Office document automatic proofreading and management method and system based on text summary algorithm

Abstract

This invention discloses an automatic review and management method and system for office documents based on text summarization algorithms, belonging to the field of semantic processing technology. The method includes the following steps: structurally parsing the office documents; screening review-focused elements and generating a candidate semantic fragment set through semantic parsing and merging operations; generating summary units, calculating their importance weights, and determining the review execution strategy accordingly; applying review rules to generate review results; and locally updating and regenerating abnormal summary units based on the results. This invention solves the pain points of low summary credibility and the inability of static rule bases to adapt to new problems. It achieves content traceability and fidelity, and realizes intelligent hierarchical and efficient allocation of review resources. It not only automatically produces highly credible summaries and detailed review reports, but also achieves simultaneous growth in processing accuracy and efficiency, providing standardized, evolvable, and reliable quality assurance for the processing of key office documents such as contracts and reports.

Description

Technical Field

[0001] This invention relates to the field of semantic processing technology, and in particular to a method and system for automatic review and management of office documents based on text summarization algorithms. Background Technology

[0002] Existing automated document review and management systems combine text summarization and intelligent review methods to achieve efficient automated processing and management of office documents. The text summarization technology, based on an Encoder-Decoder model and attention mechanism, extracts key information from documents, addresses out-of-vocabulary words and clause repetition issues, and filters redundant information through a selection network. The generated summaries are corrected by comparing their semantic relevance to the document, thereby improving semantic consistency. For document review, professional domain documents are first vectorized. Different types of prompt tags are designed to guide the large-scale model in review. Flexible combinations of these tags can check for various error types, avoiding the complexity of training independent models for each error type. The semantic understanding capabilities of the large-scale model make the review process simpler and more efficient, and can handle complex professional, knowledge-based, and time-sensitive errors. By incorporating domain knowledge, customized review is performed on documents in specific domains, ensuring document accuracy and professionalism, thereby improving the intelligence and automation level of document management.

[0003] For example, Chinese invention patent CN109145105B discloses a text summarization model generation algorithm that integrates information selection and semantic association. The algorithm includes: first, based on the Encoder-Decoder model, it combines an attention mechanism to obtain sufficient information from the input sequence; then, it uses a copy mechanism and a coverage mechanism to solve the problems of missing words and clause repetition in the generated summary; then, it designs a selection network to perform secondary encoding on the original text to filter redundant information; finally, it corrects the semantics of the summary by comparing the semantic relevance between the original text and the summary.

[0004] For example, Chinese invention patent CN117708273A discloses a method, system, and computer storage medium for intelligent text content review, which includes: collecting documents related to the content to be detected in a professional field, vectorizing them, obtaining a vector model, and storing it; designing Prompt tags for text content error types; and performing large-scale model review of the document to be tested based on the documents related to the content to be detected in a professional field, combined with the Prompt tags, and outputting the review results.

[0005] The above-mentioned technology has at least the following technical problems:

[0006] In intelligent office document review systems based on text summarization, traditional architectures often treat summary generation and document review as independent processes, making it difficult for the two to form effective two-way information interaction and collaborative optimization. The summary generation module struggles to fully anticipate review needs during information compression, resulting in the omission of key contexts. Information integrity defects discovered by the review module are difficult to be fed back to the summary generation stage in real time to trigger optimization. At the same time, the importance perceptions such as attention weights formed during summary generation fail to effectively guide the differentiated allocation of review resources. This fragmentation of processes and one-way information flow leads to limitations in system accuracy due to the lack of contextual references and losses in efficiency due to repeated iterations and resource mismatches. Summary of the Invention

[0007] On the one hand, a method for automatic review and management of office documents based on text summarization algorithms is provided, which includes:

[0008] The document is structured and assigned original text location identifiers. The elements of interest for review are read and a set of elements of interest for review are generated. Semantic parsing is performed based on the original text location identifiers and an initial set of semantic fragments is generated, forming a set of candidate semantic fragments.

[0009] Summarizing units are generated based on candidate semantic fragments. The importance weight of the summarizing units is calculated based on the attention weight distribution. The review and execution strategy for each summarizing unit is determined. During the execution of the review and execution strategy, the review and execution rules are applied to generate the review and execution results.

[0010] Based on the review results, the abstract review results are determined, the abstract unit is partially updated based on the abstract review results, and the final abstract results and document review results are output.

[0011] On the other hand, an automatic review and management system for office documents based on text summarization algorithms is provided. This system includes: an office document parsing module, which is used to perform structural parsing on office documents and assign original text location identifiers, read review-related elements, generate a set of review-related elements, perform semantic parsing based on original text location identifiers and generate an initial set of semantic fragments, and form a set of candidate semantic fragments.

[0012] The office document review and summary generation module is used to generate summary units based on candidate semantic fragments, calculate the importance weight of summary units based on attention weight distribution, determine the review execution strategy for each summary unit, and generate review results by applying review rules during the execution of the review execution strategy.

[0013] The document summary update module is used to determine the summary review result based on the review results, perform partial updates of the summary unit based on the summary review result, and output the final summary result and document review result.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0015] 1. The automatic office document review and management method based on text summarization algorithm provided by this invention improves the intelligence and automation level of office document processing by combining text summarization algorithm and automatic review mechanism. Through precise structural parsing, semantic parsing and candidate semantic fragment generation, it can deeply understand the document content, identify and handle various review issues, such as missing context, semantic inconsistency and conflicting descriptions, to ensure the consistency and accuracy of the summary content with the original text. Through adaptive review strategy, the execution strategy is dynamically adjusted according to the importance of the summary unit and review rules, which optimizes the accuracy and efficiency of review. It can continuously update the summary content in the loop process, repeatedly detect and correct potential problems until the final output of the summary and document review results that meet the requirements, which greatly improves the automation level of document review and provides efficient technical support for quality assurance of complex documents.

[0016] 2. This invention combines structural and semantic analysis of office documents. By accurately identifying chapter levels, paragraph boundaries, and sentence order information in the document, the document content can be accurately mapped to the semantic level after structural processing. This not only lays a solid foundation for the subsequent review process but also ensures the comprehensiveness and accuracy of information extraction. Especially when dealing with complex linguistic phenomena such as polysemous words, referential relationships, and cross-paragraph citations, precise dependency parsing and semantic role labeling avoid misunderstandings caused by unclear context or fragmented information. It can more efficiently generate refined candidate semantic fragments for the text, ensuring a high degree of consistency and accuracy between the review and summary content.

[0017] 3. This invention, through intelligent computing based on attention weight distribution and information entropy analysis, can dynamically evaluate the semantic coverage and concentration of each summary unit, and determine whether it enters the priority review queue according to its importance weight. This makes the summary generation more semantically deep, and can intelligently distinguish which information is more important to the overall office document. Thus, review resources are concentrated on key content, which greatly improves review efficiency and quality, and provides a more flexible and intelligent review execution strategy that can adapt to different documents and review needs.

[0018] 4. This invention achieves continuous refinement of abstract content by introducing a local update and iterative optimization mechanism. In each round of review, abstract units with consistency issues or semantic discrepancies are locally adjusted based on the review results, thereby continuously correcting potential errors or defects in the abstract. Through iterative optimization, the quality of the abstract is gradually improved, ensuring semantic consistency with the original text, avoiding unnecessary repeated reviews and resource waste, and enhancing the intelligence and efficiency of document review. This enables the generation of high-quality abstracts in a short time and minimizes the need for manual intervention. By prioritizing the retention tags of core abstract units, key information is identified as high-risk and high-importance content at the initial generation stage, thus undergoing rigorous review and real-time monitoring. This prevents any small error from evolving into a major quality problem, excessive review, or unnecessary resource waste. Only abstract units with compound defects are reprocessed, allowing for efficient concentration of resources for local updates. This ensures that the iterative optimization process only focuses on units with real problems, without affecting correctly generated parts, thereby improving abstract quality while maintaining review efficiency and system stability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of an automatic proofreading and management method for office documents based on a text digest algorithm, provided for embodiments of this application;

[0021] Figure 2 This is a flowchart illustrating the review strategy method involved in the embodiments of this application;

[0022] Figure 3 This is a diagram illustrating the generative text summarization algorithm architecture involved in the embodiments of this application;

[0023] Figure 4 This is a schematic diagram of the structure of an office document automatic review and management system based on a text digest algorithm, provided in an embodiment of this application. Detailed Implementation

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] Embodiment 1 of the present invention Figure 1The diagram shows a flowchart of an automatic office document review and management method based on a text summarization algorithm provided in this application. The method includes the following steps: performing document structure parsing, identifying chapters, paragraphs, and sentences in the document, assigning original text position identifiers, reading the definition information of review focus elements from the review rule base, filtering relevant elements based on the document business type, generating an initial set of semantic fragments through semantic parsing, further merging them into candidate semantic fragments, generating summary units based on the candidate semantic fragments, calculating the importance weight of the summary units, determining the corresponding review execution strategy, applying review rules to generate review results, performing local updates of summary units based on the review results, and if consistency anomalies are found, regenerating the summary content and updating the mapping relationship. This process continues until no new review results are generated in a consecutive loop, and finally outputting the document's summary result and review result.

