Manuscript online review auxiliary method based on natural language processing

Abstract

The invention discloses a manuscript online review auxiliary method based on natural language processing. The manuscript online review auxiliary method comprises the steps of obtaining a manuscript text, executing segmentation, sentence segmentation and dependency analysis, recognizing an entity and a logical relationship, and constructing a semantic graph; a verification rule is extracted and analyzed into a feature set, a rule potential energy field is constructed, and constraint is applied to semantic nodes; inputting the semantic graph and the potential energy field into a constraint network, executing semantic evolutionary updating, and generating an evolutionary semantic graph; inputting the evolutionary semantic nodes into an editing module to generate a continuous reversible editing field and a node editing vector; executing optimal path solution based on the editing field, and generating a manuscript revision scheme meeting rule constraints; and displaying the revised scheme, and updating rule parameters and semantic weights according to user feedback to realize optimization. According to the method, by constructing an evolvable semantic constraint graph network and a reversible editing field generation mechanism, collaborative unification of manuscript semantic understanding, rule constraint and intelligent revision is achieved.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent text review technology, and in particular to an online manuscript review assistance method based on natural language processing. Background Technology

[0002] Currently, manuscript review primarily relies on human proofreading or automated proofreading tools based on simple rules. While human proofreading can identify linguistic, semantic, and logical issues, it is time-consuming and inconsistent. Existing automated proofreading systems mostly depend on fixed grammatical rules, keyword matching, or traditional machine learning models, making it difficult to understand the contextual relationships, semantic logic, and stylistic differences between different types of manuscripts. These systems typically only detect spelling or grammatical errors, lacking the comprehensive analytical capabilities to address semantic consistency, logical coherence, and content compliance, making accurate proofreading difficult for high-requirement manuscripts such as scientific papers, policy documents, or press releases.

[0003] In existing technologies, review rules are mostly static templates, unable to dynamically participate in the model's reasoning and generation process, resulting in a disconnect between "rule detection" and "revision suggestion generation." The model often identifies errors first, and then a separate module proposes revision suggestions, causing problems such as a disconnect in the review process, delayed feedback, and inconsistent revisions. Because review rules are not embedded into the model in a structured manner, differences between different manuscript types cannot be effectively identified and adapted, and there is a lack of intelligent adjustment mechanisms tailored to domain and stylistic features.

[0004] A more prominent problem is that existing system modifications are mostly one-way replacements, lacking traceability and rollback capability. Users cannot clearly understand the basis for model modifications, the scope of their impact, or the rationale behind them. This "black box" approach to modification makes it difficult for users to conduct secondary reviews or partial restorations, severely limiting the system's credibility and controllability in high-standard review scenarios.

[0005] Therefore, how to provide an online manuscript review assistance method based on natural language processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an online manuscript review assistance method based on natural language processing. This invention fully utilizes evolvable semantic constraint graph networks, rule potential field modeling, and reversible editing field generation technologies to construct an intelligent review system integrating semantic understanding, rule constraints, and editing optimization. By mapping the semantic structure of the manuscript to review rules as a differentiable rule potential field, semantic nodes are driven to adaptively evolve under rule constraints. Based on this, a reversible editing field is generated, achieving manuscript revision with semantic consistency, rule compliance, and minimal intervention. It possesses advantages such as high semantic understanding depth, strong rule compliance, good traceability of modifications, and strong adaptive learning capabilities, significantly improving the accuracy, controllability, and intelligence level of manuscript review.

[0007] An online manuscript review assistance method based on natural language processing according to an embodiment of the present invention includes: Obtain the text of the manuscript to be reviewed and proofread, perform segmentation, sentence segmentation, lexical analysis and syntactic dependency analysis on the manuscript, identify entity and logical relationship, and construct semantic graph; The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines, and the review rules are parsed into rule feature sets. Based on the rule type and applicable conditions, a rule potential field is constructed, and semantic constraints are applied to the semantic graph nodes. The semantic graph and the rule potential field are input into the evolvable semantic constraint graph network. The semantic constraint evolution operation is performed on the semantic nodes. The node representation is dynamically updated according to the semantic propagation relationship of the nodes and the rule potential gradient, so as to obtain the set of semantic nodes after semantic constraint evolution and the corresponding evolved semantic graph. The semantic node set after semantic constraint evolution is input into the reversible edit field generation module to generate a continuous reversible edit field. The continuous reversible edit field contains multiple edit vectors, and each edit vector describes the potential edit type, operation direction and confidence of the corresponding node. Based on the optimal path solution of the continuous reversible editing field, the optimal editing sequence that satisfies the condition of minimizing the rule potential energy is determined, forming a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints, and outputting the modification location, modification content and trigger rule identifier; The revised manuscript is displayed on the online review interface. Based on the user's acceptance, rejection, or adjustment feedback, the parameters and semantic constraint weights of the rule potential field are updated to achieve adaptive optimization and continuous learning for different manuscript types and review standards.

[0008] Optionally, constructing a semantic graph refers to extracting semantic units and establishing a set of nodes based on the dependency relationships, referential relationships, and logical connections between sentences after the manuscript text is segmented into sentences, words, and syntactic dependency analysis. Then, directed edges are established on the semantically related nodes according to the syntactic master-slave relationship, contextual succession relationship, and semantic dependency relationship, forming a multi-layer semantic graph structure containing node attributes and edge weights.

