A method and system for reviewing textual content

Abstract

The application discloses a kind of text content's examination method and system, applied to text examination technical field, comprising: obtaining to-be-examined text data and examination rule base;Based on each semantic boundary, examination rule base is handled, obtains each first text segment, and determines the first semantic label corresponding to each first text segment;Each first semantic label is analyzed to be associated, to construct associated examination graph with association analysis result;Determine the second semantic label corresponding to each second text segment;From each first text segment, target text segment is screened based on second semantic label, associated examination graph and association analysis result;Based on target text segment, each second text segment is analyzed and handled, determines the local examination result corresponding to each second text segment, integrates local examination result, obtains the examination result of to-be-examined text data.The text content's examination method and system provided in the application effectively improve the reliability and accuracy of examination conclusion.

Description

Technical Field

[0001] This invention relates to the field of text review technology, and in particular to a method and system for reviewing text content. Background Technology

[0002] Systematic content review of texts can effectively identify their completeness, compliance, and logical consistency. Machine review, through structured rule parsing, intelligent semantic matching, and automated scoring, significantly improves the efficiency and consistency of text review.

[0003] Existing text content review methods, while no longer requiring human identification, still struggle to accurately understand review standards, leading to frequent misjudgments. For instance, when review rules require "providing a project implementation plan," but the text uses semantically similar but literally different expressions such as "project execution plan," "work deployment arrangement," or "implementation plan," the system often fails to associate these with the keywords in the rule due to its lack of ability to recognize synonyms and term variations. This demonstrates that this method, lacking an understanding of semantic context, results in inaccurate review outcomes, and the review conclusions fail to accurately reflect the true compliance status of the text content. Ultimately, the entire intelligent review process degenerates into a superficial keyword matching mechanism, reducing the reliability and accuracy of the review results. Summary of the Invention

[0004] This invention provides a method and system for reviewing text content to solve the technical problem of low reliability in existing review methods, thereby improving the reliability and accuracy of review results.

[0005] To address the aforementioned technical problems, this invention provides a method and system for reviewing text content, the method comprising: Obtain the text data to be reviewed and the review rule base; Extract each semantic boundary from the review rule base; The review rule base is processed based on each of the semantic boundaries to obtain each first text segment, and the first semantic tag corresponding to each first text segment is determined. Perform association analysis on each of the first semantic tags, and construct an association review graph based on the association analysis results; The text data to be reviewed is processed based on each of the semantic boundaries to obtain each second text segment, and the second semantic tag corresponding to each second text segment is determined. Based on the second semantic tag, the association review graph, and the association analysis results, the target text segment corresponding to each second semantic tag is obtained from each of the first text segments; Based on each target text segment, the corresponding second text segment is analyzed and processed to determine the local review result corresponding to each second text segment. The local review results are then integrated to obtain the review result of the text data to be reviewed.

[0006] Preferably, the step of extracting each semantic boundary from the review rule base includes: The review rule base is subjected to natural language parsing to obtain a set of semantic fragments; The semantic fragment set is embedded to generate a semantic vector; The semantic vectors are clustered to obtain the clustering results; Based on the semantic fragments of the cluster boundaries in the clustering results, the semantic switching points are obtained; Based on the semantic switching point and the clustering result, a semantic coherence scoring function is constructed to obtain the local semantic consistency score of the cluster boundary, and the semantic boundary is obtained based on the local semantic consistency score.

[0007] Preferably, the step of processing the review rule base based on each of the semantic boundaries to obtain each first text segment, and determining the first semantic tag corresponding to each first text segment, includes: Based on the semantic boundaries, the review rule base is divided to obtain each of the first text segments; Perform preliminary context association on each of the first text segments to generate text units; Multi-granular semantic feature extraction is performed on each of the first text segments to obtain the corresponding feature vector; Each feature vector is processed to obtain the corresponding first semantic label.

[0008] Preferably, the step of performing association analysis on each of the first semantic tags and constructing an association review graph based on the association analysis results includes: Each of the first semantic tags is vectorized, and the semantic similarity between any two first semantic tags is calculated to obtain a tag semantic similarity matrix; Based on the semantic similarity matrix of the tags, the association type and association weight of the associated tag pairs are determined; The association review graph is obtained based on each of the first semantic tags, the association type, and the association weight.

[0009] Preferably, the step of analyzing and processing the corresponding second text segment based on each target text segment to determine the local review result corresponding to each second text segment includes: Based on each target text segment and its corresponding second text segment, a set of text pairs is obtained; For each of the target text segments in the text pair set, perform structured parsing to obtain a rule semantic template; The logical constraints in the rule semantic template are mapped to the semantic structure of the second text segment for processing to obtain the local review result.

[0010] Another aspect of the present invention provides a text content review system, comprising: The acquisition module is used to acquire the text data to be reviewed and the review rule base. The extraction module is used to extract various semantic boundaries from the review rule base; The processing module is used to process the review rule base based on each of the semantic boundaries to obtain each first text segment and determine the first semantic tag corresponding to each first text segment. The association module is used to perform association analysis on each of the first semantic tags and construct an association review graph based on the association analysis results. The second module is used to process the text data to be reviewed based on each of the semantic boundaries to obtain each second text segment and determine the second semantic tag corresponding to each second text segment. The target text module is used to filter the target text segment corresponding to each second semantic tag from each of the first text segments based on the second semantic tag, the association review graph, and the association analysis results; The review module is used to analyze and process the corresponding second text segment based on each target text segment, determine the local review result corresponding to each second text segment, integrate the local review results, and obtain the review result of the text data to be reviewed.

[0011] Preferably, the extraction module includes: The parsing unit is used to perform natural language parsing processing on the review rule base to obtain a set of semantic fragments; An embedding unit is used to embed the set of semantic segments to generate a semantic vector; Clustering unit, used to perform clustering processing on the semantic vector to obtain clustering results; Cluster boundary unit, used to obtain semantic switching points based on semantic fragments of cluster boundaries in the clustering results; A semantic boundary unit is used to construct a semantic coherence scoring function based on the semantic switching point and the clustering result, obtain a local semantic consistency score of the clustering boundary, and obtain the semantic boundary based on the local semantic consistency score.