[0026] The system reads the original office document file, uses an office document parsing library to read the binary language structure of the document, extracts metadata tags defining paragraph attributes, character formatting, and page layout to parse its underlying format tags. Based on the extracted format tags, it uses regular expressions to match heading level styles, identify numbering sequence patterns, measure indentation distance, parse font attributes, and detect page break positions to divide the document into hierarchical logical units. It automatically identifies the document's physical and logical structure, and based on the identified features such as heading styles, numbering sequences, indentation, font, and page breaks, it uses the root node as the document's starting point, establishes parent-child node connections based on the identified heading level relationships, treats paragraphs at the same level as sibling nodes, and constructs a tree data structure with depth attributes. It traverses the paragraph text, identifies sentence-ending punctuation marks as candidate boundaries, and then checks whether the characters after punctuation are uppercase letters, numbers, or line breaks to verify the boundary validity. It also excludes false boundaries after abbreviations and filters abnormal short sentences using a sentence length threshold. Finally, it extracts valid text fragments as independent sentences and assigns a globally unique original text position identifier to each parsed chapter, paragraph, and sentence.

[0027] After completing the document structure parsing, the file extension, creator information, and template identifier are extracted from the document metadata for initial screening. Then, based on the document structure tree, title-level keywords are analyzed. If the title contains the words "contract" or "agreement" and includes clauses on the responsibilities of both parties, the business type is classified as a contract; if it contains reports, summaries, or analyses with a progressively structured chapter structure, it is classified as a report; if it contains notices, requests, or approvals and conforms to the official document format standards of Party and government organs, it is classified as an official document; all others are classified as general office documents. Based on the current office document's business type, matching review focus element definition information is triggered and read from the preset review rule base. This definition information includes structured element type codes, element feature patterns describing the key semantic and syntactic features of the elements, and associated review rule identifiers bound to them. The abstract description of the element definition information is parsed, and the regular expression skeleton, semantic constraints, and contextual rules are extracted. Then, the template compilation engine is called to convert the abstract template... The placeholder variables in the formula are replaced with specific parameters of the current office document, including document type code, business domain identifier, and timestamp information. An executable regular expression matching template and semantic recognition template for the current document instance are generated. This compilation process uses dynamic code generation technology to convert the feature pattern of the element into a recognition function object that can be directly called. The chapter structure tree, paragraph position index, and extracted named entity dictionary of the current document are injected as runtime context. Finally, the entire document is traversed, and template matching operation is performed on each text segment. When a match is successful, specific numerical values, dates, responsible parties, key clause references, and other element content are extracted and encapsulated into element instance objects containing element type code, element value, original text position identifier, and confidence score. A set of elements of concern for review is generated and a unique identifier based on document identifier and sequence number is assigned to each instance. At the same time, a hash mapping index between element instances and review rule identifiers is established to provide a clear rule foundation for subsequent automated review.

[0028] The pre-set review rule base is achieved through the following steps: Based on domain knowledge accumulation and typical document analysis, various review focus elements and their associated rules are manually defined and structuredly entered. Each rule record clearly includes the element type code, element feature pattern, and associated review rule identifier. Business type tags are established for all rules and stored in a structured database or configuration file to form a rule knowledge base that can be persistently stored and managed. This rule base supports dynamic expansion and version management, and allows new element definitions and rules to be imported in batches through a configurable interface or script according to new review needs or business scenarios.

[0029] Download the Natural Language Processing open-source toolkit directly from the official code hosting platform, visit the open-source community, search for the target toolkit name, go to the project homepage, select the corresponding operating system and programming language distribution version, download it to your local machine, and then compile and install the Natural Language Processing open-source toolkit according to the documentation instructions. Initialize the core processing modules through its application programming interface, including word segmentation and part-of-speech taggers, named entity recognizers, dependency parsing analyzers, and semantic role taggers, while loading the corresponding pre-trained models or rule bases.

[0030] The system loads word segmentation and part-of-speech tagging tools to accurately segment the current sentence and simultaneously tag the part of speech of each word. Word segmentation divides a continuous sequence of characters into the smallest linguistic units with independent semantics. Part-of-speech tagging assigns grammatical category labels such as noun, verb, and adjective to each segmented unit. Then, a named entity recognizer is invoked. Based on direct matching of text fragments using predefined regular expression patterns or an entity dictionary, the recognition process involves traversing the segmented word sequence, matching the current word against the entity dictionary, and marking a successful match as the corresponding named entity. The extraction process involves extracting the recognized entity words and their start and end positions in the original text. The data is encapsulated into named entity records containing entity type, entity value, and original document location. Simultaneously, by statistically analyzing word frequencies and calculating inverse document frequency (IVF) weights—where IVF is calculated by counting the number of times a word appears in a single sentence and dividing by the total number of sentences containing that word—words that appear frequently in the current sentence but are only found in a few sentences across the entire document set receive higher weights. Combined with the part-of-speech tagging information from the preceding steps, stop words and function words are filtered out, and keywords representing the core semantics are extracted from the sentences. Finally, the word segmentation results, part-of-speech tags, named entity tags, and keyword lists are integrated into structured semantic analysis foundational data.

[0031] Word segmentation and part-of-speech tagging are fundamental software components or algorithm libraries in natural language processing. Their core function is to segment a continuous, uninterrupted sequence of Chinese characters into a set of independent, meaningful lexical units according to certain linguistic and statistical rules. By combining dictionary matching and statistical machine learning models, they often output part-of-speech tags for each word while completing basic word segmentation. In advanced implementations, they also integrate sub-functions such as named entity recognition, thereby completing basic lexicalization, preliminary grammatical classification, and key entity extraction of text in a single scan, providing structured lexical-level input for downstream tasks such as syntactic analysis and semantic understanding.

[0032] The grammatical structure of sentences is analyzed and parsed using a dependency parser, identifying the core components of subject, verb, and object and the modification relationship. The predicate and argument structure of sentences are extracted based on a semantic role labeler, and it is determined whether a sentence contains multiple independent semantic expressions.

[0033] The sentence, after being processed by word segmentation and part-of-speech tagging, is input into the dependency parser. The dependency parser initializes the stack and input buffer, pushes the root node onto the stack, and stores the words in the sentence into the input buffer in sequence. It then enters a loop processing state, reading one word from the input buffer and pushing it onto the stack each time. When the number of words in the stack is greater than or equal to two, it reads the predefined dependency relation rule library. This rule library includes subject-verb relation rules (noun followed by verb), verb-object relation rules (verb followed by noun), attributive-head relation rules (adjective or noun followed by noun), adverbial-head relation rules (adverb followed by verb or adjective), and verb-complement relation rules (verb followed by complement marker). In dependency parsing, the core word is the ruler, and the modifier is the ruled. The specific rules are as follows: In subject-predicate and verb-object relations, the verb is the core word and the noun is the modifier; in attributive-head relations, the central noun is the core word and the attributive is the modifier; in adverbial-head relations, the central verb or adjective is the core word and the adverb is the modifier; in verb-complement relations, the verb is the core word and the complement is the modifier. The part-of-speech combination of the top two words in the stack is matched against the rule base. If a match is successful, a directed edge from the core word to the modifier is established as a dependency arc. The dependency arc is labeled with the corresponding grammatical relation type, and the modifier is removed from the stack. If no modification relation exists, a new word is pushed onto the stack. This process is repeated until the input buffer is empty and only the root node remains in the stack, forming a connected acyclic tree structure. Based on this, the semantic role labeling module is called, using the dependency parsing results as input. The dependency relation tree is traversed to identify the core predicate in the sentence. The predicate is taken as the current node, and the words in its child nodes and parent node chains are taken as candidate arguments. The predefined semantic role rule base is read. The rule base includes agent rules (nouns connected to predicates via subject-predicate relations), patient rules (nouns connected to predicates via verb-object relations), time rules (time nouns or phrases containing time prepositions), place rules (place nouns or phrases containing place prepositions), and manner rules (manner nouns or phrases containing manner prepositions). The part-of-speech of candidate arguments and their dependency relationship with predicates are matched against the rule base. If the rule conditions are met, they are marked as the corresponding semantic roles and combined to form a predicate argument structure. The system determines whether a sentence contains multiple independent semantic expressions. Specifically, the steps are: counting the number of predicates in the sentence, analyzing the dependency relationships between predicates; if there are two or more predicates connected by a parallel dependency arc or each belonging to a different subject, the sentence is determined to contain multiple independent semantic expressions; if predicates are connected by a subordinate dependency arc or share the same subject, it is determined to be a complex sentence with a single semantic expression, providing a basis for subsequent semantic segmentation.