[0009] Optionally, the step of constructing a rule potential field based on rule type and applicable conditions to impose semantic constraints on semantic graph nodes includes: The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines. Each rule is structured and parsed to determine its scope of application, triggering conditions, exception conditions, execution level, priority and version identifier, forming a rule feature set and storing it in the form of entries. Based on the constructed semantic graph, the rule feature set is mapped one-to-one with the semantic graph nodes according to the hierarchy of chapter, title, paragraph, sentence and phrase, and the rule application list, exception list and application boundary of each node are determined, and the binding of rules to nodes and the disabling mark are completed. A hierarchical description of the rule potential field is generated at three levels: node level, edge level, and region level. At the node level, the execution strength of mandatory constraints and suggested constraints is determined based on textual and structural attributes. At the edge level, a scope is established for reference consistency, logical consistency, and reference chain consistency. At the region level, a templated constraint domain is established for the same chapter or layout template. The three layers of description together constitute the rule potential expression of the semantic graph. Establish a rule conflict decomposition and dominance selection table, perform conflict detection and adjudication on multiple rules in the same node or the same scope, determine the final effective rule set in the order of mandatory over suggested, narrow domain over wide domain, and latest version over old version, and generate a traceable exception inheritance chain and conflict relationship diagram. The final set of rules that have been adjudicated and the hierarchical rule potential energy description are combined to form the rule potential energy field output, which imposes deterministic semantic constraints on each node and related edge in the semantic graph.

[0010] Optionally, obtaining the semantic node set after semantic constraint evolution and the corresponding evolved semantic graph includes: The semantic graph and rule potential field are simultaneously input into the evolvable semantic constraint graph network. A three-layer structure consisting of a rule potential alignment layer, a semantic propagation and constraint fusion layer, and an editing feasibility projection layer is initialized, and the final effective rule set, exception list, and scope boundary are loaded. In the rule potential alignment layer, each semantic node in the semantic graph is bound to a corresponding effective rule entry. Based on the node attributes, structural position and rule scope, node-level and edge-level propagation allow and prohibition markers are generated. An exception inheritance chain is established and a rule subgraph divided by chapter and template is generated. At the same time, a cross-graph reference alignment channel is established to constrain cross-paragraph associations involving charts, formulas and references. In the semantic propagation and constraint fusion layer, constrained information propagation is performed on allowed edges based on propagation tags and rule subgraphs. Convergence calculation is performed on nodes within the same scope, and the convergence result is sent back to the node through a region consistency loop, maintaining chapter-level consistency and template-level consistency, while maintaining the constraint of cross-graph reference alignment channels and eliminating the offset of reference number and reference chain. In the conflict adjudication subprocess of the semantic propagation and constraint fusion layer, multiple rules within the same node or the same local subgraph are adjudicated. The unique effective rule is determined in the order of priority of mandatory rules, priority of narrow scope, and priority of latest version. Nodes with consistent update results in two consecutive updates are marked as steady-state nodes, and nodes with differences are marked as nodes to be evolved. In the editing feasibility projection layer, the current state of the node marked as to be evolved is projected to generate candidate editing anchors that correspond one-to-one with the node. The editing type, operation direction, scope and triggering rule identifier of each anchor are given, and a traceable index from the anchor to the original semantic node and scope is established. Output the set of semantic nodes and the evolved semantic graph after semantic constraint evolution, along with the candidate edit anchor index, propagation allow flag, prohibit flag, exception inheritance chain and cross-graph reference alignment constraint.

[0011] Optionally, the step of inputting the semantically constrained evolved set of semantic nodes into the reversible edit field generation module to generate a continuous reversible edit field includes: The structure of the initial reversible edit field generation module includes a rule-driven direction derivation unit, an edit operation basis mapping unit, and a vector-script bidirectional quantization and rollback verification unit. The rule-driven direction derivation unit generates an initial edit vector for each semantic node and its corresponding candidate edit anchor point, determines the operation direction and intensity of the initial edit vector, the operation direction is determined based on the rule potential field and the current state of the node, and the intensity is determined based on the rule enforcement degree, scope of application and node confidence, and binds hard constraint references, suggested constraint references and safety threshold sets to each initial edit vector, and sets cross-scope consistency guardrails based on cross-graph reference alignment constraints; The rule-driven direction derivation unit completes the one-to-one alignment of the initial edit vector with the candidate edit anchor points, performs unique processing on the duplicate vectors generated by adjacent anchor points, forms a set of anchor point-level edit vectors, and marks each edit vector with the trigger rule identifier, scope level and exception inheritance chain location information; The anchor-level edit vector is projected onto a predefined set of edit operation bases by the edit operation base mapping unit. The set of edit operation bases includes insertion, deletion, replacement, rearrangement, and format repair. Weights of each operation base are assigned to each edit vector, and the sum of the weights is limited to no more than one. An operation allowed list and an operation prohibited list are generated based on the propagation allowed and prohibited flags. The scope and precise boundary of each edit vector are determined by a multi-granularity scope edit grid. The vector-script bidirectional quantization and rollback verification unit converts the edit vector into discrete edit script records. The edit script records include operation type, scope identifier, trigger rule identifier, operation confidence, state summary before change and state summary after change. A corresponding reverse mapping is generated to support the recovery of the edit vector from the edit script. A rollback token is generated for each edit script record and a sequentially verifiable state summary chain is constructed to complete the reversibility verification and traceability verification. Output a continuous and reversible edit field data structure, including an anchor-level edit vector set, a corresponding edit script set, an operation allowed list, an operation prohibited list, cross-scope consistency guardrails, and a multi-granularity scope edit grid.

