Intelligent office document generation system and method

Abstract

The invention discloses an intelligent office document generation system and method, and relates to the technical field of artificial intelligence, the system comprises a user input interface module, a semantic intention analysis module, a document structure dynamic modeling module, a content generation and filling module and a format adaptation and output module; the method comprises the following steps: receiving an original instruction and a material of a user, extracting key information through multi-level semantic analysis, matching and dynamically generating a document skeleton, filling text contents segment by segment, automatically adapting to a typesetting format, and outputting a final document. Through deep semantic understanding and dynamic structure expansion, high-quality document intelligent generation without template dependence is realized, semantic coherence, logic preciseness and format normalization are remarkably improved, manual intervention cost is reduced, and the defects in the aspects of dynamic response, structure processing and scene adaptation in the prior art are overcome.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent office document generation system and method. Background Technology

[0002] With the rapid growth in demand for intelligent office solutions, document generation systems have played a crucial role in improving office efficiency and reducing manual operations. Traditional document generation solutions largely rely on preset templates, historical document matching, or natural language instruction parsing, with their core principles built upon static rules and limited corpora. However, user needs in modern office scenarios are highly dynamic and diverse: document content must respond to input semantics in real time, structure must adapt to multi-level professional logic, and format must be compatible with non-standardized scenarios. Static rule systems, lacking a deep understanding of user intent and contextual awareness, often result in generated content deviating from actual needs, structural misalignment, or misuse of terminology, exacerbating document rework rates and the risk of information distortion. Furthermore, users have increasingly stringent requirements for the accuracy, flexibility, and scenario adaptability of document generation, and the rigid architecture of existing methods struggles to support adaptive generation needs under complex semantic structures and dynamic contexts.

[0003] Among them, intelligent office document generation technology focuses on achieving automated content construction through semantic understanding and structural modeling. Its basic goal is to transform fragmented information or ambiguous instructions input by users into logically rigorous, formatted, and semantically complete formal documents. This direction encompasses multiple technical branches such as natural language processing, knowledge graph reasoning, dynamic template assembly, and context awareness, aiming to bridge the semantic gap between human expression and machine generation.

[0004] In existing technologies, some solutions rely on highly relevant historical documents for content extraction and reorganization. While this ensures local semantic coherence, the lack of a real-time semantic calibration mechanism leads to lag in response to user additions or modifications, resulting in generated content that is out of touch with the current context. Another type of solution, while incorporating voice or text command parsing, is limited by pre-defined personalized organizational rules. When handling complex documents containing nested logic, technical terminology, or cross-domain references, its structural parsing capabilities are insufficient, easily leading to paragraph mismatches or semantic gaps. More importantly, neither type of solution establishes a closed-loop mechanism for dynamic content analysis and adaptive structural generation. They cannot adjust the generation strategy in real time based on document type, user role, or scenario constraints, resulting in poor adaptability, low fault tolerance, and high reconstruction costs in diverse office environments. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent office document generation system and method to solve the problem that existing technologies rely on preset templates, historical documents or fixed rules, and lack the ability to deeply understand and structure the user's real-time input intent.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] An intelligent office document generation system includes:

[0008] The user input interface module is used to receive raw instructions and material content input by users through text input, voice input, or mixed mode.

[0009] The semantic intent parsing module is used to perform multi-level semantic parsing on the original instructions and material content to extract task type, key entities, logical relationships and constraints.

[0010] The document structure dynamic modeling module is used to match a basic framework from a preset structure template library and dynamically expand it to generate a complete document skeleton based on task type, key entities, logical relationships and constraints.

[0011] The content generation and filling module is used to fill in semantically coherent and logically rigorous main text content segment by segment based on the complete document skeleton by calling the domain knowledge base and the context-aware generation model.

[0012] The format adaptation and output module is used to automatically match the typesetting specifications and style templates according to the target application scenario, perform dynamic layout and format encapsulation, and output the final document.

[0013] The document structure dynamic modeling module includes a structure template matching unit, a structure incremental generation unit, and a structure validity verification unit. The structure template matching unit divides the template library according to industry domains and uses weighted similarity calculation to match the basic framework. The structure incremental generation unit dynamically inserts structure nodes based on entity and relationship information. The structure validity verification unit performs triple verification of integrity, hierarchical validity, and reference closure.