[0034] Dependency parser is a theory used to describe the grammatical relationships between words in a sentence. Unlike traditional subject-verb-object parsing, dependency parsing does not divide phrase structures, but directly defines binary, directed, and asymmetric dependency relationships between words. In this relationship, one word is the core and another word depends on it. The relationship type is indicated by a label.

[0035] The semantic role labeling module is a tool for deep semantic parsing based on dependency parsing results. It takes the predicate in the sentence as the core analysis object, extracts candidate arguments related to the predicate by traversing the dependency relation tree, and performs pattern matching between the part-of-speech features, dependency relation types with the predicate, and preposition tags of the candidate arguments and a predefined semantic role rule base. This identifies and labels semantic roles such as agent, patient, time, place, and manner, and finally constructs a structured predicate-argument framework to reveal the deep semantic information in the sentence.

[0036] When a sentence is detected to contain multiple independent semantic expressions, it indicates that multiple complete logical units are compounded at the grammatical level. If these are not split, some information is easily lost or confused during the subsequent summary generation process, resulting in incomplete summary content. Furthermore, during the review stage, it will be impossible to accurately trace the correspondence between the summary statement and the specific events in the original text, causing semantic consistency comparison to fail. Therefore, splitting sentences into initial semantic fragments based on predicate boundaries and argument integrity standardizes the complex source text into a series of atomic, semantically self-sufficient information units. This allows each independent fact or event to be processed, summarized, and reviewed individually and accurately, and establishes a precise mapping between each fragment and its position in the original text. This lays a reliable data foundation for the subsequent generation of high-quality, verifiable summaries and the execution of precise automated review.

[0037] The sentence is split into initial semantic segments based on predicate boundaries and argument integrity. The specific steps are as follows: using each independent core predicate as an anchor point, the core arguments directly governed by that predicate are determined according to the dependency syntax tree. The dependency syntax tree is traversed to locate the node corresponding to the current core predicate. All child nodes of that node are considered as candidate arguments directly governed. Core arguments are then selected based on the dependency relationship type between the child nodes and the predicate node. Child nodes with subject-predicate dependency arcs are marked as agent arguments, and child nodes with verb-object dependency arcs are marked as recipient arguments. The argument of the predicate is a time or place argument with a prepositional dependency arc and the preposition type is time or place. The selected argument nodes and their dependency types are combined to form the core argument of the predicate. Then, by traversing the syntactic tree, all subordinate components that have direct modification or subordinate relationship with the current predicate and its core argument are completely collected, thus forming an initial semantic segment that is grammatically and semantically self-sufficient. Then, each segment is assigned a consecutive segment number and the original text position identifiers corresponding to all words in the segment are accurately recorded.

[0038] When a sentence is detected to contain only a single semantic expression, it indicates that the sentence constitutes an indivisible complete semantic unit in terms of syntax and logic. Its core feature is that it contains only one core predicate, and the subject, object, and various modifiers constructed around the predicate jointly express a coherent and single semantic event. Therefore, there is no need to split it. The entire sentence is directly treated as an initial semantic fragment. This means that the complete information carried by the sentence is retained as an atomic processing unit in the subsequent process, which is consistent with the fragments generated by the split multi-semantic expression sentence in terms of logical hierarchy.

[0039] All initial semantic fragments are sorted according to their serial numbers and the order in which they appear in the original text, forming a structured set of initial semantic fragments.

[0040] Traverse the initial set of semantic segments. For each initial semantic segment, detect adjacent initial semantic segments. If the explicit referential marker at the beginning of the latter segment is detected and verified to clearly point to the core noun or phrase that has appeared in the former segment, a referential relationship is determined. If the entity lists extracted from the two segments are compared, and if they both match the same element keyword defined in the review rule base, it is determined that they both involve the same review focus element. If, by identifying specific logical connectors and analyzing the syntactic structure, it is confirmed that the former segment contains conditional words and the latter segment contains conclusion words, and the subjects of the arguments of the two segments are consistent, it is determined that they constitute a complete logical unit.

[0041] If there is a referential relationship between adjacent initial semantic segments, it means that the two segments are logically closely interdependent and together constitute a more complete and coherent semantic representation unit. The former segment provides a clear entity that is referred to, and the latter segment expands on the entity with further statements or descriptions. Merging these two segments into the same candidate semantic segment can avoid the problems of information gaps, unclear referentials, or semantic fragmentation that would occur in subsequent summary generation due to forcibly severing this natural semantic coherence. This ensures that the semantic unit used for the final summary generation is self-sufficient and coherent in context, thereby guaranteeing the accuracy and readability of the generated summary and establishing a clear and complete semantic comparison scope for subsequent document review.

[0042] If adjacent initial semantic segments both involve the same element of review concern, it means that these two segments focus on the same core fact, clause, or entity in terms of content. They describe or limit the same element from different angles or under different conditions. Merging them into the same candidate semantic segment is to ensure that all relevant information scattered in different clauses can be centrally reviewed as a complete contextual unit during the subsequent review process. This prevents review omissions or misjudgments caused by information fragmentation, and ensures that the completeness, consistency, and internal logical conflicts of the element of review concern can be checked based on the most sufficient contextual information. This generates accurate and comprehensive document review results and ensures the completeness and accuracy of the final abstract's description of the element.

[0043] If adjacent initial semantic segments form a correspondence between conditional and concluding semantics, it means that these two segments logically constitute a complete reasoning chain of premise and result or hypothesis and inference. Merging these two segments into the same candidate semantic segment is to prevent this complete logical unit from being fragmented in the subsequent summary generation process, thereby avoiding the generated summary from appearing arbitrary or ambiguous due to the loss of key premises. At the same time, during the document review stage, the logical consistency, validity, and completeness between conditions and conclusions can be verified holistically in a unified context, thereby accurately identifying possible logical loopholes in the original text or logical distortions in the summary, ensuring the rigor and accuracy of the final output.

[0044] After merging, a set of candidate semantic segments is formed. Each candidate semantic segment corresponds to a set of continuous or non-continuous original text position markers, and each candidate semantic segment forms a logically coherent semantic whole.

[0045] like Figure 2The flowchart of the review strategy method involved in this application embodiment includes: generating corresponding summary units based on candidate semantic fragments and establishing a mapping relationship between summary units and original text positions; then, by analyzing the decoder attention weights and information entropy, calculating the importance weight of each summary unit, and classifying them into different processing levels using a first or second review execution strategy; if a summary unit involves elements of review concern, it is marked as retained; the first review strategy expands the review scope and performs cross-position consistency comparison, while the second review strategy only performs rule checks within the basic scope; anomalies triggered during the review process are recorded and associated with the summary unit. Subsequently, the review results are parsed to determine contextual missingness, semantic inconsistency, and conflicting descriptions.