[0012] Optionally, the process of forming a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints includes: Receive the set of anchor-level edit vectors, the corresponding set of edit scripts, the list of allowed operations, the list of prohibited operations, the cross-scope consistency guardrail, and the multi-granularity scope edit grid. Merge and number all candidate edit items according to the semantic node order and the trigger rule identifier. Based on the list of allowed operations and the list of prohibited operations, the feasibility of candidate edit items is screened, edit items that violate hard constraints are eliminated, and edit items that meet the cross-scope consistency guardrail and multi-granularity scope editing grid boundary conditions are retained. Sequence dependencies and mutual exclusion conflicts are generated. A comprehensive cost index is calculated for feasible editable items. The comprehensive cost index consists of three parts: the editing range, the remaining amount of suggested constraints, and the cross-scope cost. The individual cost is synthesized according to the preset weights, and constraint satisfaction labels with priority given to hard constraint satisfaction are generated at the same time. Under the condition of satisfying all hard constraints and sequential dependencies, the optimal path solving program is executed. The editing path is selected according to the criteria of minimizing the comprehensive cost and maximizing the satisfaction of the suggested constraints, so as to obtain the optimal editing path and the corresponding minimum intervention editing sequence, and maintain a one-to-one mapping relationship with the original editing script record. The manuscript revision scheme is generated based on the minimum intervention editing sequence. The scheme outputs the modification location, modification type, trigger rule identifier, operation confidence, rollback token and status summary chain fragment in the execution order. It also outputs a list of candidate editing items that were not adopted and their corresponding constraint descriptions.

[0013] Optionally, updating the parameters and semantic constraint weights of the rule potential field based on the user's acceptance, rejection, or adjustment of feedback information includes: The system receives manuscript revision plans, optimal editing paths, minimal intervention editing sequences, rollback tokens, and status summary chains. The online review interface displays the modification location, modification type, trigger rule identifier, and operation confidence level of each editing item in an itemized format. Collect the user's operation results for each edit item. The operation results include three states: adoption, rejection, and adjustment. For edit items in the adjustment state, record the user's added or deleted edit content and the corresponding change information of the scope of effect, and establish a one-to-one mapping relationship with the original editing script. Generate a set of feedback records, and mark each edit item with its final status, trigger rule identifier, scope level, rollback token, and status summary chain fragment. Mark rejected or adjusted edit items as deviation samples and adopted edit items as positive samples, and form a grouped statistical index by rule item, by chapter, and by template. Based on the feedback record set, incremental updates are performed on the parameters of the rule potential field, the execution strength, priority and exception inheritance chain of the effective rules are adjusted, hard constraints are not reduced, suggested constraints are added or removed according to feedback, the rule version identifier and timestamp are updated, and a new rule potential field is output. Based on the feedback record set, the semantic constraint weights are incrementally updated, the semantic propagation weights, conflict adjudication order and cross-graph reference alignment weights are adjusted, nodes that remain consistent for two consecutive rounds are marked as steady-state nodes, the suppression weights are added to edit modes that are rejected multiple times, and new propagation allow and prohibition tags are generated. Generate updated manuscript content and review records, and output a complete audit result including the final revision list, the list of unused edits, the rule version change record, and the status summary chain.

[0014] The beneficial effects of this invention are: This invention achieves a deep integration of semantic understanding and rule constraints by constructing an evolvable semantic constraint graph network, overcoming the core problem of "grammatically correct but semantically non-compliant" in existing proofreading techniques. Through the construction of a semantic graph and a multi-layered semantic propagation mechanism, the system enables the model to perform global reasoning at the levels of contextual semantics, logical structure, and language style, thereby accurately identifying deep-seated problems such as semantic ambiguity, logical contradictions, and inconsistencies in expression. Compared with traditional systems based on rule matching or template detection, this invention has stronger semantic understanding capabilities, automatically identifying multi-dimensional errors in complex texts and significantly improving the depth and accuracy of proofreading.

[0015] This invention introduces a rule potential field modeling method, embedding language specifications, formatting standards, and professional proofreading rules into the model in a differentiable form. This transforms rule constraints from independent detection modules into an intrinsic driving force for semantic reasoning and editing generation. The rule potential field continuously influences the evolution direction of semantic nodes during model operation, achieving synchronous constraints on grammatical norms, logical consistency, and typesetting, effectively solving the problem of "separation between rule detection and modification generation" in existing technologies. This innovative mechanism enables the system to maintain linguistic naturalness and semantic coherence while adhering to rules, improving the overall consistency and standardization of proofreading.

