An artificial intelligence-based document proofing system and method

Abstract

The application discloses a document proofreading system and method based on artificial intelligence, and relates to the technical field of natural language processing.The method comprises the following steps: analyzing a document, dividing semantic units, and marking based on a context type library to obtain a document with context labels; using a corresponding rule set to perform initial error detection, comparing the format and text quality characteristics of the document with a standard sample vector space to obtain an initial error set; constructing a defect propagation chain network, locating root error nodes that cause multiple secondary errors by tracing the defect propagation chain network; generating an intelligent proofreading report; receiving a new version of the document revised by a user, comparing the new version of the document with the original document, locating a change area, and analyzing the quality difference of the change area in related quality dimensions; based on the quality difference, outputting an incremental proofreading report, wherein the incremental proofreading report comprises newly introduced errors and unsolved root errors.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a document proofreading system and method based on artificial intelligence. Background Technology

[0002] With the widespread adoption of digital office work, the writing and review of various documents has become increasingly demanding. Traditional document proofreading relies mainly on manual labor, which is not only inefficient but also prone to omissions and errors due to proofreader fatigue. Especially for highly specialized content such as technical documents and academic papers, proofreading involves not only basic grammar and spelling checks but also a comprehensive evaluation of multiple dimensions, including logical consistency, data accuracy, and formatting compliance.

[0003] Most automated proofreading tools currently on the market are based on predefined rule bases or basic statistical models, capable of detecting only surface-level grammatical and spelling errors, and unable to deeply understand the semantic content and logical structure of documents. These tools fall short when dealing with complex logical contradictions, data inconsistencies, and other deep-seated problems. Furthermore, existing tools typically treat various errors in isolation, lacking analysis of the correlations between errors, and failing to identify and locate the root causes of a series of problems, forcing users to spend a significant amount of time on repetitive corrections. Moreover, existing intelligent proofreading systems lack continuous learning capabilities. They cannot learn from user correction behavior and cannot adapt to the proofreading needs of documents in different fields and styles, significantly limiting the system's accuracy and adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based document proofreading system and method to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a document proofreading method based on artificial intelligence, the method comprising:

[0006] S100. Parse the document, divide it into semantic units, and label the semantic unit types based on the context type library to obtain document data with context labels; the context type library is obtained by training based on the standard document sample library; the training process establishes the mapping relationship between text features and context types by associating the text features and semantic functions of semantic units in the standard document sample library.

[0007] S200. Based on context labels, perform initial error detection using the corresponding rule set, and compare the document's format and text quality features with the standard sample vector space to calculate the quality deviation of each dimension and obtain the initial error set.

[0008] S300. Using the initial set of errors as nodes, analyze the causal relationship between error nodes according to the defect propagation rule base, construct a defect propagation chain network, and locate the root cause error node that causes multiple secondary errors by tracing the network.

[0009] S400: Based on root cause error nodes and defect propagation chains, generate intelligent verification reports. The intelligent verification reports include error location, priority correction strategies for root cause errors, and global correction suggestions.

[0010] S500: Receives a new version of the document revised by the user based on the intelligent proofreading report, compares the new version of the document with the original document, locates the changed area, and calculates the quality difference of the changed area on the relevant quality dimensions.

[0011] S600: Based on quality difference, outputs incremental verification reports, which include newly introduced errors and unresolved root cause errors; continuously optimizes context type annotation, defect propagation rule base and root cause error determination logic using quality difference data.

[0012] Based on the above, step S100 includes:

[0013] S110. Perform syntactic structure and semantic dependency analysis on the input document to identify and divide the independent semantic units that constitute the logical structure of the document.

[0014] S120. Extract the text features of each semantic unit. The text features include grammatical structure pattern, keyword distribution, semantic role, and logical relationship with adjacent units. The grammatical structure pattern refers to the sentence structure and component combination of the semantic unit. The keyword distribution refers to the occurrence position and frequency characteristics of core terms in the semantic unit. The semantic role refers to the functions of each component in the semantic unit, such as statement, modification, and limitation. The logical relationship refers to the causal, progressive, and parallel relationship with adjacent semantic units.

[0015] S130. Input the text features into the pre-trained context classification model. The context classification model learns from a large number of standard document samples through machine learning algorithms to establish a mapping relationship from text features to context types. The pre-training process of the context classification model adopts a supervised learning method, using the semantic unit types manually annotated in the standard document sample library as labels to optimize model parameters and improve classification accuracy.

[0016] S140. Based on the output of the context classification model, assign a specific context type label to each semantic unit to generate document data with context labels; the context types include technical definition statements, data argument statements, reasoning argument statements, example description statements, and background introduction statements.