[0046] Taking a set of candidate semantic segments as input, a pre-trained generative text summarization algorithm is invoked. Each candidate semantic segment is encoded as an independent source text sequence, and a corresponding concise summary expression is generated through a decoder. Each generated summary expression is assigned a globally unique summary unit identifier, which is usually generated by combining an identifier, batch number, and sequence number. Based on this, two precise mapping relationships are established and maintained: first, a direct mapping from the summary unit identifier to its source candidate semantic segment to record the generation origin; second, and more importantly, a mapping from the summary unit identifier to the set of all original text position identifiers corresponding to the candidate semantic segment. This set integrates the discrete positions of all its constituent initial segments in the original text, thereby establishing a traceable location link between the summary unit and the original text that covers complete contextual support. This dual mapping relationship constitutes the fundamental basis for all subsequent review and consistency comparison.

[0047] The mapping relationship is established by initializing a hash table data structure. The first mapping relationship is established using the digest unit identifier as the key and the candidate semantic fragment content or fragment reference pointer as the value. The second mapping relationship is established using the digest unit identifier as the key and the list or bitmap of the original text position identifier set as the value. The key-value pairs are stored in memory through hash table insertion operations, the mapping relationship is retrieved through hash table lookup operations, and the mapping relationship is dynamically maintained through hash table update operations. In this way, a traceable positioning link with complete context support is established between the digest unit and the original text. This dual mapping relationship constitutes the fundamental basis for all subsequent review and consistency comparison.

[0048] A hash table data structure is a data organization method that uses hash functions to achieve fast data access. By converting keys into array indices for storing values, hash functions enable constant-level fast lookup, insertion, and deletion, ensuring data integrity and access efficiency.

[0049] like Figure 3The diagram below illustrates the generative summarization algorithm architecture of this application, comprising an input layer, an encoder, a decoder, and an output layer. The input layer converts the input words or subwords into continuous vector representations via an embedding layer and adds positional encoding to preserve the word's positional information within the sequence. The encoder processes the input sequence using a multi-head self-attention mechanism, allowing the model to establish relationships between words at different positions. Layer normalization and residual connections stabilize the training process. The encoder's output is processed by a feedforward neural network and then passed to the decoder. The decoder uses a cross-attention mechanism to focus on the encoder's output, thereby aligning the input and output sequences. The decoder's output is processed by a feedforward neural network and finally transformed into a probability distribution through a Softmax (normalized exponential function) layer to generate predicted output words.

[0050] Generative text summarization algorithms are implemented based on the Transformer encoder-decoder model. Its core mechanism is as follows: The encoder is responsible for understanding the candidate semantic segments of the input. It captures the word associations within the segments through its multi-head self-attention layer and converts them into semantic vector representations containing contextual information. The decoder then generates summary text word by word in an autoregressive manner based on this representation. When generating each new word, the cross-attention layer can dynamically examine and focus on the most relevant part of the encoder output. The attention weights generated by this focusing process accurately quantify the generative associations between the summary words and the original words. Therefore, this model not only outputs fluent summary sentences, but its internal attention weight matrix also naturally establishes a quantifiable semantic mapping from the summary unit to the original semantic segments.

[0051] The attention weight matrix corresponding to the unit is extracted from the cross-attention layer of the decoder. Then, based on this matrix, core metrics including semantic coverage score and semantic concentration score are calculated. The calculation process for the semantic coverage score is as follows: the attention weight matrix is ​​analyzed row by row, that is, the distribution of original text information relied upon by each word in the summary during generation is analyzed. For each row, the information entropy of its attention distribution is calculated. , where P i The normalized attention weight is represented at each word position in the original text in the row, A represents the information entropy of each row, i=1,2,3,...,N, and N represents the total number of words in the original text segment corresponding to the current summary unit. The entropy values ​​of all rows are averaged to obtain the semantic coverage score of the summary unit, which is used to evaluate the breadth and balance of the summary's coverage of the original text information.

[0052] The semantic concentration score is calculated as follows: by analyzing the columns of the matrix, the attention weight matrix is ​​summed column by column to obtain the original word importance vector, which reflects the total contribution of each original word to the generation of the entire summary. The values ​​in this vector are then sorted in descending order from largest to smallest. The highest-ranking values ​​are selected according to the proportion of core words and summed to obtain the total contribution of core words. Then, the sum of the importance vectors of all values ​​in this vector is calculated to obtain the contribution of all words. Finally, the sum of the core word contributions is divided by the sum of the contributions of all words to obtain the semantic concentration score.

[0053] Finally, the two scores are integrated into a unified importance weight for the summary unit by means of the average value. The higher the weight value, the more comprehensive the information coverage of the summary unit or the more focused the semantic expression is on the key information, thus providing a quantitative basis for determining its processing level and the allocation of review resources.

[0054] The steps for presetting the core vocabulary ratio in the database are as follows: First, during the database design phase, a dedicated configuration table or parameter table is created to structure and store various review algorithm parameters. Then, one or more configuration records for the core vocabulary ratio are created in this table. Each record clearly includes fields such as parameter name, parameter value, applicable business document type, and effective status. After experimental verification, experts determine which configuration table to store in the database. This design allows administrators to directly update these parameter values ​​through the configuration interface or database management tools, thereby achieving the goal of flexibly adjusting algorithm sensitivity and review strategies without modifying program code.

[0055] Match the set of original text location identifiers corresponding to each abstract unit with the set of elements of concern for review:

[0056] When the original text corresponding to a summary unit contains any element of review concern, it means that the summary unit carries and expresses a key information point in the document that is subject to rules and requires in-depth verification. This means that the unit is no longer a general summary, but is directly related to core elements such as specific business terms, numerical definitions, attribution of responsibility, or conditional conclusions. Therefore, a reserved mark is set for the summary unit.

[0057] Summary units that do not contain elements of review concern are defined as follows: The content of this summary unit does not directly involve core business entities or key clauses that are predefined by the review rule base and must be subject to mandatory verification. Therefore, its core nature is a general statement of document background information, general facts, or auxiliary descriptions, and no reservation mark is set for it.

[0058] Based on the importance weight of the summary unit, the processing level of the summary unit is divided. When the importance weight of the summary unit is greater than or equal to the preset summary importance threshold, it means that the unit has been identified as high-value or high-risk content in the algorithm evaluation. The summary unit has high information density, that is, its attention weight analysis shows that it either broadly covers multiple key parts of the original text or highly focuses on a few core statements of the original text. The semantics it carries are complete and critical. If there is a deviation in the subsequent review, it may lead to major ambiguity or factual errors in the summary as a whole. Therefore, the review execution strategy for this summary unit is determined as the first review execution strategy.

[0059] When the importance weight of a summary unit is less than the preset importance threshold, it indicates that the unit has not been identified as the core content with the highest information density or the most critical semantics in the algorithm model. Its attention weight analysis shows that the information coverage of the unit in the original text may be relatively scattered, or it may focus on relatively minor supporting or background statements. Therefore, the review execution strategy for this summary unit is determined as the second review execution strategy.

[0060] The preset importance threshold for the abstract in the database is achieved by preparing a batch of representative typical office documents as a test set, running the abstract generation and importance weight calculation process for these documents, collecting the importance weight value distribution of all abstract units, then having domain experts review these abstract units, manually marking which should be classified as high priority for review based on their business criticality, and finally analyzing the correspondence between the weight values ​​and the manual marking results, selecting a weight value that can most accurately distinguish between high and low priority units as the threshold. The value obtained through experimental analysis is determined as the abstract importance threshold and written into the parameter table of the database.

[0061] Based on the set of original text location identifiers mapped from the abstract units, the directly corresponding original text sentences are precisely located as the basic review scope. A dependency parser is invoked to analyze each sentence within this scope, identifying defining noun phrases that function as core subjects and objects, as well as conclusive verb-object combinations expressing core arguments or results. The dependency tree is traversed, searching for noun phrases with subject-verb dependency arcs as subject candidates and noun phrases with verb-object dependency arcs as object candidates. Subject and object candidates are merged into a definition keyword set. Verb-object combinations with verb-object dependency arcs are then searched, and verbs and their object combinations are merged to form a conclusion keyword set. These keywords are then used as search criteria to search the full-text index of the entire document, locating all other paragraphs containing any keyword and collecting their location identifiers. Finally, the location identifier set of the basic review scope is merged with all new location identifier sets obtained through keyword retrieval, removing duplicates to form a complete and coherent extended original text scope. This ensures that subsequent in-depth review can be conducted within a sufficient and relevant context.

[0062] Load all review rules corresponding to the set of review focus elements from the review rule base and bind them to the summary unit.