[0016] The reversible edit field generation mechanism proposed in this invention enables editing operations to be interpretable, traceable, and rollback-able. Instead of directly outputting the modified text, the system generates a continuous and reversible edit field. The direction and intensity of the operation vectors represent the modification type and confidence level, achieving intelligent revision with minimal intervention. Users can view, select, or undo modifications in the online review interface and dynamically update model parameters based on feedback results, thereby achieving adaptive optimization and continuous learning. This not only enhances the transparency and controllability of the review process but also improves review efficiency and user trust, ensuring the system has scalability and high reliability across multiple domains and text styles. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of an online manuscript review assistance method based on natural language processing proposed in this invention; Figure 2 This is a schematic diagram of the three-layer structure of the evolvable semantic constraint graph network of the online manuscript review assistance method based on natural language processing proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 An online manuscript review assistance method based on natural language processing includes: Obtain the text of the manuscript to be reviewed and proofread, perform segmentation, sentence segmentation, lexical analysis and syntactic dependency analysis on the manuscript, identify entity and logical relationship, and construct semantic graph; The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines, and the review rules are parsed into rule feature sets. Based on the rule type and applicable conditions, a rule potential field is constructed, and semantic constraints are applied to the semantic graph nodes. The semantic graph and the rule potential field are input into the evolvable semantic constraint graph network. The semantic constraint evolution operation is performed on the semantic nodes. The node representation is dynamically updated according to the semantic propagation relationship of the nodes and the rule potential gradient, so as to obtain the set of semantic nodes after semantic constraint evolution and the corresponding evolved semantic graph. The semantic node set after semantic constraint evolution is input into the reversible edit field generation module to generate a continuous reversible edit field. The continuous reversible edit field contains multiple edit vectors, and each edit vector describes the potential edit type, operation direction and confidence of the corresponding node. Based on the optimal path solution of the continuous reversible editing field, the optimal editing sequence that satisfies the condition of minimizing the rule potential energy is determined, forming a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints, and outputting the modification location, modification content and trigger rule identifier; The revised manuscript is displayed on the online review interface. Based on the user's acceptance, rejection, or adjustment feedback, the parameters and semantic constraint weights of the rule potential field are updated to achieve adaptive optimization and continuous learning for different manuscript types and review standards.

[0021] In this embodiment, constructing a semantic graph refers to extracting semantic units and establishing a set of nodes based on the dependency relationships, referential relationships, and logical connections between sentences after the manuscript text is segmented into sentences, words, and syntactic dependency analysis. Directed edges are then established between semantically related nodes according to syntactic master-slave relationships, contextual succession relationships, and semantic dependency relationships, forming a multi-layer semantic graph structure that includes node attributes and edge weights.

[0022] In this embodiment, the step of constructing a rule potential field based on rule type and applicable conditions, and applying semantic constraints to semantic graph nodes, includes: The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines. Each rule is structured and parsed to determine its scope of application, triggering conditions, exception conditions, execution level, priority and version identifier, forming a rule feature set and storing it in the form of entries. Based on the constructed semantic graph, the rule feature set is mapped one-to-one with the semantic graph nodes according to the hierarchy of chapter, title, paragraph, sentence and phrase, and the rule application list, exception list and application boundary of each node are determined, and the binding of rules to nodes and the disabling mark are completed. A hierarchical description of the rule potential field is generated at three levels: node level, edge level, and region level. At the node level, the execution strength of mandatory constraints and suggested constraints is determined based on textual and structural attributes. At the edge level, a scope is established for reference consistency, logical consistency, and reference chain consistency. At the region level, a templated constraint domain is established for the same chapter or layout template. The three layers of description together constitute the rule potential expression of the semantic graph. Establish a rule conflict decomposition and dominance selection table, perform conflict detection and adjudication on multiple rules in the same node or the same scope, determine the final effective rule set in the order of mandatory over suggested, narrow domain over wide domain, and latest version over old version, and generate a traceable exception inheritance chain and conflict relationship diagram. The final set of rules that have been adjudicated and the hierarchical rule potential energy description are combined to form the rule potential energy field output, which imposes deterministic semantic constraints on each node and related edge in the semantic graph.

[0023] In this embodiment, obtaining the semantic node set after semantic constraint evolution and the corresponding evolved semantic graph includes: The semantic graph and rule potential field are simultaneously input into the evolvable semantic constraint graph network. A three-layer structure consisting of a rule potential alignment layer, a semantic propagation and constraint fusion layer, and an editing feasibility projection layer is initialized, and the final effective rule set, exception list, and scope boundary are loaded. In the rule potential alignment layer, each semantic node in the semantic graph is bound to a corresponding effective rule entry. Based on the node attributes, structural position and rule scope, node-level and edge-level propagation allow and prohibition markers are generated. An exception inheritance chain is established and a rule subgraph divided by chapter and template is generated. At the same time, a cross-graph reference alignment channel is established to constrain cross-paragraph associations involving charts, formulas and references. In the semantic propagation and constraint fusion layer, constrained information propagation is performed on allowed edges based on propagation tags and rule subgraphs. Convergence calculation is performed on nodes within the same scope, and the convergence result is sent back to the node through a region consistency loop, maintaining chapter-level consistency and template-level consistency, while maintaining the constraint of cross-graph reference alignment channels and eliminating the offset of reference number and reference chain. In the conflict adjudication subprocess of the semantic propagation and constraint fusion layer, multiple rules within the same node or the same local subgraph are adjudicated. The unique effective rule is determined in the order of priority of mandatory rules, priority of narrow scope, and priority of latest version. Nodes with consistent update results in two consecutive updates are marked as steady-state nodes, and nodes with differences are marked as nodes to be evolved. In the editing feasibility projection layer, the current state of the node marked as to be evolved is projected to generate candidate editing anchors that correspond one-to-one with the node. The editing type, operation direction, scope and triggering rule identifier of each anchor are given, and a traceable index from the anchor to the original semantic node and scope is established. Output the set of semantic nodes and the evolved semantic graph after semantic constraint evolution, along with the candidate edit anchor index, propagation allow flag, prohibit flag, exception inheritance chain and cross-graph reference alignment constraint.