[0017] Based on the above, step S200 includes:

[0018] S210. Based on the context type tag of the semantic unit, call the special error verification rule set pre-bound to the context type tag to perform syntax rule compliance check and semantic logic consistency check on the semantic unit; each context type tag corresponds to a unique special error verification rule set, and the verification dimensions and judgment criteria of the special error verification rule set are customized according to the semantic function and quality requirements of the corresponding context type to ensure the pertinence and accuracy of the verification.

[0019] S220. Extract global formatting features and text quality statistical features from the document. Formatting features include heading hierarchy, numbering sequence consistency and chart labeling standardization. Text quality statistical features include terminology usage consistency, paragraph length distribution and logical connector density.

[0020] S230. The extracted format specification features and text quality statistical features are mapped to a multi-dimensional vector space constructed based on learning from a large number of standard document samples. The Euclidean distance between the feature vector of the document to be corrected and the cluster center of the feature vector of the standard samples in each dimension is calculated as the quality deviation of each dimension. The multi-dimensional standard sample vector space is constructed by clustering the feature data of all samples in the standard document sample library. The cluster center is the standard benchmark value of each feature dimension to ensure the objectivity and uniformity of the quality assessment.

[0021] S240. Integrate rule violation errors identified by the special error verification rule set and statistical anomaly errors identified by the quality deviation calculation to generate an initial error set; during the integration process, deduplication is performed by error location and error type, and the feature dimension data and context label corresponding to each error are recorded.

[0022] Based on the above, context types include technical definition statements, data argumentation statements, reasoning argumentation statements, example description statements, and background introduction statements; the specific error verification rule set is configured based on different context type tags;

[0023] For the rule set configured for technical definition statements, the verification focus is on the uniqueness, completeness, and unambiguity of the core terminology definitions; for the rule set configured for data argumentation statements, the verification focus is on the uniformity of units for numerical data, the standardization of data source citations, and the rationality of the supporting relationship between data and conclusions; for the rule set configured for reasoning argumentation statements, the verification focus is on the logical completeness of causal relationships in the argumentation chain and the validity of reasoning premises; for the rule set configured for example illustration statements, the verification focus is on the relevance of examples to the abstract concepts or rules being explained; and for the rule set configured for background introduction statements, the verification focus is on the objectivity of historical fact statements and basic grammatical compliance.

[0024] Based on the above, step S300 includes:

[0025] S310. Abstract each error instance in the initial error set into a network node, and record the error type, semantic unit to which the network node belongs, and context label attribute; each network node has a unique identifier, and its attribute information corresponds one-to-one with the error details in the initial error set;

[0026] S320. Call the defect propagation rule library. The defect propagation rule library defines the causal relationship logic between nodes of different error types in a specific context. The defect propagation rule library includes error type association rules, context adaptation rules and causal strength determination rules. Error type association rules clarify the propagation possibility between different error types. Context adaptation rules limit the semantic scenarios in which the propagation relationship is established. Causal strength determination rules quantify the degree of propagation association between error nodes.

[0027] S330. Based on the defect propagation rule base, traverse all error node pairs. If the node pair satisfies the causal relationship conditions defined in the defect propagation rule base, then establish a directional propagation edge between the two nodes to construct a directed defect propagation chain network. The direction of the propagation edge is from the cause error node to the effect error node. At the same time, record the rule basis and causal strength value corresponding to the propagation relationship.

[0028] S340. Calculate the out-degree of each node in the defect propagation chain network and identify cyclic dependency paths in the network; the out-degree of a node is calculated using the following formula:

[0029] OD i =∑ j A ij ;

[0030] Among them, OD iLet represent the out-degree of node i; i and j represent the indices of the nodes, and ∀i∈[1,n], meaning it applies to all nodes in the network with indices from 1 to n; when calculating the out-degree of node i, node i is fixed, while node j needs to traverse all nodes, i.e., sum from 1 to n; n represents the total number of nodes in the network; A ij Let A be an element of the adjacency matrix. If there exists a propagating edge from node i to node j, then A... ij =1, otherwise it is A ij =0; Circular dependency paths are identified by traversing the directed propagation chain network, referring to a closed path that starts from a certain node, passes through a number of propagation edges, and returns to the starting node.

[0031] S350. Nodes whose out-degree value exceeds a preset threshold or are located at the starting point of a circular dependency path are identified as root cause error nodes that trigger multiple secondary errors. The conditions for identifying a node as a root cause error node are as follows:

[0032] When the out-degree of node i is greater than the out-degree threshold θ, i.e., OD i When the value is greater than θ, the root cause error node determination condition is met. The out-degree threshold θ is determined based on the out-degree statistics of historical error nodes in the standard document sample library. The dynamic threshold is set by statistically analyzing the out-degree distribution characteristics of historical root cause error nodes to adapt to document scenarios of different complexities.

[0033] When node i belongs to the set V of starting nodes of the circular dependency path c When, i∈V c When the root cause error node determination condition is met; the set of starting nodes of the circular dependency path is identified by traversing the directed propagation chain network, and contains all elements that are the starting nodes of the circular dependency path;

[0034] A node is identified as a root cause fault node if it meets any of the above conditions, and the two conditions have equal validity.

