic transition coefficient. The version evolution intensity of each editing area is calculated based on the number of version switching and the average semantic transition coefficient. The version evolution intensity is compared with the preset evolution threshold to determine whether the current collaborative text version conflict level is a level 1 conflict, level 2 conflict, or level 3 conflict. Simultaneously acquire the editing density value and intent dispersion of each user's input, and calculate the semantic association coefficient of the current collaborative text; The semantic association coefficient is compared with the preset association threshold to determine whether the semantic association level of the current collaborative text is first-level association, second-level association, or third-level association.

5. The AI ​​multi-person writing system based on real-time collaboration according to claim 4, characterized in that, The merging complexity for a level 3 conflict is higher than that for a level 2 conflict, and the merging complexity for a level 2 conflict is higher than that for a level 1 conflict. The semantic coherence requirement for third-level associations is higher than that for second-level associations, and the semantic coherence requirement for second-level associations is higher than that for first-level associations.

6. The AI ​​multi-person writing system based on real-time collaboration according to claim 5, characterized in that, The branch generation module works as follows: The system prioritizes the semantic relevance level of the current text as the first priority factor, the version conflict level as the second priority factor, and the logical distance between the user's editing position and the main text stream as the third priority factor. The system starts merging from the current main text stream. If there are multiple editing branches with the same semantic association level near the main text stream, the branch with the lower version conflict level is selected for merging first. If the semantic association level and version conflict level of the edit branches are the same, then the branch that is logically closest to the main text stream is selected for merging; When the system reaches the logical fork point: If the branching point connects to only a single editing branch, the system directly merges the content of that branch; If a branch point connects two or more editing branches, the branch with the higher semantic association level will be merged first; if multiple branches connected by a branch point have the same semantic association level, the branch with the lower version conflict level will be merged first; if both the semantic association level and the version conflict level of the branches connected by a branch point are the same, the branches will be randomly selected for merging. If the branch point is not connected to other branches and there is unmerged content, the unmerged area with the closest logical distance will be processed first; when the system reaches any branch point, if all other areas have been merged, the initial version archive will be returned. Generate a dynamic branching strategy for text merging according to the above rules; When the system performs fusion according to the dynamic branching strategy, it adopts a differentiated semantic compensation mechanism in regions corresponding to different semantic association levels or level 3 conflicts.

7. The AI ​​multi-person writing system based on real-time collaboration according to claim 1, characterized in that, The weighted data serves as the core anchor point, benchmark influence factor, and attenuation coefficient for logical dependencies.

8. The AI ​​multi-person writing system based on real-time collaboration according to claim 7, characterized in that, The specific determination process of the conflict determination module is as follows: Calculate the edit offset between the user's real-time action sequence and the logically dependent core anchor point; Obtain the weight data corresponding to the logical dependencies between text paragraphs, extract the core anchor point, benchmark influence factor and attenuation coefficient, and obtain the correction radius of the logical dependency by multiplying the benchmark influence factor by the attenuation coefficient. A conflict detection area for logical dependencies is constructed with the core anchor point as the center and the correction radius as the coverage area. When the edit offset is greater than the correction radius, the user operation is outside the conflict detection area, and the current collaborative state is maintained. When the edit offset is less than or equal to the correction radius, the user operation enters the conflict detection area, generating a conflict correction signal.

9. The AI ​​multi-person writing system based on real-time collaboration according to claim 8, characterized in that, The specific judgment process of the difference fusion module is as follows: Set a standard semantic similarity threshold and maximum offset tolerance for text merging; If the semantic vector alignment results are within the maximum offset tolerance and the contextual coherence is within the standard threshold range, the current merging strategy is maintained. If the semantic vector comparison result exceeds the maximum offset tolerance or the contextual coherence deviates from the standard threshold range, a logical reconstruction instruction is generated.

10. The AI ​​multi-person writing system based on real-time collaboration according to claim 1, characterized in that, The parsing process of the semantic anchoring module is as follows: Capture the hash value of the user's editing behavior; Hash values ​​are mapped to version evolution graphs, a sliding window mechanism is used to eliminate transient noise, and a dynamic filtering threshold is set based on the correlation between operation frequency and version stability to extract effective editing feature sequences. A bidirectional LSTM network is applied to identify operation patterns and construct the user intent evolution path; key operation nodes are weighted through an attention mechanism to match the semantic evolution features of the text structure and filter out invalid editing interference. Synchronously acquire version difference hash values ​​of collaborative texts, set difference gradient range to identify abrupt change areas, use differential coding technology to enhance the display of version changes, combine timestamp sorting algorithm to extract version evolution trajectory, and identify potential logical conflicts; If the operation trajectory hash value and the version difference hash value both show abnormal offsets, a logical break signal is generated and the corresponding segment is marked. If neither of the two sets of data is abnormal, maintain the normal collaboration process; If any of the data has an abnormal offset, a redundant repeat signal is generated and the corresponding segment is marked.

Images