[0024] In this embodiment, the step of inputting the semantically constrained evolved set of semantic nodes into the reversible edit field generation module to generate a continuous reversible edit field includes: The structure of the initial reversible edit field generation module includes a rule-driven direction derivation unit, an edit operation basis mapping unit, and a vector-script bidirectional quantization and rollback verification unit. The rule-driven direction derivation unit generates an initial edit vector for each semantic node and its corresponding candidate edit anchor point, determines the operation direction and intensity of the initial edit vector, the operation direction is determined based on the rule potential field and the current state of the node, and the intensity is determined based on the rule enforcement degree, scope of application and node confidence, and binds hard constraint references, suggested constraint references and safety threshold sets to each initial edit vector, and sets cross-scope consistency guardrails based on cross-graph reference alignment constraints; The rule-driven direction derivation unit completes the one-to-one alignment of the initial edit vector with the candidate edit anchor points, performs unique processing on the duplicate vectors generated by adjacent anchor points, forms a set of anchor point-level edit vectors, and marks each edit vector with the trigger rule identifier, scope level and exception inheritance chain location information; The anchor-level edit vector is projected onto a predefined set of edit operation bases by the edit operation base mapping unit. The set of edit operation bases includes insertion, deletion, replacement, rearrangement, and format repair. Weights of each operation base are assigned to each edit vector, and the sum of the weights is limited to no more than one. An operation allowed list and an operation prohibited list are generated based on the propagation allowed and prohibited flags. The scope and precise boundary of each edit vector are determined by a multi-granularity scope edit grid. The vector-script bidirectional quantization and rollback verification unit converts the edit vector into discrete edit script records. The edit script records include operation type, scope identifier, trigger rule identifier, operation confidence, state summary before change and state summary after change. A corresponding reverse mapping is generated to support the recovery of the edit vector from the edit script. A rollback token is generated for each edit script record and a sequentially verifiable state summary chain is constructed to complete the reversibility verification and traceability verification. Output a continuous and reversible edit field data structure, including an anchor-level edit vector set, a corresponding edit script set, an operation allowed list, an operation prohibited list, cross-scope consistency guardrails, and a multi-granularity scope edit grid.

[0025] In this embodiment, the process of forming a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints includes: Receive the set of anchor-level edit vectors, the corresponding set of edit scripts, the list of allowed operations, the list of prohibited operations, the cross-scope consistency guardrail, and the multi-granularity scope edit grid. Merge and number all candidate edit items according to the semantic node order and the trigger rule identifier. Based on the list of allowed operations and the list of prohibited operations, the feasibility of candidate edit items is screened, edit items that violate hard constraints are eliminated, and edit items that meet the cross-scope consistency guardrail and multi-granularity scope editing grid boundary conditions are retained. Sequence dependencies and mutual exclusion conflicts are generated. A comprehensive cost index is calculated for feasible editable items. The comprehensive cost index consists of three parts: the editing range, the remaining amount of suggested constraints, and the cross-scope cost. The individual cost is synthesized according to the preset weights, and constraint satisfaction labels with priority given to hard constraint satisfaction are generated at the same time. Under the condition of satisfying all hard constraints and sequential dependencies, the optimal path solving program is executed. The editing path is selected according to the criteria of minimizing the comprehensive cost and maximizing the satisfaction of the suggested constraints, so as to obtain the optimal editing path and the corresponding minimum intervention editing sequence, and maintain a one-to-one mapping relationship with the original editing script record. The manuscript revision scheme is generated based on the minimum intervention editing sequence. The scheme outputs the modification location, modification type, trigger rule identifier, operation confidence, rollback token and status summary chain fragment in the execution order. It also outputs a list of candidate editing items that were not adopted and their corresponding constraint descriptions.