[0035] Based on the above, step S400 includes:

[0036] S410. Generate error location information, which includes detailed information about the root cause error node and all secondary error nodes directly and indirectly associated with the root cause error node in the defect propagation chain network. The detailed information includes error content, error type, location information and the context type to which it belongs.

[0037] S420. Generate a priority correction strategy for root cause errors. The priority correction strategy includes determining the priority order of correction operations based on the attributes of the root cause error node and the centrality of the root cause error node in the defect propagation chain network, as well as a description of how secondary error nodes associated with the root cause error node will be automatically eliminated or transformed into a pending state after the root cause error is corrected. The centrality of the root cause error node is calculated based on its out-degree, propagation path coverage, and the number of associated error nodes in the defect propagation chain network. The higher the centrality of the root cause error node, the higher the correction priority.

[0038] S430. Generate global correction suggestions. Based on the topological analysis of the defect propagation chain network, the global correction suggestions generate the systemic logical defects and structural problems existing in the document, and provide modification directions for overall document quality optimization.

[0039] Based on the above, step S500 includes:

[0040] S510: Receive the revised new version of the document submitted by the user and align the new version of the document with the original document; based on the text difference algorithm, compare the new version of the document with the original document line by line to identify and locate the text change areas in the document where addition, deletion and modification operations have occurred.

[0041] S520. Within the text change area, recalculate the format specification features and text quality statistical features, and re-execute error detection to obtain the revised error status and quality deviation value.

[0042] S530. Compare the revised error status and quality deviation value of the changed area with the status of the corresponding area in the original document to calculate the quality difference. The quality difference includes the amount of change in the quality deviation value of the fixed error, the newly introduced error, and the quality deviation value. Fixed errors are determined by comparing the difference in error status before and after the revision. Newly introduced errors are obtained by identifying the error instances added after the revision. The amount of change in the quality deviation value is calculated by subtracting the deviation value of the corresponding area in the new version from that in the original version.

[0043] Based on the above, step S600 includes:

[0044] S610. Generate an incremental proofreading report, which includes the newly introduced set of errors and unresolved root cause error nodes in the quality difference representation, as well as the precise location information of the root cause error nodes in the document.

[0045] S620. Associate the quality difference data with the corresponding document revision records and store them to build a standard document sample library for context type library optimization;

[0046] S630. Based on the standard document sample library, incrementally train the context classification model used for context type annotation to optimize the classification accuracy of the context classification model.

[0047] S640. Based on examples of error propagation in the standard document sample library, optimize the defect propagation rule library by adding new propagation rules or adjusting the confidence level of existing rules.

[0048] S650. Based on the root cause error determination results in the standard document sample library, optimize the out-degree threshold and the sensitivity of identifying circular dependency paths in the root cause error determination logic.

[0049] An artificial intelligence-based document proofreading system includes: a document parsing module, an error detection module, a proofreading report module, a version comparison module, and a model optimization module;

[0050] The document parsing module includes a semantic parsing unit and a feature extraction unit. The semantic parsing unit is used to parse the document, divide it into semantic units, and label the semantic unit types based on the context type library to generate document data with context labels. The feature extraction unit is used to extract global format specification features and text quality statistical features from the document.

[0051] The error detection module includes an initial detection unit, a defect propagation chain unit, and a root cause error unit. The initial detection unit is used to call the corresponding rule set based on the context label to perform error detection, and at the same time compares the extracted features with the standard sample vector space to calculate the quality deviation and obtain an initial error set. The defect propagation chain unit is used to construct a defect propagation chain network based on the defect propagation rule base, using the initial error set as nodes. The root cause error unit is used to locate the root cause error node that causes multiple secondary errors by tracing the defect propagation chain network.

[0052] The proofreading report module includes a report generation unit and an incremental report generation unit. The report generation unit is used to generate an intelligent proofreading report containing error location, priority correction strategy and global correction suggestions based on root cause error nodes and defect propagation chains. The incremental report generation unit is used to output an incremental proofreading report containing newly introduced errors and unresolved root cause errors based on quality differences.

[0053] The version comparison module includes a change location unit and a quality difference unit. The change location unit is used to receive the revised new version of the document and compare it with the original document to locate the changed area. The quality difference unit is used to calculate the quality difference of the changed area.

[0054] The model optimization module includes a sample library unit and an optimization unit. The sample library unit is used to store quality difference data and document revision records to build a standard document sample library. The optimization unit is used to continuously optimize the context type annotation, defect propagation rule library, and root cause error determination logic using the quality difference data.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] 1. This invention constructs a defect transmission chain network, analyzes the causal relationship between different errors, accurately locates the root cause of multiple secondary errors, and thus provides a corrective solution that addresses both the symptoms and the root cause, significantly improving document correction efficiency;

[0057] 2. This invention combines rule-based error detection and vector space-based quantitative evaluation to comprehensively assess document quality from multiple dimensions such as syntax, format, logic, and data consistency, thereby improving the accuracy and comprehensiveness of proofreading.