[0012] Preferably, the processing module includes: A segmentation unit is used to segment the review rule base based on the semantic boundaries to obtain each of the first text segments; A text unit is used to perform preliminary context association on each of the first text segments to generate a text unit; The feature unit is used to extract multi-granularity semantic features from each of the first text segments to obtain the corresponding feature vector. The first semantic unit is used to process each of the feature vectors to obtain the corresponding first semantic label.

[0013] Preferably, the associated module includes: The similarity unit is used to vectorize each of the first semantic tags, calculate the semantic similarity between any two first semantic tags, and obtain the tag semantic similarity matrix; The determining unit is used to determine the association type and association weight of the associated tag pairs based on the tag semantic similarity matrix. The graph unit is used to obtain the association review graph based on each of the first semantic tags, the association type, and the association weight.

[0014] Preferably, the step of analyzing and processing the corresponding second text segment based on each target text segment to determine the local review result corresponding to each second text segment includes: A text pair unit is used to obtain a set of text pairs based on each target text segment and the corresponding second text segment; The parsing unit is used to perform structured parsing on the target text segment in each of the text pair sets to obtain a rule semantic template; The mapping unit is used to map the logical constraints in the rule semantic template to the semantic structure of the second text segment for processing, so as to obtain the local review result.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: This invention extracts semantic boundaries from a defined review rule base and segments both the review rule base and the text data to be reviewed based on these semantic boundaries, resulting in a first text segment and a second text segment with semantic consistency. This effectively avoids the misjudgment problem caused by traditional methods that rely solely on literal keyword matching without semantic context. Furthermore, by performing association analysis on the first semantic tags to construct an association review graph, the system can understand the semantic relationships between rules and support generalized matching of non-explicit expressions such as synonyms and term variations. For example, when the rule requires "provide a project implementation plan," even if the text uses "implementation plan," the system can still perform generalized matching. Even phrases like "work deployment arrangements" can be used to identify their equivalence through the mapping relationship of semantic tags in the association review graph. Based on this, target text segments are selected from each of the first text segments using the second semantic tag, the association review graph, and the association analysis results, achieving precise alignment of rules and text at the semantic granularity. Finally, each second text segment is analyzed and processed based on the target text segment to determine the corresponding local review result. These local review results are then integrated to obtain the review result of the text data to be reviewed, ensuring that the review conclusion truly reflects the compliance status of the text content. This invention significantly improves the accuracy and semantic understanding of text review, overcomes the limitations of keyword matching mechanisms, and achieves fine-grained, interpretable, and highly consistent intelligent content review. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a text content review method according to one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a text content review system in one embodiment of the present invention; Figure label: Among them, 11. Acquisition module; 12. Extraction module; 13. Processing module; 14. Association module; 15. Second module; 16. Target text module; 17. Review module. Detailed Implementation

[0017] 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. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] One embodiment of the present invention provides a method for reviewing text content. For details, please refer to [link / reference]. Figure 1 , Figure 1 The diagram illustrates a flowchart of a text content review method according to one embodiment of the present invention, including: S1. Obtain the text data to be reviewed and the review rule base; S2. Extract each semantic boundary from the review rule base; S3. Process the review rule base based on each semantic boundary to obtain each first text segment, and determine the first semantic tag corresponding to each first text segment; S4. Perform association analysis on each primary semantic tag and construct an association review graph based on the association analysis results; S5. Process the text data to be reviewed based on each semantic boundary to obtain each second text segment, and determine the second semantic tag corresponding to each second text segment. S6. Based on the second semantic tag, the association review graph, and the association analysis results, the target text segment corresponding to each second semantic tag is obtained from each first text segment; S7. Analyze and process the corresponding second text segment based on each target text segment, determine the local review result corresponding to each second text segment, integrate the local review results, and obtain the review result of the text data to be reviewed.

[0022] First, the text data to be reviewed and the review rule base are acquired. Acquiring these is a fundamental prerequisite for the text review process. The text data to be reviewed refers to various types of text information that require compliance and security verification, such as user-posted comments, platform-generated content, and documents submitted by third parties. The review rule base contains the complete set of rules required for text review, covering specific content such as prohibited semantic standards, compliance judgment criteria, and semantic boundary definitions. This is done to provide processing objects and judgment benchmarks for subsequent review work. Specifically, this can be achieved by combining data interface calls and local file reading. The text data to be reviewed is obtained from the business system through an API (Application Programming Interface) to ensure data real-time performance and integrity. The review rule base is read using database queries and file parsing techniques. The rule base file is parsed line by line and structured, converting unstructured rule text into processable structured data. Simultaneously, the acquired text data to be reviewed and the review rule base are deduplicated and noise-reduced to remove invalid data and redundant information, ensuring data quality. This ensures that subsequent review processes have clearly defined targets and unified judgment standards, thereby improving review efficiency and accuracy and laying a solid foundation for subsequent semantic boundary extraction, text segment processing, and review result determination.

[0023] Preferably, semantic boundaries are extracted from the review rule base. Natural language parsing is performed on the review rule base to obtain a set of semantic fragments; the set of semantic fragments is embedded to generate semantic vectors; the semantic vectors are clustered to obtain clustering results; semantic switching points are obtained based on the semantic fragments at the cluster boundaries in the clustering results; a semantic coherence scoring function is constructed based on the semantic switching points and the clustering results to obtain the local semantic consistency score of the cluster boundaries, and the semantic boundaries are obtained based on the local semantic consistency score.

[0024] Extracting semantic boundaries from the review rule base is crucial for clarifying the boundaries between different semantic categories within the review rule base. Semantic boundaries refer to the distinguishing nodes between different semantic categories in the review rule base, used to define the semantic scope of different review rules. This is done to accurately split the review rule base and the text data to be reviewed, ensuring the review is targeted.