[0063] First, the set of review focus elements identified for the summary unit is read. Each element in the set has a unique element identifier. Using these element identifiers as query keys, a quick search is performed in the mapping index of the review rule base to retrieve all review rule identifiers pre-associated with these elements. Then, based on these rule identifiers, the corresponding specific review rules are fully loaded from the rule base. Finally, these rules are bound to the summary unit, that is, the rule set is stored in memory or a temporary database as a dedicated review task list for the unit. This ensures that in the subsequent rule-by-rule verification process, the review engine can accurately know which specific checks need to be performed sequentially for this summary unit, thus achieving precise and dynamic review and differentiated in-depth verification of different content.

[0064] Following the preset execution order in the review rule base, each loaded review rule is sequentially invoked within the extended original text scope. Each rule acts as an independent verification function, receiving text within the current original text scope as input. Based on its built-in logic (such as regular expression matching, numerical calculation, or semantic pattern detection), it scans and analyzes the text. After execution, the rule engine immediately generates a structured intermediate result for rule judgment. This result includes at least the rule's unique identifier, rule triggering status (such as triggered or passed), matched key text fragments, calculated matching score, and a detailed description of the triggering anomaly. All intermediate results are recorded in real time and stored in memory or a temporary database, and associated with the current summary unit identifier and the corresponding original text position to form a complete review process traceability chain. This provides detailed, verifiable evidence for subsequent summary judgments, correlation analysis, and the generation of the final review result.

[0065] The execution of the review rules is essentially an automated verification program unit that transforms and encapsulates a specific business review requirement (such as checking whether the payment amount is consistent with the capitalized version or verifying whether the signing date is earlier than the performance period). Each rule clearly defines three core parts: a unique rule identifier, specific matching or calculation logic (such as using regular expressions to extract the amount or using a date comparator to determine the order), and clear verification standards (such as matching thresholds). When the review is executed, it does not make a vague judgment, but strictly plays the role of a tireless professional auditor: receiving text, calling the calculation module embedded in the rule to perform precise pattern matching, numerical calculation or logical reasoning, and finally outputting a quantitative conclusion of pass or failure and its complete chain of evidence according to the preset standards.

[0066] The regular expression matching calculation module is pre-set in the review rule base. It reads the predefined regular expression pattern library, which contains regular expression strings and matching weight coefficients corresponding to each rule type. It matches the text within the extended original text range with the regular expression, counts the number of successfully matched characters and the total number of characters in the original text, and calculates the matching percentage by dividing the number of successfully matched characters by the total number of characters in the original text and then multiplying by 100%. If there are multiple matchable substrings in the original text, the longest matching substring is taken to calculate the matching percentage. Finally, the matching degree is output.

[0067] For each review rule applied to the text within the extended original text range, an independent matching score calculation process is initiated. This process first calls the corresponding regular expression matching calculation module based on the rule type, performs pattern matching using regular expressions, and calculates the matching percentage. This score is compared with a preset matching threshold, and simultaneously checks whether the key fields required by the rule exist in the text. If the rule matching score is lower than the preset matching threshold, or after traversing the extended original text range, the extended original text range is matched and verified against the format regular expressions corresponding to the field types. If no candidate segments are found after the traversal, or if the candidate segments fail to match the format regular expressions, the necessary fields are identified. The missing information indicates that the original text being reviewed fails to meet the formal requirements or complete business constraints defined by the rule. A matching degree lower than the preset matching threshold means that there is a significant deviation between the original text expression and the expected pattern of the rule, which may be due to vague expression, incorrect format, or incomplete information. Missing necessary fields directly indicate that key information items do not exist at all. Both of these situations cause the automated program to be unable to make a definite conclusion of approval on this rule dimension. Both of these situations are judged as triggering a single point of failure, and a structured intermediate result record of the failure is generated. Its content includes the unique identifier of the triggering rule, the calculated specific matching degree value, and a clear list of missing fields.

[0068] Conversely, if the rule matching degree is higher than or equal to the preset matching threshold and no necessary field is missing, it indicates that the original text being reviewed has fully met the formal requirements and business constraints defined by the rule in terms of expression format and information completeness. A matching degree higher than or equal to the preset matching threshold means that the original text features are highly consistent with the expected pattern of the rule. The absence of necessary field is ensured that all key data points are accurately identified. In this case, the rule is deemed to have passed, and a normal result containing the rule identifier and the passing status is recorded accordingly.

[0069] Taking a typical review of a purchase contract (an office document type) regarding liquidated damages as an example, the specific form and execution process of the rule are as follows: A rule is predefined, with its unique identifier being 001. The review focus element associated with this rule is the liquidated damages ratio. Its verification logic clearly stipulates that the two keywords "liquidated damages" and "ratio" and a percentage value must be extracted from the original text simultaneously, and this value must not exceed 20% of the total contract amount. When executing this rule within the expanded scope of the original text, the text is scanned using a regular expression module to attempt to match the pattern of liquidated damages | breach of contract compensation | penalty ratio. If the key components such as "liquidated damages," "daily payment," and "0.5%" are successfully matched in the sentence "If Party B delays delivery, it shall pay liquidated damages of 0.5% of the total contract amount per day," the confidence level of this match is calculated (for example, due to the complete match of keywords and percentage format, the matching degree reaches 66%), and compared with the preset matching threshold (such as 60%). If the matching degree is higher than the matching threshold, the formal verification is deemed successful, and finally, a record of rule approval is generated.

[0070] After completing rule-by-rule validation, all positional identifiers within the expanded original text are traversed, and their texts are parsed and extracted into three categories: numerical expressions, conditional expressions, and conclusion expressions. For numerical expressions, numbers, units, and context are directly extracted and converted into standard quantification formats. The specific conversion steps are: recognizing numeric character sequences, converting them into floating-point or integer values, recognizing unit words after numbers, and combining values ​​and units into standardized numeric strings. For conditional and conclusion expressions, semantic parsing tools are used to convert them into structured logical forms. The specific conversion steps are: recognizing conditional and conclusion conjunctions, extracting conditional and conclusion clauses, marking conditional clauses as logical premises, marking conclusion clauses as logical inferences, and combining them into a premise-inference structure. All extracted similar expressions are paired and cross-positional consistency comparison is performed. For numerical expressions, their values ​​are directly compared. Within the allowable error range, for the expression of conditions and conclusions, on the one hand, a logical reasoning model is used to determine whether there is an implicit logical contradiction between different conditions; on the other hand, a semantic similarity model is used to convert each conclusion expression text to be compared into a fixed-dimensional semantic vector. Then, the cosine similarity between these two vectors of different conclusion expressions is calculated and recorded as the semantic vector similarity. Finally, a judgment is made based on the comparison results: if any set of unequal values, logical contradictions, or semantic similarities below the preset similarity threshold are detected, it indicates that the office document has internal inconsistencies, logical conflicts, or semantic ambiguities in these key information points. These situations all point to a core problem: different parts of the office document have irreconcilable differences in their expression of the same fact or the same logical relationship, causing the office document as a whole to fail to form a self-consistent and rigorous semantic whole. In this case, an association anomaly is triggered and a detailed record is generated; otherwise, it is determined that no anomaly has been triggered.

[0071] The construction of a semantic similarity model typically follows this process: First, a large-scale text pair dataset is collected and cleaned, containing manually labeled samples with high, medium, and low semantic similarity. Next, a pre-trained language model with a multi-layer Transformer encoder architecture is selected as the foundation. Pre-trained weights are obtained by performing a masked language model training task on a large-scale historical unlabeled text dataset. Then, a pooling layer is added to this model to output fixed-dimensional semantic vectors, and a loss function is learned. This function uses cosine similarity to measure the semantic closeness of sentence pairs. The model parameters are optimized by narrowing the distance between positive sample pairs and widening the distance between negative sample pairs. This optimizes the model to learn how to map semantically similar sentences to similar positions in the vector space. Subsequently, the model is fine-tuned in a supervised manner using a prepared dataset of historically reviewed office documents, continuously optimizing its parameters to accurately quantify the semantic similarity between any two sentences. After training, the model is encapsulated as an independent service or function, which can be called by reviewers when they need to determine the semantic consistency between the summary and the original text. An interpretable semantic vector similarity is obtained by calculating the cosine similarity between the two semantic vectors.