[0058] 3. This invention focuses on newly introduced problems and unresolved root cause errors in a new round of proofreading by comparing versions and calculating quality differences, providing users with accurate incremental proofreading reports and optimizing the workflow of multi-round proofreading. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the steps of an artificial intelligence-based document proofreading method according to the present invention.

[0060] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based document proofreading system according to the present invention. Detailed Implementation

[0061] 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.

[0062] Example: Figures 1-2 As shown, the present invention provides a technical solution, an artificial intelligence-based document proofreading method, the method comprising:

[0063] S100. Parse the document, divide it into semantic units, and label the semantic unit types based on the context type library to obtain document data with context labels; the context type library is obtained by training based on the standard document sample library; the training process establishes the mapping relationship between text features and context types by associating the text features and semantic functions of semantic units in the standard document sample library.

[0064] Specifically, step S100 includes:

[0065] S110. Perform syntactic structure and semantic dependency analysis on the input document to identify and divide the independent semantic units that constitute the logical structure of the document.

[0066] S120. Extract the text features of each semantic unit. The text features include grammatical structure pattern, keyword distribution, semantic role, and logical relationship with adjacent units. The grammatical structure pattern refers to the sentence structure and component combination of the semantic unit. The keyword distribution refers to the occurrence position and frequency characteristics of core terms in the semantic unit. The semantic role refers to the functions of each component in the semantic unit, such as statement, modification, and limitation. The logical relationship refers to the causal, progressive, and parallel relationship with adjacent semantic units.

[0067] For example: the document to be proofread is parsed and divided into 128 semantic units; the text feature vector of each semantic unit is extracted, where: the grammatical structure pattern feature dimension is 15 dimensions, the keyword distribution feature dimension is 20 dimensions, the semantic role feature dimension is 8 dimensions, and the logical relationship feature dimension is 6 dimensions; this is just an example and is not a limitation.

[0068] S130. Input the text features into the pre-trained context classification model. The context classification model learns from a large number of standard document samples through machine learning algorithms to establish a mapping relationship from text features to context types. The pre-training process of the context classification model adopts a supervised learning method, using the semantic unit types manually annotated in the standard document sample library as labels to optimize model parameters and improve classification accuracy.

[0069] S140. Based on the output of the context classification model, assign a specific context type label to each semantic unit to generate document data with context labels; the context types include technical definition statements, data argument statements, reasoning argument statements, example description statements, and background introduction statements;

[0070] For example, when a 49-dimensional feature vector is input into a pre-trained context classification model, the model outputs the type probability distribution of each semantic unit. For example, the probability distribution of the 25th semantic unit is as follows: technical definition statement: 0.12, data argument statement: 0.08, reasoning argument statement: 0.65, example description statement: 0.10, background introduction statement: 0.05. Based on the principle of maximum probability, this unit is labeled as the reasoning argument statement type. This is only an example and is not a limitation.

[0071] S200. Based on context labels, perform initial error detection using the corresponding rule set, and compare the document's format and text quality features with the standard sample vector space to calculate the quality deviation of each dimension and obtain the initial error set.

[0072] Specifically, step S200 includes:

[0073] S210. Based on the context type tag of the semantic unit, call the special error verification rule set pre-bound to the context type tag to perform syntax rule compliance check and semantic logic consistency check on the semantic unit; each context type tag corresponds to a unique special error verification rule set, and the verification dimensions and judgment criteria of the special error verification rule set are customized according to the semantic function and quality requirements of the corresponding context type to ensure the pertinence and accuracy of the verification.

[0074] For example: Semantic unit No. 25, the reasoning and argument statement type, calls the rule set configured for the reasoning and argument statement type and finds an error of incomplete argument chain, with a severity score of 0.75; This is only an example and is not a limitation.

[0075] S220. Extract global formatting features and text quality statistical features from the document. Formatting features include heading hierarchy, numbering sequence consistency and chart labeling standardization. Text quality statistical features include terminology usage consistency, paragraph length distribution and logical connector density.

[0076] S230. The extracted format specification features and text quality statistical features are mapped to a multi-dimensional vector space constructed based on learning from a large number of standard document samples. The Euclidean distance between the feature vector of the document to be corrected and the cluster center of the feature vector of the standard samples in each dimension is calculated as the quality deviation of each dimension. The multi-dimensional standard sample vector space is constructed by clustering the feature data of all samples in the standard document sample library. The cluster center is the standard benchmark value of each feature dimension to ensure the objectivity and uniformity of the quality assessment.