[0025] First, the review rule base is processed using natural language parsing to obtain a set of semantic segments. Natural Language Processing (NLP) techniques, combined with word segmentation tools, are used to segment the entire review rule base. After removing stop words, the segments are split according to semantic relevance to obtain independent semantic segments, forming a semantic segment set. Next, the semantic segment set is embedded to generate semantic vectors. Embedding models such as GloVe (Global Vectors for Word Representation) are used to convert each semantic segment into a fixed-dimensional real-valued vector. The vector dimension can be set to 128 or 256 dimensions to ensure accurate mapping of semantic information. Finally, the semantic vectors are clustered to obtain clustering results using DBSCAN (Density-Based Spatial Clustering of Applications with...). Noise (density-based spatial clustering with noise) algorithm determines the optimal number of clusters using the elbow rule and silhouette coefficient. Semantic vectors with similar meanings are grouped into one cluster, resulting in multiple clusters. Then, semantic switching points are identified based on semantic segments at the cluster boundaries. These segments represent weak semantic connections between different clusters. Cosine similarity is calculated between semantic vectors in adjacent clusters, and segments with similarity below a preset threshold (e.g., 0.3) are designated as semantic switching points. A semantic coherence scoring function is then constructed based on these switching points and the clustering results, using a combination of cosine similarity and semantic overlap. Inputting the semantic vectors corresponding to the switching points and the semantic vectors within the clusters, a local semantic consistency score is calculated for the cluster boundaries. Finally, cluster boundaries with scores below a preset threshold are defined as semantic boundaries. This approach accurately delineates the semantic categories of the review rule base, providing a precise basis for subsequent text segmentation and semantic tagging, thus improving the accuracy and efficiency of the review process.

[0026] Furthermore, the review rule base is processed based on various semantic boundaries to obtain various first text segments, and the first semantic label corresponding to each first text segment is determined. Based on semantic boundaries, the review rule base is divided to obtain various first text segments; preliminary context association is performed on each first text segment to generate text units; multi-granular semantic features are extracted from each first text segment to obtain corresponding feature vectors; each feature vector is processed to obtain corresponding first semantic labels. The first text segment refers to a rule fragment with independent and complete semantics obtained after splitting the review rule base according to semantic boundaries, such as rule text containing a single prohibition type judgment standard and rule text describing a single compliance requirement. The first semantic label refers to a label used to identify the core semantics of the first text segment, such as a prohibited word label, a compliance standard label, and a risk level label. This is done to decompose the chaotic review rule base into individually identifiable semantic units, facilitating subsequent association analysis and review matching.

[0027] First, based on the extracted semantic boundaries, the review rule base is divided using a boundary location and segmentation algorithm. This algorithm traverses the entire review rule base to locate the start and end positions of each semantic boundary. Using these boundaries as segmentation points, the review rule base is divided into multiple continuous and semantically independent segments, each of which is a first text segment. During the segmentation process, it is crucial to ensure that adjacent first text segments do not have semantic overlap or content omissions. Next, preliminary context association is performed on each first text segment using a context semantic concatenation algorithm. This algorithm extracts the semantic boundary information of each preceding and following first text segment and semantically concatenates this boundary information with the first text segment itself, generating structurally complete and semantically clear text units, avoiding semantic ambiguity in single text segments. Then, multi-granularity semantic feature extraction is performed on each first text segment to obtain corresponding feature vectors. This multi-granularity semantic feature extraction considers three levels: word-level, phrase-level, and sentence-level. At the word level, the term frequency-inverse document frequency algorithm is used to extract keyword features; at the phrase level… Algorithms are used to extract 2-gram and 3-gram phrase features, and sentence-level semantic features are extracted using a model. The features from these three levels are then fused and combined into a unified feature vector using feature concatenation technology. The vector dimension can be set to 256 dimensions to ensure feature comprehensiveness. Finally, each feature vector is processed to obtain the corresponding first semantic label. A classification algorithm combined with a pre-set label library is used. The classification algorithm can be support vector machine, random forest, or neural network. The feature vector is input into the trained classification model, and the model matches the corresponding label from the pre-set label library based on the semantic features of the feature vector, outputting the first semantic label for each first text segment. The pre-set label library needs to be constructed in advance based on the core semantics of the review rule base, covering all possible rule semantic types. The advantage of this is that it enables the structured decomposition and semantic labeling of the review rule base, improving the efficiency and accuracy of subsequent association analysis and review matching, and laying the foundation for building an association review graph.

[0028] Next, association analysis is performed on each primary semantic tag, and an association review graph is constructed based on the association analysis results. Each primary semantic tag is vectorized, and the semantic similarity between any two primary semantic tags is calculated to obtain a tag semantic similarity matrix. Based on the tag semantic similarity matrix, the association type and association weight of the associated tag pairs are determined. Based on each primary semantic tag, association type, and association weight, the association review graph is obtained.

[0029] Association analysis is performed on each primary semantic tag, and an association review graph is constructed based on the association analysis results. The association review graph is a semantic association network formed by primary semantic tags as nodes, the association relationships between tags as edges, and the edges as association weights. The tag semantic similarity matrix is ​​a two-dimensional numerical matrix that records the degree of semantic similarity between any two primary semantic tags. The associated tag pair is two primary semantic tags whose semantic similarity reaches a preset threshold. The association type refers to the semantic association attributes between tags, such as causal association, subordinate association, and parallel association. The association weight is a numerical value that quantifies the degree of association between tags. This is done in order to explore the inherent association patterns of semantic tags in the review rule base and form a structured rule association system.

[0030] First, each primary semantic label is vectorized. Each primary semantic label is input into a pre-trained model, which outputs a fixed-dimensional label semantic vector. The vector dimension can be set to 128 or 256 dimensions to ensure accurate mapping of label semantic information. Next, the semantic similarity between any two primary semantic labels is calculated to obtain a label semantic similarity matrix. A combination of cosine similarity algorithm and Pearson correlation coefficient method is used to calculate pairwise similarity for all label semantic vectors. The cosine similarity algorithm measures semantic similarity in the vector space, while the Pearson correlation coefficient method corrects for linear correlation bias in the vectors. The calculated similarity values ​​are filled into a two-dimensional matrix according to the label order. The rows and columns of the matrix correspond to the first semantic label, and each value in the matrix represents the semantic similarity between two corresponding labels. Then, based on the label semantic similarity matrix, the association type and association weight of the associated label pairs are determined. A preset similarity threshold is first set, typically between 0.5 and 0.7. Label pairs with similarity values ​​higher than this threshold are identified as associated label pairs. Next, the association type is determined through semantic rule matching and contextual semantic analysis. Semantic rule matching involves calling a preset association type determination rule base and matching the semantic content of the associated label pairs against the rule base. In this contextual semantic analysis, the semantic association between tags is determined by combining the contextual information of the first text segment corresponding to each tag. Then, normalization and hierarchical weighting are used to determine the association weights. First, the similarity values ​​are normalized to the range of 0 to 1 using Min-Max normalization. Then, different basic weight coefficients are assigned according to the association type: causal association is set to 1.2, subordinate association to 1.0, and parallel association to 0.8. The normalized similarity values ​​are multiplied by the corresponding basic weight coefficients to obtain the final association weights. Finally, an association review diagram is obtained based on the association type and association weight of each first semantic tag. The spectrum is constructed using graph theory, with each first semantic label as an independent node and the association between related label pairs as edges between nodes. The association type is labeled in the edge attributes, and the association weight is used as the edge weight value. The graph is constructed and stored using a graph computing library. At the same time, the graph is indexed to nodes and edges to facilitate subsequent fast retrieval and matching. The advantage of this approach is that it transforms discrete review rule semantic labels into a structured association network, clearly presenting the semantic associations between rules, improving the comprehensiveness and accuracy of matching the text to be reviewed with the review rules, and providing a traceable semantic association path for review reasoning.