[0072] The similarity threshold is preset in the database through the following steps: First, based on the analysis of the business scenario and the evaluation of the performance of the semantic model on the test set, initial threshold suggestions are set for different types of semantic comparison tasks. Then, experts create corresponding records in the database through configuration tools, clearly writing the parameter name, comparison type, threshold value and applicable business scope. This threshold supports dynamic optimization. Based on the feedback of misjudgments in the actual review results, its value can be fine-tuned by updating the database record. Finally, when performing cross-location semantic consistency comparison, the threshold will be queried in real time according to the current comparison type and used as a strict standard for judging whether the conclusion expression is consistent.

[0073] When any review rule determines that an intermediate result or consistency comparison result meets the abnormal triggering condition, a corresponding review result record is generated and associated with the summary unit identifier.

[0074] The abnormal triggering conditions refer to triggering a single point of failure and triggering a related failure.

[0075] The basic review scope is determined based on the original text location identifier set. All review rules corresponding to the review focus elements involved in the summary unit are loaded from the review rule base and bound to the summary unit. The set of review focus elements identified for the summary unit is read. Each element in the set has a unique element identifier. Using these element identifiers as query keys, a fast search is performed in the mapping index of the review rule base to retrieve all review rule identifiers pre-associated with these elements. Then, based on these rule identifiers, the corresponding specific review rules are fully loaded from the rule base. Finally, these rules are bound to the summary unit, that is, the rule set is stored in memory or a temporary database as a dedicated review task list for the unit. This ensures that in the subsequent rule-by-rule verification process, the review engine can accurately know which specific checks need to be performed sequentially for this summary unit, realizing the precision and dynamism of the review, and performing differentiated in-depth verification of different content.

[0076] Following the preset execution order in the review rule base, each loaded review rule is called sequentially within the basic review scope. Each rule acts as an independent verification function, receiving text within the current original text scope as input. Based on its built-in logic (such as regular expression matching, numerical calculation, or semantic pattern detection), it scans and analyzes the text. After execution, the rule engine immediately generates a structured intermediate result for rule judgment. This result includes at least the rule's unique identifier, rule triggering status (such as triggered or passed), matched key text fragments, calculated matching score, and a detailed description of the triggering anomaly. All intermediate results are recorded in real time and stored in memory or a temporary database, and associated with the current summary unit identifier and the corresponding original text position to form a complete review process traceability chain. This provides detailed, verifiable evidence for subsequent summary judgments, correlation analysis, and the generation of the final review result.

[0077] For each rule, the regular expression matching module calculates the rule matching degree. If the rule matching degree is lower than the preset matching threshold or a necessary field required by the rule is missing, it means that the original text being reviewed fails to meet the minimum requirements defined by the rule in terms of form or completeness. A matching degree lower than the preset matching threshold indicates that there is a significant deviation between the expression pattern of the original text and the expectation of the rule, which may be due to formatting errors, mismatched wording, or ambiguity. Missing fields directly mean that the core information constituting the business elements is not provided at all. Since sufficient evidence cannot be obtained to make a definite judgment in this case, the rule is determined to trigger a single point of failure. The intermediate result of the rule judgment, including the rule identifier, rule matching degree, and list of missing fields, is recorded. Conversely, if the matching degree is higher than the preset threshold, it means that the original text being reviewed has fully met the requirements of the rule in terms of form matching and information completeness. In this case, it is determined that a single point of failure has not been triggered, and the rule is recorded as passed.

[0078] When the intermediate result of any core review rule triggers a single point of failure, a corresponding review result record is generated and associated with the identifier of the summary unit.

[0079] Each input review result record is structured and parsed. First, the standard format of the review result record is identified and read. Then, based on predefined data patterns or delimiters, three core fields are precisely matched and extracted: anomaly type, original text location identifier that triggered the anomaly, and associated summary unit identifier. Based on rule-based regular expressions, key phrases and identifiers are located and captured. The extracted information is immediately converted into an internally unified structured data object, completing the conversion from unstructured or semi-structured raw records to standardized information that can be used directly and accurately. Based on the parsed structured information, logical judgments, including context missing judgment, semantic inconsistency judgment, and conflict description judgment, are automatically executed.

[0080] Context missing determination involves locating the corresponding abstract text based on the parsed abstract unit identifiers, and identifying three types of key elements in the abstract through dependency parsing: citation elements, definition elements, and conditional elements. Citation elements are external documents mentioned in the abstract, identified by matching quotation marks; definition elements are defining statements of core concepts or terms in the abstract, identified by identifying defining conjunctions; and conditional elements are parts of the abstract expressing premises or constraints, identified by identifying conditional conjunctions and conditional clauses. Using the identified key elements as anchor points, the abstract unit is then associated with... The system uses a set of original text location identifiers to locate and read the complete text of all these locations in the document, forming a set of original text contexts to be searched. Then, each anchor point extracted from the summary is semantically matched with this set of original text contexts. If any key element has a highest similarity score with all sentences in the original text set that is lower than a preset similarity threshold, it is determined that the key element lacks necessary contextual support. Finally, for all key elements that have not found support, a summary-related review result of type "Necessary Context Missing" is generated, which records in detail the content of the missing element, the associated summary unit, and the original text location.

[0081] The semantic inconsistency determination process includes obtaining the abstract expression text based on the abstract unit identifier, reading and integrating the original text content of all corresponding positions according to the set of original text position identifiers associated with it, inputting the abstract sentence and the integrated original text sentence into a pre-trained semantic similarity model, calculating their vector similarity score, and if the overall semantic similarity is lower than the preset similarity threshold, it is determined that there is a semantic mismatch relationship, and a review result of type "speech expression and original text semantic inconsistency" is generated, which records in detail the specific content of the mismatch, the type of constraint involved, and the comparison basis.

[0082] The conflict description determination process involves extracting the numerical expression, conditional expression, and conclusion expression from each target summary unit based on its associated multiple original text location identifiers. Similar expressions are then standardized and structured, followed by cross-comparison of the structured expressions: for numerical values, equality is directly checked; for conditions and conclusions, a logical reasoning model is used to determine if different conditions are mutually exclusive or contradictory, and a semantic similarity model is used to calculate the vector similarity of different conclusions. If any comparison reveals unequal numerical values, direct logical conflicts in conditions, or a semantic similarity of the conclusion below a preset similarity threshold, an original text conflict description is determined, generating a summary-related review result of type "Original Text Conflict Description," which details the conflicting original text locations, the specific content of the conflict, and the type of conflict, thereby revealing internal inconsistencies within the document.

[0083] When any review result related to a summary unit is detected, the state of the summary unit is marked as inconsistent. The set of original text position identifiers initially mapped to the unit is reread, and it is confirmed whether it has a reserved mark. When regenerating the summary, the reserved mark and its corresponding key elements are input as strong constraints into the generative text summarization algorithm to guide the algorithm to accurately cover these elements when reconstructing the summary. After generating new summary content, the text content of the summary unit and its mapping relationship with the original text position are updated synchronously, and its inconsistent state is cleared. Summary units that are not marked as inconsistent remain unchanged and do not participate in this update cycle.

[0084] After completing the partial update of the summary unit, a complete re-evaluation and review cycle is initiated. The specific steps and cycle termination mechanism are as follows: Based on the updated summary unit content and its new mapping relationship with the original text, the core calculation is re-executed. That is, the importance weight of each summary unit is calculated using the attention weight distribution of the generative text summarization algorithm, and the review execution strategy for each unit is re-determined accordingly. Then, according to the newly determined strategy, a complete review process is performed on the office document corresponding to each summary unit, including applying all relevant review rules within the expanded original text scope and performing cross-position consistency comparison, thereby generating a new review result record. Based on this, three judgments are made again: missing context, semantic inconsistency, and conflict description, to determine whether any new summary-related review results are generated. This process constitutes a complete loop of evaluation, review, and judgment, and the output of this loop is continuously monitored: if no new summary-related review results are generated in all summary units in a complete loop, it indicates that a stable state has been reached, all known summary consistency issues have been fixed, at which point the iterative judgment converges, the loop process is terminated, and all the latest summary texts are integrated into the final summary result. At the same time, all accumulated review results are summarized into a complete document review report and output together.

[0085] like Figure 4 The diagram shows the structure of an office document automatic review and management system based on a text summarization algorithm provided in this application, including: an office document parsing module, an office document review and summary generation module, and a document summary update module.