[0077] S240. Integrate rule violation errors identified by the special error verification rule set and statistical anomaly errors identified by the quality deviation calculation to generate an initial error set; during the integration process, deduplication is performed by error location and error type, and the feature dimension data and context label corresponding to each error are recorded.

[0078] For example: Calculating the quality deviation of a document in the terminology consistency dimension: Standard sample feature vector cluster centers: 12.5, 8.3, 15.7, ..., 9.2; Document feature vector to be checked: 13.1, 7.8, 14.9, ..., 8.7 (20 dimensions); The calculated quality deviation D in this dimension is 2.34, exceeding the threshold of 1.5, and is recorded as a statistical anomaly error; This is only an example and is not a limitation.

[0079] Furthermore, the context types include technical definition statements, data argument statements, reasoning argument statements, example description statements, and background introduction statements. Specific error verification rule sets are configured based on different context type tags. For the rule set configured for technical definition statements, the verification focus is on the uniqueness, completeness, and unambiguity of core term definitions. For the rule set configured for data argument statements, the verification focus is on the uniformity of numerical data units, the standardization of data source citations, and the rationality of the supporting relationship between data and conclusions. For the rule set configured for reasoning argument statements, the verification focus is on the logical completeness of causal relationships in the argument chain and the validity of reasoning premises. For the rule set configured for example description statements, the verification focus is on the relevance of the example to the abstract concept or rule being explained. For the rule set configured for background introduction statements, the verification focus is on the objectivity of historical fact statements and basic grammatical compliance.

[0080] S300. Using the initial set of errors as nodes, analyze the causal relationship between error nodes according to the defect propagation rule base, construct a defect propagation chain network, and locate the root cause error node that causes multiple secondary errors by tracing the network.

[0081] Specifically, step S300 includes:

[0082] S310. Abstract each error instance in the initial error set into a network node, and record the error type, semantic unit to which the network node belongs, and context label attribute; each network node has a unique identifier, and its attribute information corresponds one-to-one with the error details in the initial error set;

[0083] S320. Call the defect propagation rule library. The defect propagation rule library defines the causal relationship logic between nodes of different error types in a specific context. The defect propagation rule library includes error type association rules, context adaptation rules and causal strength determination rules. Error type association rules clarify the propagation possibility between different error types. Context adaptation rules limit the semantic scenarios in which the propagation relationship is established. Causal strength determination rules quantify the degree of propagation association between error nodes.

[0084] S330. Based on the defect propagation rule base, traverse all error node pairs. If the node pair satisfies the causal relationship conditions defined in the defect propagation rule base, then establish a directional propagation edge between the two nodes to construct a directed defect propagation chain network. The direction of the propagation edge is from the cause error node to the effect error node. At the same time, record the rule basis and causal strength value corresponding to the propagation relationship.

[0085] S340. Calculate the out-degree of each node in the defect propagation chain network and identify cyclic dependency paths in the network; the out-degree of a node is calculated using the following formula:

[0086] OD i =∑ j A ij ;

[0087] Among them, OD i Let represent the out-degree of node i; i and j represent the indices of the nodes, and ∀i∈[1,n], meaning it applies to all nodes in the network with indices from 1 to n; when calculating the out-degree of node i, node i is fixed, while node j needs to traverse all nodes, i.e., sum from 1 to n; n represents the total number of nodes in the network; A ij Let A be an element of the adjacency matrix. If there exists a propagating edge from node i to node j, then A... ij =1, otherwise it is A ij =0; Circular dependency paths are identified by traversing the directed propagation chain network, referring to a closed path that starts from a certain node, passes through a number of propagation edges, and returns to the starting node.

[0088] S350. Nodes whose out-degree value exceeds a preset threshold or are located at the starting point of a circular dependency path are identified as root cause error nodes that trigger multiple secondary errors. The conditions for identifying a node as a root cause error node are as follows:

[0089] When the out-degree of node i is greater than the out-degree threshold θ, i.e., OD i When the value is greater than θ, the root cause error node determination condition is met. The out-degree threshold θ is determined based on the out-degree statistics of historical error nodes in the standard document sample library. The dynamic threshold is set by statistically analyzing the out-degree distribution characteristics of historical root cause error nodes to adapt to document scenarios of different complexities.

[0090] When node i belongs to the set V of starting nodes of the circular dependency path c When, i∈V c When the root cause error node determination condition is met; the set of starting nodes of the circular dependency path is identified by traversing the directed propagation chain network, and contains all elements that are the starting nodes of the circular dependency path;

[0091] A node is identified as a root cause fault node if it meets any of the above conditions, and the two conditions have equal validity.

[0092] For example: 35 error instances are abstracted as network nodes, constructing a 35×35 adjacency matrix A; where: node 17 has a vague core concept definition, points to node 25, and the argument chain is incomplete; node 25 points to node 31, and the conclusion lacks support; corresponding matrix elements: A 17,25 =1, A 25,31 =1;

[0093] Calculate the out-degree (OD) of nodes 17 =4, OD 25 =2; Identify the circular dependency path: node 8-node 12-node 19-node 8, forming a circular path; This is just an example and is not a limitation.