[0031] Then, based on each semantic boundary, the text data to be reviewed is processed to obtain each second text segment, and a second semantic tag is determined for each second text segment. The second text segment refers to a text fragment with independent and complete semantics obtained after splitting the text data to be reviewed according to semantic boundaries, such as a single viewpoint in a user comment, a single paragraph in a document, or a single expression in a message. The second semantic tag refers to a tag used to identify the core semantics of the second text segment, following a unified tagging standard with the first semantic tag, such as tags related to prohibited words, compliance-related tags, and tags with no explicit review association. This is done to decompose the text to be reviewed into semantic units that can be precisely matched with the review rules, achieving segment-by-segment review and improving accuracy.

[0032] First, based on the extracted semantic boundaries, the text data to be reviewed is segmented. Using the same boundary location and segmentation algorithm as the one used to split the review rule base, NLP technology is first used to traverse the entire text to be reviewed, identifying and locating text positions semantically consistent with the semantic boundaries in the review rule base. These positions are used as segmentation points to cut the text data into multiple continuous and semantically independent segments, each segment being a second text segment. During the segmentation process, blank segments and invalid character segments are removed to ensure that each second text segment has actual semantic meaning. Next, preliminary context association is performed on each second text segment. A context semantic concatenation algorithm is used to extract key information from adjacent text segments before and after each second text segment, and semantically fuse it with the second text segment itself to generate semantically complete text units, avoiding semantic ambiguity caused by missing context in a single text segment. Then, multi-granularity semantic feature extraction is performed on each second text segment, using the same method as for extracting features from the first text segment, taking into account the word, phrase, and sentence levels. The word level uses the term frequency-inverse document frequency algorithm for extraction. Keyword features are extracted at the keyword level, followed by 2-gram and 3-gram phrase features. At the sentence level, semantic features are extracted using a model. These three levels of features are then fused into a unified feature vector using feature concatenation technology. The vector dimension is set to 256, consistent with the feature vector of the first text segment, to ensure comparability. Finally, each feature vector is processed to obtain a corresponding second semantic label. The same classification model used for labeling the first semantic label is employed, combining support vector machine, random forest, and neural network algorithms. The feature vector of the second text segment is input into the trained classification model, which, combined with a pre-defined label library, matches the corresponding label based on the semantic features of the feature vector and outputs a second semantic label for each second text segment. This ensures that the labeling standards of the second semantic label and the first semantic label are consistent. This approach enables structured decomposition and semantic identification of the text to be reviewed, allowing it to accurately correspond to the first text segment in the review rule library. This provides a foundation for subsequent selection of target text segments and segment-by-segment review, improving the targeting and efficiency of the review process.

[0033] Furthermore, based on the second semantic tag, the association review graph, and the association analysis results, the target text segment corresponding to each second semantic tag is selected from each first text segment. The target text segment refers to the first text segment that has a semantic relationship with the second semantic tag and can be used to determine the review result of the corresponding second text segment. For example, when the second semantic tag is a prohibited word tag, the corresponding target text segment is the first text segment containing the rule for determining that prohibited word. The association analysis results refer to the tag semantic similarity matrix obtained from the previous association analysis of the first semantic tags, including the association type and association weight of the tag pairs. This is done to match each second text segment with a corresponding review rule basis, ensuring that the review is supported by clear rules.

[0034] First, for each second semantic tag, first semantic tags with direct and indirect associations are retrieved from the association review graph. A graph retrieval algorithm, combined with node and edge information from the association review graph, is used via the graph computing library's retrieval interface. The second semantic tag is input as the retrieval node, and the association path length threshold is set to 1 to 2. All first semantic tags connected to this node are retrieved, and the association type and weight corresponding to these first semantic tags are extracted. Then, based on the tag semantic similarity matrix in the association analysis results, the retrieved first semantic tags are filtered, retaining those with semantic similarity higher than a preset threshold (generally 0.5) and removing tags with weak semantic associations. Next, based on the filtered first semantic tags, the corresponding first text segments are matched. Each first semantic tag corresponds to a unique first text segment. All matching first text segments are obtained directly through the tag-text segment association mapping relationship. Then, the obtained first text segments are... A second screening process is performed on each text segment, further optimizing it by combining association type and association weight. First text segments with causal or subordinate association types and association weights higher than 0.8 are prioritized for retention. Parallel associations with lower association weights are eliminated, ensuring the selected first text segments have the strongest correlation with the second semantic tag. Finally, the first text segments after this second screening are determined as the target text segments corresponding to the second semantic tag. If a second semantic tag corresponds to multiple target text segments, they are sorted from highest to lowest association weight to facilitate subsequent priority use of the rule text with the strongest correlation for review. Simultaneously, the association type and association weight of each target text segment are recorded for reference in subsequent analysis and processing. The advantage of this approach is that it accurately matches the review rules corresponding to each second text segment, avoids interference from irrelevant rules, reduces review redundancy, improves the accuracy and efficiency of review, and provides a reliable rule basis for determining the local review results of the second text segment.

[0035] Finally, based on each target text segment, the corresponding second text segment is analyzed and processed to determine the local review result for each second text segment. These local review results are then integrated to obtain the review result for the text data to be reviewed. Based on each target text segment and its corresponding second text segment, a set of text pairs is obtained. The target text segments in each text pair set are then subjected to structured parsing to obtain a rule semantic template. The logical constraints in the rule semantic template are mapped to the semantic structure of the second text segment for processing to obtain the local review result.