[0014] A method for generating intelligent office documents, comprising:

[0015] Step S1: Receive the user's original instructions and material content through the user input interface module. The original instructions and material content include text input, voice input, and mixed-mode input.

[0016] Step S2: Perform multi-level semantic parsing on the original instructions and material content to extract task type, key entities, logical relationships and constraints;

[0017] Step S3: Based on the task type, key entities, logical relationships and constraints, match the basic framework from the preset structure template library and dynamically expand to generate a complete document skeleton;

[0018] Step S4: Based on the complete document skeleton, the domain knowledge base and context-aware generation model are called to fill in the semantically coherent and logically rigorous main text content segment by segment.

[0019] Step S5: Automatically match layout specifications and style templates according to the target application scenario, and perform dynamic layout and format encapsulation;

[0020] Step S6: Output the final document that meets industry standards and user requirements.

[0021] The original instructions and material content undergo multi-level semantic parsing to extract task type, key entities, logical relationships, and constraints, including:

[0022] Perform word segmentation, part-of-speech tagging, and named entity recognition on the input text to output a tagged word sequence;

[0023] Construct a syntactic dependency tree structure to identify subject-verb-object attributive-adverbial-complement grammatical components and their interrelationships;

[0024] Based on the dependency structure, the core arguments of each predicate and their semantic roles are marked. The semantic roles include agent, patient, time, place, and manner.

[0025] The word sequence, dependency structure, and semantic role annotation results are input into a multilayer perceptron classifier to output document task category labels.

[0026] Task types, key entities, logical relationships, and constraints are matched with a basic framework from a pre-defined structure template library and dynamically expanded to generate a complete document skeleton, including:

[0027] Retrieve basic frameworks matching the task category from a structure template library categorized by industry sector, including government affairs, finance, law, education, healthcare, manufacturing, scientific research, and the internet.

[0028] Based on the entity list and relationship graph output by semantic parsing, sub-chapter, list item, table container or appendix area are dynamically inserted on the basic framework;

[0029] Perform integrity checks, hierarchical nesting validity verification, and reference relationship closure verification on the generated document skeleton.

[0030] The complete document skeleton calls upon the domain knowledge base and context-aware generation model to fill in semantically coherent and logically rigorous main text content segment by segment, including:

[0031] Based on each content node in the document skeleton, and combined with the domain knowledge graph, plan the core arguments, supporting evidence, and data citation paths that this paragraph should cover;

[0032] A sequence-to-sequence model based on an attention mechanism is used to transform the semantic planning results into natural language sentences;

[0033] During the generation process, the topic distribution, term usage frequency, and referential chains of the generated content are tracked in real time, and semantic coherence is ensured through implicit state caching and backtracking mechanisms.

[0034] Automatically match layout specifications and style templates based on the target application scenario to perform dynamic layout and formatting encapsulation, including:

[0035] Based on document type, industry, and user preferences, select font, font size, line spacing, indentation, header / footer, and chart style parameter sets from the style rule library;

[0036] Automatic pagination, element positioning, and complex structure numbering are applied to text content, tables, and images based on the selected style parameters.

[0037] It supports the automatic numbering and jump link generation of multi-level numbered lists, cross-references, footnotes, and endnotes.

[0038] Voice input is achieved through an end-to-end acoustic modeling structure with a sampling frequency of 16,000 Hz, a frame length of 25 milliseconds, and a frame shift of 10 milliseconds. Feature extraction uses Mel frequency cepstral coefficients combined with first-order and second-order differences to form a 45-dimensional feature vector.

[0039] Named entity recognition uses a conditional random field model, and the recognition scope includes personal names, place names, organization names, dates, times, currencies, percentages, product models, and legal clause numbers.

[0040] The sequence-to-sequence model uses a six-layer bidirectional transformer as the encoder and a six-layer unidirectional transformer as the decoder. The vocabulary size is 50,000, the embedding dimension is 768, and the number of attention heads is 12. An adversarial discriminator is introduced during training.

[0041] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0042] This invention achieves deep structured deconstruction of user's original input by constructing a semantic intent parsing module, thus eliminating the dependence on historical documents or fixed templates and significantly improving the system's adaptability to diverse inputs.