[0026] In this embodiment, updating the parameters and semantic constraint weights of the rule potential field based on the user's acceptance, rejection, or adjustment of feedback information includes: The system receives manuscript revision plans, optimal editing paths, minimal intervention editing sequences, rollback tokens, and status summary chains. The online review interface displays the modification location, modification type, trigger rule identifier, and operation confidence level of each editing item in an itemized format. Collect the user's operation results for each edit item. The operation results include three states: adoption, rejection, and adjustment. For edit items in the adjustment state, record the user's added or deleted edit content and the corresponding change information of the scope of effect, and establish a one-to-one mapping relationship with the original editing script. Generate a set of feedback records, and mark each edit item with its final status, trigger rule identifier, scope level, rollback token, and status summary chain fragment. Mark rejected or adjusted edit items as deviation samples and adopted edit items as positive samples, and form a grouped statistical index by rule item, by chapter, and by template. Based on the feedback record set, incremental updates are performed on the parameters of the rule potential field, the execution strength, priority and exception inheritance chain of the effective rules are adjusted, hard constraints are not reduced, suggested constraints are added or removed according to feedback, the rule version identifier and timestamp are updated, and a new rule potential field is output. Based on the feedback record set, the semantic constraint weights are incrementally updated, the semantic propagation weights, conflict adjudication order and cross-graph reference alignment weights are adjusted, nodes that remain consistent for two consecutive rounds are marked as steady-state nodes, the suppression weights are added to edit modes that are rejected multiple times, and new propagation allow and prohibition tags are generated. Generate updated manuscript content and review records, and output a complete audit result including the final revision list, the list of unused edits, the rule version change record, and the status summary chain.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to an academic publishing management platform of a university. This platform has long faced problems such as low review efficiency, weak semantic consistency detection capabilities, and inconsistent formatting requirements among different departments. Traditional systems rely on template matching and keyword verification, leading to frequent occurrences of "grammatically correct but semantically incorrect," "fragmented modifications," and "lagging rule updates." The system proposed in this invention combines evolvable semantic constraint graph networks, rule potential field modeling, and continuous reversible editing field generation technology, significantly improving the above-mentioned problems in automatic manuscript review.

[0028] In practical applications, the system first performs text preprocessing and semantic modeling on the uploaded manuscripts. Taking science and engineering academic papers as an example, a single paper is approximately 42,000 words long. During the parsing phase, the system constructs a semantic graph through syntactic dependency analysis and named entity recognition, generating an average of about 11,000 semantic nodes. Subsequently, the system loads journal standards and proofreading guidelines, generates a rule potential field, and implements three layers of constraint propagation in the evolvable semantic constraint graph network: the rule potential alignment layer ensures that the manuscript follows academic norms; the semantic propagation and constraint fusion layer strengthens contextual logic and sentence consistency; and the editing feasibility projection layer generates candidate editing anchors. Through this mechanism, the system can automatically detect multi-level semantic conflicts, incoherent expressions, and format violations.

[0029] The generated reversible edit field contains candidate edit vectors and corresponding discrete edit scripts. Each script identifies the modification type, location, rule source, and confidence level. The system displays modification suggestions visually in the interface, allowing users to adopt, reject, or fine-tune modifications in real time. The system dynamically updates rule potential parameters and semantic constraint weights based on user feedback, gradually adapting them to the expression habits and format requirements of different disciplines. During the testing period, the system processed 156 papers, detecting approximately 2,960 semantic and format issues, with 82.7% of the modifications adopted by users. The average review time per manuscript was 3 minutes and 12 seconds, significantly lower than the average 45 minutes for manual review. Statistics show that the overall standardization of manuscripts improved by 27.4%, logical consistency issues decreased by 38.2%, and repetition errors dropped to 7.1%.

[0030] By applying the method of this invention, the system not only achieves efficient automatic review but also optimizes the model with each user feedback, forming a continuous learning loop. Users generally believe that the system outperforms existing technologies in semantic understanding, rule compliance, and modification traceability, especially in the abstract, methods, and conclusion sections of research papers, where the system can accurately identify logical breaks and inconsistencies in terminology.

[0031] Table 1. Statistical Table of Experimental Results of Academic Journal Paper Review System

[0032] As shown in Table 1, the online manuscript review assistance method based on natural language processing proposed in this invention significantly outperforms traditional review methods in several key indicators. In terms of efficiency, the average review time of this invention is only 3 minutes and 12 seconds, which is approximately 93% shorter than manual review and approximately 65% ​​faster than template-based review systems, demonstrating algorithm-level efficiency and automation advantages. This is attributed to the system's collaborative processing in semantic graph construction, rule potential constraints, and editing path optimization, enabling the semantic understanding and rule matching of large-scale manuscripts to be completed in a short time.

[0033] In terms of accuracy and intelligence in proofreading, the method of this invention achieves a semantic error detection rate of 94.8% and a format and specification error recognition rate of 96.3%, both significantly higher than the detection levels of existing template proofreading systems. The system utilizes an evolvable semantic constraint graph network to achieve contextual logical consistency and semantic coherence analysis, effectively solving the problem of "grammatically correct but semantically non-compliant" proofreading in the past. The logical consistency problem reduction rate reaches 38.2%, indicating that the system can automatically identify logical breaks in the paper's structure and between chapters, and provide reasonable revision suggestions, thereby significantly improving the overall logical quality of the manuscript.