[0036] Based on each target text segment, the corresponding second text segment is analyzed and processed to determine the local review result for each second text segment. The local review results are then integrated to obtain the review result of the text data to be reviewed. The local review result refers to the compliant, non-compliant, or pending review result of a single second text segment based on the review rules of the corresponding target text segment. The text pair set refers to the one-to-one correspondence of a single second text segment and all its corresponding target text segments. The rule semantic template refers to the standardized rule model containing review logic constraints, judgment standards, and semantic elements extracted after the target text segment is structured and parsed. The logic constraints refer to the review judgment conditions clearly defined in the rule semantic template, such as prohibited content matching conditions, compliance element missing conditions, and semantic threshold limitation conditions. This is done in order to accurately judge each segment of the split text to be reviewed based on the review rules, and then achieve a comprehensive review of the entire text through result integration.

[0037] First, based on each target text segment and its corresponding second text segment, a text pair set is constructed according to the principle of one-to-one matching. One second text segment corresponds to one text pair set, which contains the second text segment and all the target text segments obtained from its selection. A unique identifier is added to each text pair set to ensure traceability in subsequent processing. Next, the target text segments in each text pair set are subjected to structured parsing to obtain rule semantic templates. Using NLP technology combined with rule parsing algorithms, syntactic analysis and semantic extraction are first performed on the target text segments. Dependency parsing is used to determine the subject-verb-object structure and logical connectors in the rules, extracting core semantic elements such as the review subject, judgment action, and judgment threshold. Finally, the rules are processed according to the preset rule template. The framework standardizes and fills in the extracted semantic elements and logical relationships. It includes four core modules: review type, judgment conditions, violation results, and compliance standards. This ultimately generates a structured rule semantic template, ensuring clear logic and complete elements. Then, the logical constraints in the rule semantic template are mapped to the semantic structure of the second text segment for processing, yielding partial review results. First, a semantic structure parsing algorithm is used to process the second text segment, extracting its semantic elements and structural relationships and transforming them into a semantic structure consistent with the rule semantic template. Then, a semantic matching algorithm is used to match each logical constraint in the rule semantic template with the semantic structure of the second text segment one by one. The matching process combines precise matching and fuzzy matching. The method employs precise matching for explicit elements such as prohibited words, and fuzzy matching using the cosine similarity algorithm for ambiguous elements such as semantic descriptions, with a similarity threshold set to 0.7. After matching, a partial review result is derived based on the matching results and the judgment criteria of the rule semantic template. If all logical constraints match, it is judged as a violation; if none match, it is judged as compliant; if there is a partial match or the matching result is questionable, it is judged as pending review. Finally, the partial review results corresponding to all second text segments are integrated to obtain the review result of the text data to be reviewed. A result fusion algorithm is used: first, all partial review results are statistically analyzed; if any second text segment is judged as a violation, the entire text is initially judged as a violation; if all second text segments are judged as violations, the entire text is initially judged as a violation. If all sections are deemed compliant, the entire text is deemed compliant. If there are pending review results and no violations are found, the entire text is deemed pending review. Then, the text initially deemed non-compliant undergoes a second verification, checking the corresponding target text sections and matching process. After confirming there are no matching errors, the final overall review result is determined, and a review result report is generated, including the review results of each section, the overall judgment conclusion, the location of the violation, and the rule basis. The advantage of this approach is that it enables precise segment-by-segment judgment of the text to be reviewed, ensuring the meticulousness and accuracy of the review results. By integrating the results, it takes into account both the partial judgments and the overall conclusion, avoiding review errors caused by partial omissions. At the same time, it completely preserves the review basis and process, improving the traceability and credibility of the review results.

[0038] It is important to note that while the aforementioned text content review methods can effectively achieve refined rule matching based on semantic boundaries and the integration of local review results, they may still be limited by the static coverage and semantic generalization capabilities of the rule base when dealing with highly specialized, terminology-dense, and logically complex engineering texts (such as bidding documents). Without deep modeling and dynamic reasoning mechanisms for domain knowledge, relying solely on preset semantic boundaries and tag mappings makes it difficult to accurately capture the implicit connections and compliance requirements between technical clauses. Therefore, this invention, based on a general text review framework, further provides a two-stage intelligent bidding method that deeply integrates structured retrieval and semantic enhancement generation, specifically addressing the unique needs of professional bidding scenarios. This method, through the construction of a multi-level knowledge graph-driven tag expansion mechanism and vector enhancement strategy, not only compensates for the shortcomings of traditional rule bases in terms of knowledge breadth and semantic depth but also provides a scalable technical path for high-precision and high-efficiency intelligent review. The following section will focus on how this evaluation method, through the collaborative architecture of initial screening using SQL (Structured Query Language) and fine-tuning using RAG (Retrieval-Augmented Generation), as well as the introduction of a knowledge graph for evaluation-assisted decision-making, achieves accurate compliance judgment on complex engineering bids, thereby ensuring the organic unity between the generality and professionalism of the overall technical solution and the feasibility of engineering implementation.

[0039] Another embodiment of this invention proposes a two-stage intelligent bidding evaluation method and system that combines SQL structured retrieval and RAG semantic retrieval for professional bidding review scenarios. This solution aims to address the problems of retrieval omissions, judgment biases, and response delays caused by incomplete knowledge coverage and insufficient semantic understanding in traditional intelligent bidding evaluation systems when faced with highly specialized, terminology-dense, diverse, and implicitly complex engineering bids. Its core innovation lies in: constructing a multi-level knowledge graph-driven tag expansion mechanism to achieve a leap from surface keyword matching to deep semantic association reasoning; and designing an efficient two-stage retrieval architecture that balances retrieval breadth, accuracy, and speed.

[0040] I. Overall System Architecture and Data Flow The system operates in a typical distributed environment consisting of user terminals, a communication network, and a server. Users submit two types of key inputs via terminals such as smartphones and PCs: review content, such as "Does the power distribution system comply with Standard A?", and the tender documents to be reviewed, i.e., the supplier's bid documents. Upon receiving the review request, the server initiates a multi-module collaborative processing flow: Tag extraction: A pre-trained tagging model, trained on a large number of manually annotated knowledge slices and tag samples, automatically identifies corresponding review tags, such as "power distribution system," from the review content. The model's output can be reviewed and adjusted by experts to ensure tag accuracy.