[0043] This invention supports flexible expansion from basic templates to complex incremental structures through a document structure dynamic modeling module, effectively solving the problem of automatic generation of multi-level logical documents. Through the context consistency maintenance mechanism in the content generation and filling module, it ensures the global coherence of long documents in terms of semantics, style and reference, avoiding the inconsistencies or terminology drift problems common in the prior art.

[0044] This invention achieves fully automatic adaptation and accurate output of professional-grade document formats through a dynamic typesetting engine in the format adaptation and output module, significantly reducing the cost of manual post-adjustment. Without relying on high-quality historical samples, it can automatically generate office documents with rigorous structure, accurate semantics, and standardized format based on real-time user input, comprehensively improving the efficiency and quality of document generation in intelligent office scenarios. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the architecture of an intelligent office document generation system proposed in this invention;

[0046] Figure 2 This is a schematic diagram of the process of an intelligent office document generation method proposed in this invention. Detailed Implementation

[0047] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0049] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0050] Please refer to Figure 1 , Figure 1 This is an architecture diagram of an intelligent office document generation system provided by an embodiment of the present invention, including:

[0051] The module includes a user input interface module, a semantic intent parsing module, a document structure dynamic modeling module, a content generation and filling module, and a format adaptation and output module.

[0052] The user input interface module is deployed on the client side to receive raw instructions and content input by users through text, voice, or mixed modes; it adopts a responsive web design, supports access from desktop and mobile browsers, has an input latency of less than 200 milliseconds, and a voice recognition response time of less than 1.5 seconds.

[0053] The semantic intent parsing module is deployed on a cloud-based inference server to perform multi-level semantic parsing on the original instructions and material content to extract task types, key entities, logical relationships, and constraints. It adopts containerized deployment, with each instance occupying eight gigabytes of memory, a CPU utilization rate of 35%, and an average processing latency of 0.8 seconds.

[0054] The document structure dynamic modeling module is deployed on the application server. It is used to match the basic framework from the preset structure template library based on task type, key entities, logical relationships and constraints, and dynamically expand to generate a complete document skeleton. It uses an in-memory database to cache the template library, with a query response time of less than 50 milliseconds and an average structure generation time of 0.3 seconds.

[0055] The document structure dynamic modeling module includes a structure template matching unit, a structure incremental generation unit, and a structure validity verification unit. The structure template matching unit divides the template library according to industry domains and uses weighted similarity calculation to match the basic framework. The structure incremental generation unit dynamically inserts structure nodes based on entity and relationship information. The structure validity verification unit performs triple verification of integrity, hierarchical validity, and reference closure.

[0056] The content generation and filling module is deployed on a dedicated graphics processor cluster. It is used to fill semantically coherent and logically rigorous text content segment by segment based on the complete document skeleton, calling the domain knowledge base and the context-aware generation model. Each generation request is allocated one graphics processor card, with a video memory usage of 6 gigabytes, and the average time for single-segment generation is 1.2 seconds.

[0057] The format adaptation and output module is deployed on the file server. It is used to automatically match the typesetting specifications and style templates according to the target application scenario, perform dynamic layout and format encapsulation, and output the final document. It supports concurrent export requests, with an average processing time of 0.5 seconds per request and a maximum concurrency of 200.

[0058] The modules communicate with each other via a Hypertext Transfer Protocol interface. The data format is a lightweight data exchange format, the interface response time is less than 100 milliseconds, and the error rate is less than one in ten thousand.

[0059] Please refer to Figure 2 , Figure 2 This is a flowchart of an intelligent office document generation method provided by an embodiment of the present invention, including the following steps:

[0060] Step S1: Receive the user's original instructions and material content through the user input interface module. The original instructions and material content include text, voice, or mixed mode input.

[0061] In the step of receiving the user's original instructions and material content, the system receives multimodal input information from the user through the user input interface module; this input interface module supports text input, voice input, and mixed-mode input.

[0062] Text input is completed through a standard character input interface, supporting three methods: copy and paste, drag and drop to import files, and real-time typing.

[0063] Voice input is accomplished through a built-in speech recognition engine, which adopts an end-to-end acoustic modeling structure and supports real-time transcription of three languages: Mandarin, Cantonese, and English. The sampling frequency is 16,000 Hz, the frame length is 25 milliseconds, the frame shift is 10 milliseconds, and the feature extraction uses Mel frequency cepstral coefficients combined with first-order and second-order differences to form a 45-dimensional feature vector.