[0094] S400: Based on root cause error nodes and defect propagation chains, generate intelligent verification reports. The intelligent verification reports include error location, priority correction strategies for root cause errors, and global correction suggestions.

[0095] Specifically, step S400 includes:

[0096] S410. Generate error location information, which includes detailed information about the root cause error node and all secondary error nodes directly and indirectly associated with the root cause error node in the defect propagation chain network. The detailed information includes error content, error type, location information and the context type to which it belongs.

[0097] S420. Generate a priority correction strategy for root cause errors. The priority correction strategy includes determining the priority order of correction operations based on the attributes of the root cause error node and the centrality of the root cause error node in the defect propagation chain network, as well as a description of how secondary error nodes associated with the root cause error node will be automatically eliminated or transformed into a pending state after the root cause error is corrected. The centrality of the root cause error node is calculated based on its out-degree, propagation path coverage, and the number of associated error nodes in the defect propagation chain network. The higher the centrality of the root cause error node, the higher the correction priority.

[0098] S430. Generate global correction suggestions. Based on the topological analysis of the defect propagation chain network, the global correction suggestions generate the systemic logical defects and structural problems existing in the document, and provide modification directions for overall document quality optimization.

[0099] For example: Generate error location information, including root cause error node 17 and its 8 associated secondary error nodes; calculate the centrality index of node 17: out-degree: 4, propagation path coverage: 12 nodes, number of associated error nodes: 9, centrality score: 8.5, correction priority: highest; This is just an example and is not a limitation.

[0100] S500: Receives a new version of the document revised by the user based on the intelligent proofreading report, compares the new version of the document with the original document, locates the changed area, and calculates the quality difference of the changed area on the relevant quality dimensions.

[0101] Specifically, step S500 includes:

[0102] S510: Receive the revised new version of the document submitted by the user and align the new version of the document with the original document; based on the text difference algorithm, compare the new version of the document with the original document line by line to identify and locate the text change areas in the document where addition, deletion and modification operations have occurred.

[0103] S520. Within the text change area, recalculate the format specification features and text quality statistical features, and re-execute error detection to obtain the revised error status and quality deviation value.

[0104] S530. Compare the revised error status and quality deviation value of the changed area with the status of the corresponding area in the original document to calculate the quality difference. The quality difference includes the amount of change in the quality deviation value of the fixed error, the newly introduced error, and the quality deviation value. Fixed errors are determined by comparing the difference in error status before and after the revision. Newly introduced errors are obtained by identifying the error instances added after the revision. The amount of change in the quality deviation value is calculated by subtracting the deviation value of the corresponding area in the new version from that in the original version.

[0105] For example: After the user makes revisions, the quality indicators of the changed area are recalculated: Fixed errors: 15 errors including nodes 17, 25, and 31; Newly introduced errors: 2 formatting errors; Changes in quality deviation: The terminology usage consistency dimension decreased from 2.34 to 1.12.

[0106] S600: Based on quality difference, outputs incremental verification reports, which include newly introduced errors and unresolved root cause errors; continuously optimizes context type annotation, defect propagation rule base and root cause error determination logic using quality difference data;

[0107] Specifically, step S600 includes:

[0108] S610. Generate an incremental proofreading report, which includes the newly introduced set of errors and unresolved root cause error nodes in the quality difference representation, as well as the precise location information of the root cause error nodes in the document.

[0109] S620. Associate the quality difference data with the corresponding document revision records and store them to build a standard document sample library for context type library optimization;

[0110] S630. Based on the standard document sample library, incrementally train the context classification model used for context type annotation to optimize the classification accuracy of the context classification model.

[0111] S640. Based on error propagation examples in the standard document sample library, perform strong optimization on the defect propagation rule library, add new propagation rules or adjust the confidence of existing rules;

[0112] S650. Based on the root cause error determination results in the standard document sample library, optimize the out-degree threshold and the sensitivity of identifying circular dependency paths in the root cause error determination logic.

[0113] The present invention provides another technical solution, an artificial intelligence-based document proofreading system, which includes: a document parsing module, an error detection module, a proofreading report module, a version comparison module, and a model optimization module;

[0114] The document parsing module includes a semantic parsing unit and a feature extraction unit. The semantic parsing unit is used to parse the document, divide it into semantic units, and label the semantic unit types based on the context type library to generate document data with context labels. The feature extraction unit is used to extract global format specification features and text quality statistical features from the document.