[0041] Knowledge graph-driven tag expansion: Using review tags as the starting point for queries, associated entities are mined within the tag-related knowledge graph to obtain a set of expanded tags such as "10kV transformer," "circuit breaker," and "current transformer." This process is a key preliminary step in this solution, used to compensate for knowledge gaps that may result from relying solely on the original tags.

[0042] Phase 1: SQL Structured Initial Screening: Based on the tag set formed by the combination of review tags and extended tags, all related knowledge slices are quickly screened out in the first SQL database to form a filtered knowledge slice set. This database adopts a three-table index structure (tag index table, content index table, tag-content mapping table), supports millisecond-level multi-tag joint queries, which is significantly better than the unstructured storage method of traversing each record one by one.

[0043] Phase 2: RAG Semantic Refinement: The filtered knowledge slice set and the original review content (or its enhanced version) are input into the RAG retrieval model. Through vector similarity calculation, the most relevant target knowledge slice set is further accurately located.

[0044] Intelligent evaluation decision-making: Finally, the evaluation model comprehensively reviews the content, the full text of the bid to be evaluated or its key segments, and the target knowledge segment set to generate structured evaluation results such as "compliant / non-compliant", "deduction item explanation" or "suggested clarification issues", and pushes them to the user.

[0045] II. Detailed Explanation of Core Innovative Components (1) Construction and application of tag-related knowledge graph This graph forms the basis for semantic generalization. Its construction does not rely entirely on manual expert definition, but rather employs a hybrid strategy of data-driven approach combined with expert validation. Data sources: covering four major categories of multi-source heterogeneous data: technical parameters extracted from industrial material bidding documents, unified evaluation standards and enterprise rules, historical winning / unsuccessful bidding cases, and national / industry technical specifications such as GB and DL standards.

[0046] Preprocessing: Data is cleaned through a two-step process of rule filtering and manual verification to remove noise, deduplication, and correct formatting errors.

[0047] Slicing and labeling: The cleaned knowledge base content is divided into semantically complete knowledge slices, and one or more slice labels are assigned to each slice through a labeling model.

[0048] Graph construction logic: First, core tags are selected from all slice tags based on indicators such as high frequency of occurrence and many co-occurrence tags, serving as hub nodes for map construction.

[0049] For each pair of core labels and other labels, semantic similarity and co-occurrence frequency are calculated based on a pre-trained language model.

[0050] If both exceed the threshold, a correlation is determined, and the correlation degree is calculated according to the weighted formula, such as 0.8 indicating a strong correlation and 0.2 indicating a weak correlation.

[0051] The remaining tags that are not covered by the core tags are called supplementary tags, and their relationships are completed using the same logic.

[0052] This method avoids the infeasibility of exhaustively enumerating all pairwise relationships in a massive label space, and realizes dynamic growth and inferential completion of the graph. For example, although "power distribution system" and "Type A transformer" do not co-occur directly, if both co-occur with "10kV power distribution network" at high frequency, an indirect relationship can be established through graph inference.

[0053] During the application phase, the system efficiently obtains extended tags by limiting the number of path hops (e.g., ≤3 hops) and using a correlation aggregation strategy to obtain the maximum path correlation among multiple paths, ensuring that it neither overgeneralizes nor misses key correlations.

[0054] (2) The significance of the two-stage retrieval mechanism The first stage (SQL initial screening): Leveraging the efficient indexing capabilities of structured databases, the candidate set of millions of knowledge points is narrowed down to thousands or even hundreds within milliseconds. The key to this step is expanding the initial screening scope through tag extension, preventing relevant knowledge from being prematurely excluded due to tag annotation bias.

[0055] The second stage (RAG fine-tuning): In the narrowed high-quality candidate set, the RAG model can focus more on deep semantic matching, avoiding wasting computational resources on irrelevant noise, thereby improving retrieval accuracy while reducing latency.

[0056] This strategy of prioritizing broad search followed by refined search perfectly balances recall and efficiency, which is the core advantage of this solution that distinguishes it from pure vector search or pure keyword search.

[0057] (3) Knowledge graph and extended evaluation vector for bid evaluation and decision support To further enhance RAG's semantic understanding capabilities, the system introduces a second knowledge graph—the evaluation-assisted decision-making knowledge graph. This graph stores domain knowledge in a structured manner using triples (<entity, relation, entity>), for example: <110kV environmentally friendly oil-immersed transformer, involving no-load loss> The DL / T 2552-2024 standard includes no-load loss <6.4 kW> When processing the review content "Whether the 110 kV environmentally friendly oil-immersed transformer of Supplier X complies with the DL / T 2552-2024 standard", the system not only uses the original text but also: Extract technical indicator entities related to "110 kV transformer" from the knowledge graph; Extract specific clause entities related to "DL / T 2552-2024"; Fuse these entities, their attributes, and relationships into extended information; Finally, generate an extended review vector, which can be vectorized as a whole by splicing the original text and the extended information, or perform weighted fusion of the original vector and the extended vector, giving higher weight to the original content to retain the core intention.

[0058] This vector enhancement mechanism enables the RAG model to "understand" the complete technical context behind the review during retrieval, greatly enhancing the ability to identify implicit compliance requirements.

[0059] III. System Extensibility and Engineering Implementation Processing of the tender itself: For large tenders, the system can also slice, label, and store them in the second SQL database, execute the same two-stage retrieval process, and obtain the target tender slice set, enabling the bid evaluation model to cross-compare external standards with the tender responses.

[0060] Modular design: The system is divided into modules such as construction, marking, acquisition, screening, and bid evaluation, facilitating independent development, testing, and deployment.

[0061] Software and hardware compatibility: It can be implemented entirely or partially through software computer program products, hardware, or firmware, and supports running on general servers, clusters, or dedicated devices.

[0062] Security and communication: Support encryption protocols to ensure the security of review data during transmission and storage; the communication network can be wired Ethernet, USB, or wireless cellular, Wi-Fi, Bluetooth.