[0064] Mixed-mode input allows users to submit both speech segments and text paragraphs simultaneously. The system automatically aligns timestamps and merges them into a unified semantic stream. All input content undergoes unified encoding conversion before entering subsequent processing modules to ensure the character set is in Unicode encoding format, avoiding garbled characters and character loss. Input content is encapsulated into a structured data packet containing five types of metadata: original character sequence, input timestamp, input device identifier, input language type, and input mode type.

[0065] Step S2: Perform multi-level semantic parsing on the original instructions and material content to extract task type, key entities, logical relationships and constraints;

[0066] In the multi-level semantic parsing step of the original input, the system calls the semantic intent parsing module to deconstruct the input data packet layer by layer. The semantic intent parsing module includes: lexical analysis unit, syntactic dependency analysis unit, semantic role labeling unit and intent classification unit.

[0067] The lexical analysis unit first performs word segmentation on the input text, employing a hybrid segmentation algorithm combining bidirectional maximum matching and a hidden Markov model. The dictionary contains 380,000 entries, covering general vocabulary, professional terminology, and emerging internet slang. The segmentation output is a sequence of words with part-of-speech tags. The tagging system uses the national standard part-of-speech tag set, comprising forty-two basic part-of-speech tags. Named entity recognition uses a conditional random field model, recognizing twelve types of entities including names of people, places, organizations, dates, times, currencies, percentages, product models, and legal clause numbers.

[0068] The syntactic dependency parsing unit receives the lexical analysis results, constructs a dependency syntax tree, and adopts a transition-based dependency parser. The feature template includes the current top word of the stack, the first word of the buffer, the root node words of the left and right subtrees of the generated dependency arcs, and their parts of speech. A total of 128 sparse features are extracted. The transition operation is predicted by a linear classifier, and finally a complete dependency structure tree is output, which labels the grammatical relationship type between each pair of words, such as subject-verb relationship, verb-object relationship, attributive-head relationship, adverbial-head relationship, prepositional-object relationship, etc.

[0069] Based on the dependency tree, the semantic role labeling unit locates the core argument of each predicate and labels its semantic role type, including sixteen roles such as agent, patient, instrument, time, place, cause, purpose, manner, result, and condition. The labeling model adopts a sequence labeling architecture that combines a bidirectional long short-term memory network with a conditional random field. The input is a concatenation of word vectors, part-of-speech vectors, and dependency path vectors, and the output is the semantic role label corresponding to each word.

[0070] The intent classification unit concatenates all the above analysis results into a high-dimensional feature vector, which is then input into a three-layer fully connected neural network classifier. The activation function is a modified linear unit, and the output layer uses a soft maximum function. The classification label space includes 36 document task types such as meeting minutes, project reports, draft contracts, notices and announcements, technical solutions, financial budgets, personnel appointments and removals, legal opinions, scientific research papers, and product manuals.

[0071] Step S3: Based on the task type, key entities, logical relationships and constraints, match the basic framework from the preset structure template library and dynamically expand to generate a complete document skeleton;

[0072] In the step of matching the basic framework from the structure template library based on the parsing results and dynamically expanding it to generate a complete document skeleton, the system calls the document structure dynamic modeling module, which includes: a structure template matching unit, a structure incremental generation unit, and a structure validity verification unit.

[0073] The structural template matching unit maintains a structured template database, which is divided into eight main categories based on industry sectors: government affairs, finance, law, education, healthcare, manufacturing, scientific research, and the internet. Each main category has several subcategories, such as the law category which includes subcategories for civil contracts, criminal defense, intellectual property, and labor arbitration. Each template defines a complete chapter hierarchy, including naming rules for first-level, second-level, and third-level headings, paragraph order constraints, required field identifiers, optional field identifiers, maximum number of repeated fields, and logical dependencies between fields. The matching process uses a weighted similarity calculation, with the following weights: 40% for task type matching, 30% for key entity matching, 20% for logical relationship matching, and 10% for constraint matching. The template with the highest matching score is selected as the basic framework.