[0115] The error detection module includes an initial detection unit, a defect propagation chain unit, and a root cause error unit. The initial detection unit is used to call the corresponding rule set based on the context label to perform error detection, and at the same time compares the extracted features with the standard sample vector space to calculate the quality deviation and obtain an initial error set. The defect propagation chain unit is used to construct a defect propagation chain network based on the defect propagation rule base, using the initial error set as nodes. The root cause error unit is used to locate the root cause error node that causes multiple secondary errors by tracing the defect propagation chain network.

[0116] The proofreading report module includes a report generation unit and an incremental report generation unit. The report generation unit is used to generate an intelligent proofreading report containing error location, priority correction strategy and global correction suggestions based on root cause error nodes and defect propagation chains. The incremental report generation unit is used to output an incremental proofreading report containing newly introduced errors and unresolved root cause errors based on quality differences.

[0117] The version comparison module includes a change location unit and a quality difference unit. The change location unit is used to receive the revised new version of the document and compare it with the original document to locate the changed area. The quality difference unit is used to calculate the quality difference of the changed area.

[0118] The model optimization module includes a sample library unit and an optimization unit. The sample library unit is used to store quality difference data and document revision records to build a standard document sample library. The optimization unit is used to continuously optimize the context type annotation, defect propagation rule library, and root cause error determination logic using the quality difference data.

[0119] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An artificial intelligence-based document proofreading method, characterized by: The method comprises: S100, parsing the document, dividing semantic units, and obtaining a document with context labels based on context type library labeling; the context type library is obtained by training based on a standard document sample library; Step S100 comprises: S110, performing syntax structure and semantic dependency analysis on the input document to identify and divide independent semantic units constituting the logical structure of the document; S120, extracting text features of each semantic unit, including syntax structure mode, keyword distribution, semantic role, and logical association relationship with adjacent units; S130, inputting the text features into a pre-trained context classification model, the context classification model learns a large number of standard document samples through a machine learning algorithm to establish a mapping relationship from text features to context types; S140, based on the output of the context classification model, a specific context type label is assigned to each semantic unit to generate a document with context labels; the context types include technical definition statements, data argument statements, argument statements, instance explanation statements, and background introduction statements; S200, based on the context labels, use the corresponding rule set to perform initial error detection, and compare the format and text quality features of the document with the standard sample vector space to obtain an initial error set; S300, taking the initial error set as a node, analyzing the causal relationship between error nodes, constructing a defect propagation chain network, and locating the root error node causing multiple secondary errors by tracing the defect propagation chain network; Step S300 comprises: S310, abstracting each error instance in the initial error set as a network node, and recording the error type, semantic unit, and context label attributes of the network node; S320, calling a defect propagation rule library, which defines the causal relationship logic between different error type nodes in a specific context; S330, based on the defect propagation rule library, traversing all error node pairs, if the node pair meets the causal relationship condition defined in the defect propagation rule library, a directional propagation edge is established between the two nodes to construct a directed defect propagation chain network; S340, calculating the out-degree of each node in the defect propagation chain network, and identifying the circular dependency path in the network; S350, determining the root error node causing multiple secondary errors as a node with an out-degree value exceeding a preset threshold or located at the starting point of a circular dependency path; the determination conditions include: the first determination condition is that the out-degree value of the node is greater than the preset out-degree threshold, and the out-degree threshold is determined based on the out-degree statistical data of historical error nodes in the standard document sample library; the second determination condition is that the node belongs to the starting point node set of the circular dependency path, and the starting point node set of the circular dependency path is identified by traversing the directed propagation chain network and contains all elements as starting nodes of the circular dependency path; wherein, the node is determined as the root error node if it meets any of the determination conditions, and the two determination conditions have the same effect. S400, generating an intelligent proof report based on the root cause error node and the defect propagation chain network, the intelligent proof report including error positioning, a priority correction strategy for the root cause error, and a global correction suggestion; S500, receiving a new version of the document revised by the user, comparing the new version of the document with the original version of the document, positioning the changed area, and analyzing the quality difference of the changed area in the related quality dimension; S600, based on the quality difference, outputting an incremental proof report, the incremental proof report including newly introduced errors and unsolved root cause errors; using the quality difference data to optimize the context type library, the defect propagation rule library, and the root cause error judgment logic.

2. The method of claim 1, wherein: Step S200 includes: S210, according to the context type label of the semantic unit, calling a special error checking rule set pre-bound with the context type label, and performing syntax rule compliance check and semantic logic consistency check on the semantic unit; S220, extracting global format specification features and text quality statistical features from the document, the format specification features including title hierarchy structure, numbering sequence consistency, and chart labeling specification, and the text quality statistical features including term usage consistency, paragraph length distribution, and logical connection word density; S230, mapping the extracted format specification features and text quality statistical features to a multi-dimensional vector space learned and constructed based on a large number of standard document samples, calculating the Euclidean distance between the feature vector of the document to be proofread and the clustering center of the standard sample feature vector in each dimension as the quality deviation degree of each dimension; S240, integrating the rule violation errors identified by the special error checking rule set and the statistical abnormal errors identified by the quality deviation degree calculation, to generate an initial error set.