[0063] In summary, the embodiments of the present invention not only provide a complete set of intelligent bid evaluation technical solutions, but also construct a next-generation intelligent bid evaluation infrastructure with professional depth, semantic breadth, and engineering efficiency through three pillars: knowledge graph-driven label reasoning, organic integration of structured and unstructured retrieval, and vector enhanced expression of multi-source knowledge. Its design concept can be widely applied to the tendering and bidding fields with strict technical compliance requirements such as power, transportation, construction, and medical equipment, and has significant practical value and promotion potential.

[0064] This invention stores knowledge slices and their corresponding slice tags in the form of an SQL database. That is, the database can store multiple index relationship tables that can improve subsequent retrieval speed, including, but not limited to, the following three index relationship tables (Note: the data in the tables are for illustrative purposes only): Table 1. Slice Label Index Relationship Table Table 2 Knowledge Slice Index Relationship Table Table 3 Relationship between Tag Index and Content Index Understandably, Tables 1, 2, and 3 together constitute an efficient multi-level indexing mechanism: Table 1 records the mapping relationship between slice tags and tag index values, Table 2 stores the correspondence between the specific content of a knowledge slice and its content index value, and Table 3 establishes the association between tag index values ​​and content index values. When it is necessary to find all knowledge slices related to a certain slice tag, the system first obtains its corresponding tag index value through Table 1, then retrieves all associated content index values ​​through Table 3, and finally uses Table 2 to restore these index values ​​to the specific knowledge slice content. This structured index design avoids the inefficient operation of traversing all data rows in the traditional single-table storage mode, significantly improving the speed of the first-stage screening of the knowledge slice set, thus effectively supporting the overall performance and feasibility of the two-stage retrieval architecture of SQL initial screening and RAG fine-tuning.

[0065] This invention proposes a two-stage retrieval mechanism combining SQL structured retrieval and RAG semantic retrieval. First, knowledge slices and their tags are efficiently indexed and stored in an SQL database, achieving rapid initial screening. Then, the selected subset of knowledge is passed to the RAG model for refined semantic matching in the second stage, significantly reducing the scope and computational cost of RAG retrieval and improving overall response speed. Crucially, the system introduces a tag-associated knowledge graph to intelligently expand the original review tags. Even if some knowledge slices do not explicitly contain keywords such as "power distribution system" in the text, if their content, such as mentioning "Type A transformer," is highly relevant to the review topic, their potential semantic association with the review tags can be identified through graph reasoning, and they can be included in the initial screening. This recall strategy based on expanded tags effectively avoids the omission of relevant knowledge due to annotation limitations, ensuring the sufficiency and accuracy of subsequent RAG retrieval. Meanwhile, this knowledge graph does not need to exhaustively enumerate all the relationships between tags in advance. Instead, it relies on existing core association data and dynamically infers potential relationships that are not explicitly labeled through indicators such as co-occurrence frequency and semantic similarity. This enables the self-growth and continuous completion of the association network, which greatly reduces the complexity of the early knowledge construction and provides solid support for the efficient and robust operation of the entire two-stage retrieval architecture.

[0066] Another embodiment of the present invention provides a text content review system; for details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates the structure of a text content review system according to one embodiment of the present invention, comprising: Module 11 is used to acquire the text data to be reviewed and the review rule base; Extraction module 12 is used to extract various semantic boundaries from the review rule base; The processing module 13 is used to process the review rule base based on each semantic boundary to obtain each first text segment and determine the first semantic tag corresponding to each first text segment. The association module 14 is used to perform association analysis on each first semantic tag and construct an association review graph based on the association analysis results; The second module 15 is used to process the text data to be reviewed based on each semantic boundary to obtain each second text segment and determine the second semantic tag corresponding to each second text segment. The target text module 16 is used to filter the target text segment corresponding to each second semantic tag from each first text segment based on the second semantic tag, the association review graph and the association analysis results; The review module 17 is used to analyze and process the corresponding second text segment based on each target text segment, determine the local review result corresponding to each second text segment, integrate the local review results, and obtain the review result of the text data to be reviewed.

[0067] Preferably, the extraction module 12 includes: The parsing unit is used to perform natural language parsing on the review rule base to obtain a set of semantic fragments; Embedding units are used to embed a set of semantic fragments to generate semantic vectors; Clustering units are used to cluster semantic vectors to obtain clustering results; Cluster boundary units are used to obtain semantic switching points based on semantic fragments of cluster boundaries in the clustering results; The semantic boundary unit is used to construct a semantic coherence scoring function based on semantic switching points and clustering results, obtain the local semantic consistency score of the cluster boundary, and obtain the semantic boundary based on the local semantic consistency score.

[0068] Preferably, the processing module 13 includes: The segmentation unit is used to divide the review rule base based on semantic boundaries to obtain each first text segment; Text units are used to perform preliminary contextual association on each first text segment to generate text units; The feature unit is used to extract multi-granular semantic features from each first text segment to obtain the corresponding feature vector. The first semantic unit is used to process each feature vector to obtain the corresponding first semantic label.

[0069] Preferably, the associated module 14 includes: Similarity units are used to vectorize each first semantic label, calculate the semantic similarity between any two first semantic labels, and obtain the label semantic similarity matrix; The unit is used to determine the association type and association weight of associated tag pairs based on the tag semantic similarity matrix. The graph unit is used to obtain the association review graph based on each first semantic label, association type, and association weight.

[0070] Preferably, the corresponding second text segment is analyzed and processed based on each target text segment to determine the local review result corresponding to each second text segment, including: A text pair unit is used to obtain a set of text pairs based on each target text segment and its corresponding second text segment; The parsing unit is used to perform structured parsing of the target text segments in each text pair set to obtain the rule semantic template; The mapping unit is used to map the logical constraints in the rule semantic template to the semantic structure of the second text segment for processing, so as to obtain the local review result.

[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0072] Accordingly, embodiments of the present invention provide a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform steps in the text content review method of the above embodiments, for example... Figure 1 Steps S1 to S7 as described above.