[0086] The office document parsing module is used to perform structural parsing on office documents and assign original text location identifiers, read the elements of interest for review and editing, generate a set of elements of interest for review and editing, perform semantic parsing based on the original text location identifiers and generate an initial set of semantic fragments, forming a set of candidate semantic fragments.

[0087] The office document review and summary generation module is used to generate summary units based on candidate semantic fragments, calculate the importance weight of summary units based on attention weight distribution, determine the review execution strategy for each summary unit, and generate review results by applying review rules during the execution of the review execution strategy.

[0088] The document summary update module is used to determine the summary review result based on the review results, perform partial updates of the summary unit based on the summary review result, and output the final summary result and document review result.

[0089] In Embodiment 2 of the present invention, while remaining otherwise unchanged from Embodiment 1, the method for performing a partial update of the digest unit based on the digest review result in Embodiment 1 further includes: determining a consistency anomaly state. The specific analysis method is as follows:

[0090] The system receives all generated summary-related review results and then iterates through each result. For each result, based on the summary unit identifier associated with the summary-related review result, it checks whether the unit has a pre-set reservation flag. If it has, it means that the summary unit was identified as containing core business elements that require mandatory review during initial generation. Its content has high business importance and high risk. Therefore, when any summary-related review result for such core units appears, regardless of the size or number of issues, it is considered an unacceptable major quality defect and must be immediately marked as a consistency anomaly to absolutely ensure the accuracy of key information. If it has not been set, it means that the summary unit was not determined to contain mandatory review elements in the initial evaluation. Its importance mainly depends on the dynamic weight calculated by the algorithm. In this case, it will not be immediately marked as an anomaly based on a single summary-related review result.

[0091] During the traversal, real-time statistics are performed to record how many results point to each summary unit. If the number of results pointing to the same unit is greater than or equal to the preset pointing threshold, it indicates that the summary unit has multi-dimensional and multi-faceted complex quality problems. This means that the summary unit fails to meet the standards in multiple core dimensions such as accuracy, completeness, and reliability. Its problems are no longer isolated accidental errors, but reflect deeper generation defects or deviations in the understanding of the original text. In this case, it is also marked as a consistency anomaly. Conversely, if the number of results pointing to the unit does not reach the pointing threshold, it means that the summary unit either did not trigger any summary-related review results, or only triggered a very small number of problems that may be occasional or minor. The current problems have not yet constituted serious defects that require triggering a full regeneration. Therefore, it will not be marked as a consistency anomaly just because of one or two results, and the unit will maintain its original state.

[0092] The initial value of the target threshold is usually determined based on business experience and trial operation data analysis. By analyzing historical review data, the distribution of the average number of anomalies triggered in a single round of review for a summary unit is used as a reference to set an initial value that can effectively identify compound grid defects. Then, a configuration record for the target threshold is created in the database through the expert configuration interface, specifying its parameter name, threshold value and applicable scenarios. This threshold can be dynamically adjusted and optimized according to the actual operation effect after going live. The strategy can be fine-tuned by updating the database record, thereby finding the best balance between accurately capturing problems and maintaining operational efficiency.

[0093] After completing the traversal and analysis of all abstract-related review results, a list driving the update is generated and an optimization loop is started: an empty list of abstract units in a consistency anomaly state is initialized. During the traversal and analysis process, if an abstract unit is marked immediately because it has been set with a retention flag or the number of results it points to has reached the pointing threshold, its unique identifier is added to this list. After the traversal is completed, this list contains the target set of all units that need to be reprocessed. All abstract units that have not been added remain unchanged, and their identifiers will not appear in the list. Using this list as the core input, subsequent local update operations such as rereading the original text and regenerating the abstract under the retention flag constraint are precisely performed only on the units in the list. This achieves that the iterative optimization resources are completely focused on units that have been proven to have quality defects, without interfering with other units that have passed the review and are in good condition. This achieves an efficient, targeted, and automated closed loop for improving abstract quality.

[0094] Embodiment 3 of the present invention, while remaining unchanged from Embodiment 1, further includes: updating and adjusting the review rule base, the specific analysis method of which is as follows:

[0095] After each round of review, all summary units marked as inconsistent in this round and their corresponding review result records are collected. The text features of all anomalous summary units, the types of review rules triggered, and their mapping deviation from the original text are clustered and statistically analyzed.

[0096] The K-means++ algorithm (k-means clustering algorithm) is used to cluster the text features of all abnormal summary units. The text content is represented by the term frequency inverse document frequency feature vector. The range of the number of clusters k is set and the optimal number of clusters is automatically determined by the silhouette coefficient. At the same time, the frequency statistics of the triggered review rule types and their mapping deviation with the original text are performed. Based on the clustering and statistical analysis results, the types of summary quality defects that occur frequently are identified.

[0097] The range of values ​​for k is pre-stored in the rule parameter table of the configuration database. This preset logic is determined by experts after statistical analysis of historical review data: the lower limit corresponds to the minimum classification dimension of the abstract quality defects, ensuring that at least three basic defect patterns—terminology bias, missing facts, and logical inconsistencies—can be distinguished; the upper limit is set according to the constraints of computing resource load and processing efficiency to prevent excessive clustering from leading to redundancy in the rule base; the database also stores a dynamic adjustment mechanism for this range. When the terminology complexity of a document in a certain domain exceeds the preset threshold, the upper limit can be automatically expanded to 15, and the contour coefficient traversal calculation module iterates and tests within the range or the expanded range, selecting the k value corresponding to the maximum contour coefficient as the optimal number of clusters to be output to the clustering engine for execution.

[0098] The core improvement of the K-means++ algorithm lies in its probabilistic selection of initial cluster centers, which significantly improves clustering quality and accelerates convergence. Specifically, the algorithm first randomly selects the first cluster center, then calculates the minimum distance from each data point to the selected center, and selects the next center point with a probability proportional to the square of the minimum distance. That is, the farther a point is from the existing center, the greater the probability of it being selected. This process is repeated until k initial centers are selected. Finally, the iterative optimization steps of the standard K-means algorithm are executed. This initialization strategy effectively avoids the local optimum problem caused by random initial centers in the traditional K-means algorithm, making the final clustering results more stable and the sum of squared errors smaller. It is particularly suitable for clustering analysis of text feature vectors.

[0099] Based on the statistical analysis results, we identified the types of frequently occurring abstract quality defects, including but not limited to: language usage deviations for specific domain terminology, missing information on key facts or data, redundant repetition across abstract units, and unclear logic of causal or conditional relationships.

[0100] Based on the identified defect types, the review rule base is dynamically adjusted. For language usage deviations in specific domain terms, specific language standardization rules containing correct terms, typical errors, and their contextual features will be generated. For missing information in key facts or data, mandatory checks on specific syntactic patterns will be added. For redundant repetitions across summary units, semantic similarity comparison rules will be added to detect and warn of highly repetitive content within a preset distance. For unclear causal or conditional relationships, the integrity check rules for logical connectors will be strengthened.

[0101] The updated review rule base is applied to the next round or targeted incremental review process. All abstract units are re-evaluated based on the updated rules. Based on the continuously accumulated review result data, the calculation model of the importance weight of abstract units and the threshold for the division of review execution strategies are optimized, thereby forming a closed loop from review to problem discovery to rule optimization to re-review, continuously improving the overall quality and accuracy of abstract generation and automatic review.

[0102] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automatic proofreading and management method for office documents based on text summarization algorithms, characterized in that: Includes the following steps: The document structure is parsed and original text location identifiers are assigned. The elements of interest for review are read and a set of elements of interest for review are generated. Semantic parsing is performed based on the original text location identifiers and an initial set of semantic fragments is generated, forming a set of candidate semantic fragments. Summarization units are generated based on candidate semantic fragments. The importance weight of the summarization units is calculated based on the attention weight distribution. The review and execution strategy for each summarization unit is determined. During the execution of the review and execution strategy, the review and execution rules are applied to generate the review and execution results. Based on the review results, the abstract review results are determined, the abstract unit is partially updated based on the abstract review results, and the final abstract results and document review results are output.

2. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The method for performing structural parsing based on office documents, assigning original text location identifiers, reading elements of interest for review, and generating a set of elements of interest for review is as follows: Receive office documents to be processed, perform structural parsing operations on the office documents, identify the chapter levels, paragraph boundaries and sentence order information in the documents, and assign a unique original text position identifier to each paragraph and sentence; After completing the document structure parsing, the definition information of the elements of interest for review is read from the preset review rule base; The definition information of the elements of interest for review includes element type code, element feature pattern and associated review rule identifier; Based on the business type of the current office document, a subset of matching business-related elements is selected from the element definition information. The selected element subset is instantiated, and the element feature pattern is compiled into an executable semantic matching template. A set of review-focused elements containing numerical information, time information, description of responsible parties, definitional statements, and cross-paragraph reference relationships is generated. A unique element identifier is assigned to each review-focused element, and a mapping index between the element identifier and the associated review rule identifier is established.

3. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The method for semantic parsing based on the original text location identifiers and generating an initial set of semantic fragments is as follows: The current sentence is segmented and part-of-speech tagging is performed using a word segmentation tool to identify named entities and keywords in the sentence; The grammatical structure of a sentence is parsed using a dependency parser, identifying the core components of subject, verb, and object and the modification relationship. The predicate and argument structure of the sentence are extracted based on semantic role labeling, and it is determined whether the sentence contains multiple independent semantic expressions. When a sentence is detected to contain multiple independent semantic expressions, the sentence is split into initial semantic segments based on predicate boundaries and argument integrity. Each initial semantic segment contains a complete subject-verb-object structure and expresses a single semantic event. Each initial semantic segment is assigned a segment number, and the corresponding original text position identifier is recorded. When a sentence is detected to contain only a single semantic expression, the entire sentence is treated as an initial semantic fragment. Sort all initial semantic fragments according to their original text position identifiers to form an initial semantic fragment set.

4. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The specific analysis method for forming the candidate semantic fragment set is as follows: Traverse the initial set of semantic segments. If there is a referential relationship between adjacent initial semantic segments, merge them into the same candidate semantic segment. If adjacent initial semantic segments involve the same review focus element, they are merged into the same candidate semantic segment; If adjacent initial semantic segments form a correspondence between conditional semantics and conclusion semantics, they are merged into the same candidate semantic segment. After merging, a set of candidate semantic segments is formed, and each candidate semantic segment corresponds to a set of continuous or non-continuous original text position identifiers.

5. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The process involves generating summarization units based on candidate semantic fragments, calculating the importance weight of each summarization unit based on the attention weight distribution, determining the review execution strategy for each summarization unit, and applying review rules to generate review results during the execution of the review execution strategy. The specific analysis method is as follows: Taking the set of candidate semantic segments as input, a summary generation process is performed on each candidate semantic segment to obtain the corresponding summary expression content. Each summary expression content is defined as a summary unit. A unique summary unit identifier is assigned to each summary unit, and a mapping relationship is established between the summary unit identifier and the candidate semantic segment, and between the summary unit identifier and its corresponding set of original text position identifiers. In the process of generating summary units, the semantic coverage score is analyzed based on the cross-attention weight distribution of the decoder, the semantic concentration score is analyzed based on the information entropy, and the importance weight of the generated summary units is fused together. Match the set of original text location identifiers corresponding to each abstract unit with the set of elements of concern for review: When the original text location corresponding to the abstract unit contains any element of review concern, a retention flag is set for that abstract unit; Based on the importance weight of the abstract unit, the processing level is divided into abstract units. When the importance weight of the abstract unit is greater than or equal to the preset abstract importance threshold, the review execution strategy of the abstract unit is determined as the first review execution strategy. When the importance weight of a summary unit is less than the preset importance threshold, the review execution strategy for that summary unit is determined as the second review execution strategy.

6. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 5, characterized in that: The specific analysis method for the first review and execution strategy is as follows: The basic scope of review is determined based on the set of original text location identifiers, and the definition keywords and conclusion keywords in the current scope are identified based on dependency parsing. Related paragraphs containing the same keywords are searched in the office document to form an expanded scope of the original text. Load all review rules corresponding to the set of review focus elements from the review rule base and bind them to the summary unit; The loaded review rules are executed line by line in the extended original text in a preset order, and the intermediate results of the rule judgment for each rule are recorded. For each rule, calculate the rule matching degree. If the rule matching degree is lower than the preset matching threshold or the necessary fields required by the rule are missing, the rule is determined to trigger a single point of failure. Record the intermediate result of the rule judgment, which includes the rule identifier, rule matching degree, and list of missing fields. Otherwise, it is determined that a single point of failure has not been triggered and the rule is recorded as passed. After completing the rule-by-rule verification, a cross-position consistency comparison operation is performed on the numerical expression, conditional expression and conclusion expression among multiple original text position identifiers within the extended original text range. If there are cross-position numerical inconsistencies, logical contradictions, or semantic similarity lower than the preset similarity threshold, it is determined that an association anomaly has been triggered; otherwise, it is determined that no association anomaly has been triggered. When any review rule determines that an intermediate result or consistency comparison result meets the abnormal triggering condition, a corresponding review result record is generated and associated with the summary unit identifier.

7. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 5, characterized in that: The second review and execution strategy is analyzed in the following way: The scope of basic review is determined based on the set of original text location identifiers; Load all review rules corresponding to the review focus elements involved in this summary unit from the review rule base and bind them to this summary unit; Execute all loaded review rules on each item of the original text within the scope of basic review, and record the intermediate results of the rule judgments; For each rule, calculate the rule matching degree. If the rule matching degree is lower than the preset matching threshold or the necessary fields required by the rule are missing, the rule is determined to trigger a single point of failure. Record the intermediate result of the rule judgment, which includes the rule identifier, rule matching degree, and list of missing fields. Otherwise, it is determined that a single point of failure has not been triggered and the rule is recorded as passed. When the intermediate result of any core review rule meets the abnormal triggering condition, the corresponding review result record is generated and associated with the summary unit identifier.

8. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The specific analysis method for determining the abstract review results based on the review results is as follows: Each input review result record is structured and parsed to extract the anomaly type, the original text location identifier that triggered the anomaly, and the associated summary unit identifier. Based on the parsed structured information, logical judgments including context missing judgment, semantic inconsistency judgment, and conflict description judgment are automatically executed. The context missing determination includes checking whether the references, definitions or conditions mentioned in the summary unit can be found in all the document locations associated with them. If not, the necessary context missing summary-related review results are generated. The semantic inconsistency determination includes generating a review result for a summary that is semantically inconsistent with the original text when the semantic similarity between the summary expression of the summary unit and the corresponding original text content in the set of original text position identifiers is lower than the similarity threshold. The conflict description determination includes checking whether there are contradictions in the values, conditions or conclusions between different original text positions associated with the same summary unit. If there are conflicting values, conditions or conclusions, the summary-related review results of the original text conflict description are generated.

9. The method for automatic review and management of office documents based on text summarization algorithm as described in claim 1, characterized in that: The process of performing partial updates to the summary unit based on the summary review results, and outputting the final summary result and document review result, is analyzed in the following way: If any summary-related review result is found, the summary unit is marked as a consistency anomaly. The original text location identifier set corresponding to the summary unit is reread. Under the condition of retaining the labeling constraints, the summary expression content of the summary unit is regenerated, and the mapping relationship between the summary unit and the original text is updated. Summary units not marked as having a consistency anomaly do not participate in the update; After completing the local update of the summary unit, the importance weight of the summary unit is recalculated based on the attention weight distribution to determine the review execution strategy for each summary unit. During the execution of the review execution strategy, the review rules are applied to generate the review results. The summary review result is determined based on the review results. When no new summary-related review results are generated in a complete loop, the loop process is terminated, and the final summary result and document review result are output.

10. An automatic office document review and management system based on text summarization algorithms, characterized in that: include: The document parsing module, the document review and summary generation module, and the document summary update module are all included. The office document parsing module is used to perform structural parsing based on the office document and assign original text location identifiers, read the elements of interest for review and editing, generate a set of elements of interest for review and editing, perform semantic parsing based on the original text location identifiers and generate an initial set of semantic fragments, and form a set of candidate semantic fragments. The office document review and summary generation module is used to generate summary units based on candidate semantic fragments, calculate the importance weight of summary units based on attention weight distribution, determine the review execution strategy for each summary unit, and generate review results by applying review rules during the execution of the review execution strategy. The document summary update module is used to determine the summary review result based on the review result, perform a partial update of the summary unit based on the summary review result, and output the final summary result and the document review result.

Images