3. The document proofreading method based on artificial intelligence according to claim 2, characterized in that: The context types include technical definition statements, data argument statements, reasoning argument statements, instance description statements, and background introduction statements; and the special error checking rule set is configured based on different context type labels.

4. The document proofreading method based on artificial intelligence according to claim 1, characterized in that: The out-degree of the node is calculated as follows: OD i =∑ j A ij ; where OD i represents the out-degree of node i; i and j represent indices of nodes, and V i e [1, n] applies to all nodes with indices from 1 to n in the network; n represents the total number of nodes in the network; A ij represents an element of the adjacency matrix, and A ij = 1 if there is a conductive edge from node i pointing to node j, otherwise A ij = 0.

5. The method of claim 1, wherein: Step S400 includes: S410, generating error positioning information, the error positioning information including detailed information of the root cause error node and all secondary error nodes directly associated and indirectly associated with the root cause error node in the defect propagation chain network, the detailed information including error content, error type, location information, and belonging context type; S420, generating a priority correction strategy for the root cause error, the priority correction strategy including determining a priority order of correction operations based on the attributes of the root cause error node and the centrality of the root cause error node in the defect propagation chain network, and describing secondary error nodes associated with the root cause error node that will be automatically eliminated or transformed into a pending state after the root cause error is corrected. S430, generating a global correction suggestion, the global correction suggestion being generated based on a topology analysis of the defect propagation chain network, generating systematic logical defects and structural problems existing in the document, and providing a modification direction for overall document quality optimization.

6. The method of claim 1, wherein: Step S500 includes: S510, receiving a user-submitted revised new version document, and version aligning the new version document with the original version document; based on a text difference algorithm, comparing the new version document with the original version document line by line, identifying and locating the text change area in the document where the increase, deletion and modification operations occur; S520, in the text change area, re-computing the format specification features and text quality statistical features, and re-executing error detection to obtain the revised error state and quality deviation value; S530, comparing the revised error state and quality deviation value of the change area with the state of the corresponding area in the original version document, and calculating the quality difference; the quality difference includes repaired errors, newly introduced errors and the change amount of quality deviation value.

7. The method of claim 1, wherein: Step S600 includes: S610, generating an incremental proofreading report, the incremental proofreading report including a newly introduced error set represented by quality difference and an unsolved root cause error node, and accurate positioning information of the root cause error node in the document; S620, storing the quality difference data and the corresponding document revision record in association to construct a standard document sample library for context type library optimization; S630, based on the standard document sample library, incrementally training a context classification model used for context type labeling to optimize the classification accuracy of the context classification model; S640, based on the error propagation instances in the standard document sample library, strengthening and expanding the defect propagation rule library, adding new propagation rules or adjusting the confidence of existing rules; S650, based on the determination result of the root cause error in the standard document sample library, optimizing the out-degree threshold and the recognition sensitivity of the cyclic dependency path in the root cause error determination logic.

8. An artificial intelligence-based document proofreading system for implementing the artificial intelligence-based document proofreading method of any one of claims 1-7, characterized in that: The system includes a document analysis module, an error detection module, a proofreading report module, a version comparison module and a model optimization module; The document analysis module includes a semantic analysis unit and a feature extraction unit; the semantic analysis unit is used to analyze the document, divide semantic units, and label the semantic unit types based on the context type library to generate document data with context labels; the feature extraction unit is used to extract global format specification features and text quality statistical features from the document; The error detection module comprises an initial detection unit, a defect propagation chain unit and a root cause error unit; the initial detection unit is configured to call a corresponding rule set based on the context label to perform error detection, and to calculate a quality deviation degree by comparing the extracted features with a standard sample vector space, thereby obtaining an initial error set; the defect propagation chain unit is configured to take the initial error set as a node, and to construct a defect propagation chain network according to a defect propagation rule library; and the root cause error unit is configured to locate a root cause error node causing a plurality of secondary errors by tracing the defect propagation chain network. The proofreading report module comprises a report generation unit and an incremental report generation unit; the report generation unit is configured to generate an intelligent proofreading report containing error positioning, a priority correction strategy and a global correction suggestion based on the root cause error node and the defect propagation chain; and the incremental report generation unit is configured to output an incremental proofreading report containing newly introduced errors and unsolved root cause errors based on a quality difference. The version comparison module comprises a change positioning unit and a quality difference unit; the change positioning unit is configured to receive a revised new version document, and to compare the new version document with an original version document to position a changed area; and the quality difference unit is configured to calculate a quality difference of the changed area. The model optimization module comprises a sample library unit and an optimization unit; the sample library unit is configured to store quality difference data and document revision records, and to construct a standard document sample library; and the optimization unit is configured to continuously optimize context type labeling, a defect propagation rule library and root cause error determination logic by using the quality difference data.

Images