[0034] Regarding sustainable optimization and user interaction, this invention supports a self-learning mechanism based on user feedback. It automatically updates rule potential parameters and semantic constraint weights after each task, achieving adaptive optimization across different disciplines and review standards. The user adoption rate for modifications is as high as 82.7%, indicating that the generated review suggestions have high practicality and credibility. The system's duplicate error rate is only 7.1%, demonstrating continuous learning and error correction capabilities. This invention not only improves the accuracy and consistency of manuscript review but also achieves dynamic optimization and long-term stable operation through feedback learning, possessing significant engineering application value and promotion potential.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assisting online manuscript review based on natural language processing, characterized in that, include: Obtain the text of the manuscript to be reviewed and proofread, perform segmentation, sentence segmentation, lexical analysis and syntactic dependency analysis on the manuscript, identify entity and logical relationship, and construct semantic graph; The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines, and the review rules are parsed into rule feature sets. Based on the rule type and applicable conditions, a rule potential field is constructed, and semantic constraints are applied to the semantic graph nodes. The semantic graph and the rule potential field are input into the evolvable semantic constraint graph network. The semantic constraint evolution operation is performed on the semantic nodes. The node representation is dynamically updated according to the semantic propagation relationship of the nodes and the rule potential gradient, so as to obtain the set of semantic nodes after semantic constraint evolution and the corresponding evolved semantic graph. The semantic node set after semantic constraint evolution is input into the reversible edit field generation module to generate a continuous reversible edit field. The continuous reversible edit field contains multiple edit vectors, and each edit vector describes the potential edit type, operation direction and confidence of the corresponding node. Based on the optimal path solution of the continuous reversible editing field, the optimal editing sequence that satisfies the condition of minimizing the rule potential energy is determined, forming a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints, and outputting the modification location, modification content and trigger rule identifier; The revised manuscript is displayed on the online review interface. Based on the user's acceptance, rejection, or adjustment feedback, the parameters and semantic constraint weights of the rule potential field are updated to achieve adaptive optimization and continuous learning for different manuscript types and review standards.

2. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The construction of the semantic graph refers to extracting semantic units and establishing a set of nodes based on the dependency relationships, referential relationships and logical connections between sentences after the manuscript text is segmented into sentences, words and syntactic dependency analysis. The semantically related nodes are then arranged into directed edges according to the syntactic master-slave relationship, contextual succession relationship and semantic dependency relationship, forming a multi-layer semantic graph structure that includes node attributes and edge weights.

3. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The construction of a rule potential field based on rule type and applicable conditions, and the application of semantic constraints to semantic graph nodes, includes: The review rules are extracted from the review guidelines library, style standards library and journal submission guidelines. Each rule is structured and parsed to determine its scope of application, triggering conditions, exception conditions, execution level, priority and version identifier, forming a rule feature set and storing it in the form of entries. Based on the constructed semantic graph, the rule feature set is mapped one-to-one with the semantic graph nodes according to the hierarchy of chapter, title, paragraph, sentence and phrase, and the rule application list, exception list and application boundary of each node are determined, and the binding of rules to nodes and the disabling mark are completed. A hierarchical description of the rule potential field is generated at three levels: node level, edge level, and region level. At the node level, the execution strength of mandatory constraints and suggested constraints is determined based on textual and structural attributes. At the edge level, a scope is established for reference consistency, logical consistency, and reference chain consistency. At the region level, a templated constraint domain is established for the same chapter or layout template. The three layers of description together constitute the rule potential expression of the semantic graph. Establish a rule conflict decomposition and dominance selection table, perform conflict detection and adjudication on multiple rules in the same node or the same scope, determine the final effective rule set in the order of mandatory over suggested, narrow domain over wide domain, and latest version over old version, and generate a traceable exception inheritance chain and conflict relationship diagram. The final set of rules that have been adjudicated and the hierarchical rule potential energy description are combined to form the rule potential energy field output, which imposes deterministic semantic constraints on each node and related edge in the semantic graph.

4. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The process of obtaining the semantic node set after semantic constraint evolution and the corresponding evolved semantic graph includes: The semantic graph and rule potential field are simultaneously input into the evolvable semantic constraint graph network. A three-layer structure consisting of a rule potential alignment layer, a semantic propagation and constraint fusion layer, and an editing feasibility projection layer is initialized, and the final effective rule set, exception list, and scope boundary are loaded. In the rule potential alignment layer, each semantic node in the semantic graph is bound to a corresponding effective rule entry. Based on the node attributes, structural position and rule scope, node-level and edge-level propagation allow and prohibition markers are generated. An exception inheritance chain is established and a rule subgraph divided by chapter and template is generated. At the same time, a cross-graph reference alignment channel is established to constrain cross-paragraph associations involving charts, formulas and references. In the semantic propagation and constraint fusion layer, constrained information propagation is performed on allowed edges based on propagation tags and rule subgraphs. Convergence calculation is performed on nodes within the same scope, and the convergence result is sent back to the node through a region consistency loop, maintaining chapter-level consistency and template-level consistency, while maintaining the constraint of cross-graph reference alignment channels and eliminating the offset of reference number and reference chain. In the conflict adjudication subprocess of the semantic propagation and constraint fusion layer, multiple rules within the same node or the same local subgraph are adjudicated. The unique effective rule is determined in the order of priority of mandatory rules, priority of narrow scope, and priority of latest version. Nodes with consistent update results in two consecutive updates are marked as steady-state nodes, and nodes with differences are marked as nodes to be evolved. In the editing feasibility projection layer, the current state of the node marked as to be evolved is projected to generate candidate editing anchors that correspond one-to-one with the node. The editing type, operation direction, scope and triggering rule identifier of each anchor are given, and a traceable index from the anchor to the original semantic node and scope is established. Output the set of semantic nodes and the evolved semantic graph after semantic constraint evolution, along with the candidate edit anchor index, propagation allow flag, prohibit flag, exception inheritance chain and cross-graph reference alignment constraint.

5. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The step of inputting the semantically constrained evolved set of semantic nodes into the reversible edit field generation module to generate a continuous reversible edit field includes: The structure of the initial reversible edit field generation module includes a rule-driven direction derivation unit, an edit operation basis mapping unit, and a vector-script bidirectional quantization and rollback verification unit. The rule-driven direction derivation unit generates an initial edit vector for each semantic node and its corresponding candidate edit anchor point, determines the operation direction and intensity of the initial edit vector, the operation direction is determined based on the rule potential field and the current state of the node, and the intensity is determined based on the rule enforcement degree, scope of application and node confidence, and binds hard constraint references, suggested constraint references and safety threshold sets to each initial edit vector, and sets cross-scope consistency guardrails based on cross-graph reference alignment constraints; The rule-driven direction derivation unit completes the one-to-one alignment of the initial edit vector with the candidate edit anchor points, performs unique processing on the duplicate vectors generated by adjacent anchor points, forms a set of anchor point-level edit vectors, and marks each edit vector with the trigger rule identifier, scope level and exception inheritance chain location information; The anchor-level edit vector is projected onto a predefined set of edit operation bases by the edit operation base mapping unit. The set of edit operation bases includes insertion, deletion, replacement, rearrangement, and format repair. Weights of each operation base are assigned to each edit vector, and the sum of the weights is limited to no more than one. An operation allowed list and an operation prohibited list are generated based on the propagation allowed and prohibited flags. The scope and precise boundary of each edit vector are determined by a multi-granularity scope edit grid. The vector-script bidirectional quantization and rollback verification unit converts the edit vector into discrete edit script records. The edit script records include operation type, scope identifier, trigger rule identifier, operation confidence, state summary before change and state summary after change. A corresponding reverse mapping is generated to support the recovery of the edit vector from the edit script. A rollback token is generated for each edit script record and a sequentially verifiable state summary chain is constructed to complete the reversibility verification and traceability verification. Output a continuous and reversible edit field data structure, including an anchor-level edit vector set, a corresponding edit script set, an operation allowed list, an operation prohibited list, cross-scope consistency guardrails, and a multi-granularity scope edit grid.

6. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The process of creating a manuscript revision scheme that meets the requirements of semantic consistency and rule constraints includes: Receive the set of anchor-level edit vectors, the corresponding set of edit scripts, the list of allowed operations, the list of prohibited operations, the cross-scope consistency guardrail, and the multi-granularity scope edit grid. Merge and number all candidate edit items according to the semantic node order and the trigger rule identifier. Based on the list of allowed operations and the list of prohibited operations, the feasibility of candidate edit items is screened, edit items that violate hard constraints are eliminated, and edit items that meet the cross-scope consistency guardrail and multi-granularity scope editing grid boundary conditions are retained. Sequence dependencies and mutual exclusion conflicts are generated. A comprehensive cost index is calculated for feasible editable items. The comprehensive cost index consists of three parts: the editing range, the remaining amount of suggested constraints, and the cross-scope cost. The individual cost is synthesized according to the preset weights, and constraint satisfaction labels with priority given to hard constraint satisfaction are generated at the same time. Under the condition of satisfying all hard constraints and sequential dependencies, the optimal path solving program is executed. The editing path is selected according to the criteria of minimizing the comprehensive cost and maximizing the satisfaction of the suggested constraints, so as to obtain the optimal editing path and the corresponding minimum intervention editing sequence, and maintain a one-to-one mapping relationship with the original editing script record. The manuscript revision scheme is generated based on the minimum intervention editing sequence. The scheme outputs the modification location, modification type, trigger rule identifier, operation confidence, rollback token and status summary chain fragment in the execution order. It also outputs a list of candidate editing items that were not adopted and their corresponding constraint descriptions.

7. The online manuscript review assistance method based on natural language processing according to claim 1, characterized in that, The step of updating the parameters and semantic constraint weights of the rule potential field based on the user's acceptance, rejection, or adjustment of feedback information includes: The system receives manuscript revision plans, optimal editing paths, minimal intervention editing sequences, rollback tokens, and status summary chains. The online review interface displays the modification location, modification type, trigger rule identifier, and operation confidence level of each editing item in an itemized format. Collect the user's operation results for each edit item. The operation results include three states: adoption, rejection, and adjustment. For edit items in the adjustment state, record the user's added or deleted edit content and the corresponding change information of the scope of effect, and establish a one-to-one mapping relationship with the original editing script. Generate a set of feedback records, and mark each edit item with its final status, trigger rule identifier, scope level, rollback token, and status summary chain fragment. Mark rejected or adjusted edit items as deviation samples and adopted edit items as positive samples, and form a grouped statistical index by rule item, by chapter, and by template. Based on the feedback record set, incremental updates are performed on the parameters of the rule potential field, the execution strength, priority and exception inheritance chain of the effective rules are adjusted, hard constraints are not reduced, suggested constraints are added or removed according to feedback, the rule version identifier and timestamp are updated, and a new rule potential field is output. Based on the feedback record set, the semantic constraint weights are incrementally updated, the semantic propagation weights, conflict adjudication order and cross-graph reference alignment weights are adjusted, nodes that remain consistent for two consecutive rounds are marked as steady-state nodes, the suppression weights are added to edit modes that are rejected multiple times, and new propagation allow and prohibition tags are generated. Generate updated manuscript content and review records, and output a complete audit result including the final revision list, the list of unused edits, the rule version change record, and the status summary chain.

Images