[0073] This invention addresses the fundamental problem of inaccurate matching in existing text review methods due to a lack of contextual understanding, starting from the perspective of semantic structure alignment. Unlike traditional methods that perform coarse-grained comparisons between rules and the entire text, this invention first synchronously segments the review rule base and the text data to be reviewed into first and second text segments based on a unified semantic boundary, ensuring that rules and text interact at the same semantic granularity. By assigning first and second semantic tags to each text segment and constructing a correlation review graph using the results of association analysis, the system can capture the logical relationships and semantic generalization paths between rules, thereby effectively identifying the semantic equivalence behind differences in expressions such as "implementation plan" and "project implementation plan." On this basis, target text segments that are truly relevant to the content to be reviewed are selected based on the second semantic tags, the correlation review graph, and the results of association analysis, avoiding interference from irrelevant rules. Finally, by performing fine-grained analysis on each second text segment and its corresponding target text segment to generate local review results, these are integrated into an overall conclusion, achieving a paradigm shift from "surface matching" to "semantic alignment." This invention not only improves the tolerance and understanding of synonyms, term variations, and unstructured expressions, but also enhances the logical consistency, traceability, and overall reliability of the review process.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for reviewing text content, characterized in that, include: Obtain the text data to be reviewed and the review rule base; Extract each semantic boundary from the review rule base; The review rule base is processed based on each of the semantic boundaries to obtain each first text segment, and the first semantic tag corresponding to each first text segment is determined. Perform association analysis on each of the first semantic tags, and construct an association review graph based on the association analysis results; The text data to be reviewed is processed based on each of the semantic boundaries to obtain each second text segment, and the second semantic tag corresponding to each second text segment is determined. Based on the second semantic tag, the association review graph, and the association analysis results, the target text segment corresponding to each second semantic tag is obtained from each of the first text segments; Based on each target text segment, the corresponding second text segment is analyzed and processed to determine the local review result corresponding to each second text segment. The local review results are then integrated to obtain the review result of the text data to be reviewed.

2. The text content review method as described in claim 1, characterized in that, The extraction of semantic boundaries from the review rule base includes: The review rule base is subjected to natural language parsing to obtain a set of semantic fragments; The semantic fragment set is embedded to generate a semantic vector; The semantic vectors are clustered to obtain the clustering results; Based on the semantic fragments of the cluster boundaries in the clustering results, the semantic switching points are obtained; Based on the semantic switching point and the clustering result, a semantic coherence scoring function is constructed to obtain the local semantic consistency score of the cluster boundary, and the semantic boundary is obtained based on the local semantic consistency score.

3. The text content review method as described in claim 1, characterized in that, The process of processing the review rule base based on each of the semantic boundaries to obtain each first text segment, and determining the first semantic tag corresponding to each first text segment, includes: Based on the semantic boundaries, the review rule base is divided to obtain each of the first text segments; Perform preliminary context association on each of the first text segments to generate text units; Multi-granular semantic feature extraction is performed on each of the first text segments to obtain the corresponding feature vector; Each feature vector is processed to obtain the corresponding first semantic label.

4. The text content review method as described in claim 1, characterized in that, The step of performing association analysis on each of the first semantic tags and constructing an association review graph based on the association analysis results includes: Each of the first semantic tags is vectorized, and the semantic similarity between any two first semantic tags is calculated to obtain a tag semantic similarity matrix; Based on the semantic similarity matrix of the tags, the association type and association weight of the associated tag pairs are determined; The association review graph is obtained based on each of the first semantic tags, the association type, and the association weight.

5. The text content review method as described in claim 1, characterized in that, The step of analyzing and processing the corresponding second text segment based on each target text segment to determine the local review result corresponding to each second text segment includes: Based on each target text segment and its corresponding second text segment, a set of text pairs is obtained; For each of the target text segments in the text pair set, perform structured parsing to obtain a rule semantic template; The logical constraints in the rule semantic template are mapped to the semantic structure of the second text segment for processing to obtain the local review result.

6. A text content review system, characterized in that, include: The acquisition module is used to acquire the text data to be reviewed and the review rule base. The extraction module is used to extract various semantic boundaries from the review rule base; The processing module is used to process the review rule base based on each of the semantic boundaries to obtain each first text segment and determine the first semantic tag corresponding to each first text segment. The association module is used to perform association analysis on each of the first semantic tags and construct an association review graph based on the association analysis results. The second module is used to process the text data to be reviewed based on each of the semantic boundaries to obtain each second text segment and determine the second semantic tag corresponding to each second text segment. The target text module is used to filter the target text segment corresponding to each second semantic tag from each of the first text segments based on the second semantic tag, the association review graph, and the association analysis results; The review module is used to analyze and process the corresponding second text segment based on each target text segment, determine the local review result corresponding to each second text segment, integrate the local review results, and obtain the review result of the text data to be reviewed.

7. The text content review system as described in claim 6, characterized in that, The extraction module includes: The parsing unit is used to perform natural language parsing processing on the review rule base to obtain a set of semantic fragments; An embedding unit is used to embed the set of semantic segments to generate a semantic vector; Clustering unit, used to perform clustering processing on the semantic vector to obtain clustering results; Cluster boundary unit, used to obtain semantic switching points based on semantic fragments of cluster boundaries in the clustering results; A semantic boundary unit is used to construct a semantic coherence scoring function based on the semantic switching point and the clustering result, obtain a local semantic consistency score of the clustering boundary, and obtain the semantic boundary based on the local semantic consistency score.

8. The text content review system as described in claim 6, characterized in that, The processing module includes: A segmentation unit is used to segment the review rule base based on the semantic boundaries to obtain each of the first text segments; A text unit is used to perform preliminary context association on each of the first text segments to generate a text unit; The feature unit is used to extract multi-granularity semantic features from each of the first text segments to obtain the corresponding feature vector. The first semantic unit is used to process each of the feature vectors to obtain the corresponding first semantic label.

9. The text content review system as described in claim 6, characterized in that, The associated module includes: The similarity unit is used to vectorize each of the first semantic tags, calculate the semantic similarity between any two first semantic tags, and obtain the tag semantic similarity matrix; The determining unit is used to determine the association type and association weight of the associated tag pairs based on the tag semantic similarity matrix. The graph unit is used to obtain the association review graph based on each of the first semantic tags, the association type, and the association weight.

10. The text content review system as described in claim 6, characterized in that, The step of analyzing and processing the corresponding second text segment based on each target text segment to determine the local review result corresponding to each second text segment includes: A text pair unit is used to obtain a set of text pairs based on each target text segment and the corresponding second text segment; The parsing unit is used to perform structured parsing on the target text segment in each of the text pair sets to obtain a rule semantic template; The mapping unit is used to map the logical constraints in the rule semantic template to the semantic structure of the second text segment for processing, so as to obtain the local review result.

Images