[0074] The incremental structure generation unit, based on the entity list and relationship graph output by the semantic parsing module, dynamically inserts new structural nodes within the basic framework. If multiple time entities are detected and have a chronological relationship, a "Time Node" sub-section is automatically inserted; if multiple numerical entities are detected and have a comparison relationship, a "Data Comparison Table" container is automatically inserted; if legal clause reference entities are detected, they are automatically inserted into the "Legal Basis" appendix area. The insertion operation follows preset structural expansion rules, including insertion position constraints, insertion depth limits, and the upper limit on the number of sibling nodes after insertion.

[0075] The structural validity verification unit performs three checks on the generated document skeleton: the first is completeness verification, which checks whether all required nodes have been generated, and if any are missing, it triggers default value filling or user prompts; the second is hierarchy validity verification, which checks whether the nesting of parent and child nodes conforms to the preset depth rules, and if there is illegal nesting, it automatically promotes or demotes the node level; the third is reference closure verification, which checks whether all cross-reference markers have corresponding target nodes, and if the reference is not closed, it automatically creates placeholder nodes or deletes invalid references.

[0076] Step S4: Based on the complete document skeleton, the domain knowledge base and context-aware generation model are called to fill in the semantically coherent and logically rigorous main text content segment by segment.

[0077] In the step of filling in semantically coherent main text content segment by segment based on the document skeleton by calling the domain knowledge base and the generation model, the system calls the content generation and filling module, which includes: paragraph-level semantic planning unit, sentence-level language generation unit and context consistency maintenance unit.

[0078] The paragraph-level semantic planning unit receives each content node from the document skeleton and, combined with a pre-loaded domain knowledge graph, plans the core arguments, supporting evidence, and data citation paths that the paragraph should cover. The knowledge graph adopts an attribute graph model, with node types including five categories: concepts, entities, events, rules, formulas, and cases, and edge types including six categories: hierarchical relationships, synonymous relationships, causal relationships, temporal relationships, citation relationships, and comparative relationships. The planning process uses a graph traversal algorithm, starting from the core concept corresponding to the current node, expanding the range by two hops along causal and temporal edges, extracting relevant entities and events as supporting materials, and generating a semantic planning vector with a dimension of 256, including four sub-dimensions: argument strength value, expected value of evidence quantity, data citation density value, and professional terminology density value.

[0079] The sentence-level language generation unit employs an attention-based sequence-to-sequence model. The encoder is a six-layer bidirectional transformer, and the decoder is a six-layer unidirectional transformer. The vocabulary size is 50,000 words, the embedding dimension is 768, and the number of attention heads is 12. The input is a concatenation of the semantic planning vector and the current paragraph title text, and the output is a sequence of natural language sentences. An adversarial discriminator is introduced during model training. The discriminator is a three-layer convolutional neural network. The input is the word vector sequences of the generated sentence and the real sentence, and the output is the true / false probability. The generator and discriminator are jointly trained, and the loss function is a weighted sum of cross-entropy loss and adversarial loss with a weight ratio of 7:3.

[0080] The context consistency maintenance unit tracks the topic distribution, term usage frequency, and referential chains of the generated content in real time during the generation process. Topic distribution uses a latent Dirichlet assignment model with 16 topics; the topic distribution vector is updated for each new sentence generated. Term usage frequency is maintained using a dynamic dictionary, recording the cumulative occurrences of each term in the current document. If the term frequency in a new sentence deviates from the historical average by more than 20%, term replacement or explanation insertion is triggered. Referential chains employ a coreference resolution algorithm, maintaining a referential linked list that records the antecedent of each pronoun or omitted subject. If a new sentence has ambiguous or broken referential structures, explicit noun phrases are inserted for clarification. All generated content undergoes grammar and spelling checks before output. Grammar checks use a rule-based error detection engine, covering 28 common errors such as subject-verb agreement, tense consistency, missing articles, and preposition misuse. Spelling checks use an edit distance algorithm with a dictionary of 400,000 words and a maximum edit distance of 2; erroneous words are automatically replaced with the most similar words in the dictionary.

[0081] Step S5: Automatically match layout specifications and style templates according to the target application scenario, and perform dynamic layout and format encapsulation;

[0082] In the process of automatically matching layout styles and executing dynamic layout based on the target application scenario, the system calls the format adaptation and output module, which includes: style rule matching unit, dynamic layout engine and multi-format export unit.

[0083] The style rule matching unit maintains a style rule library, with rules indexed by three dimensions: document type, industry attribute, and user preference. The document type dimension includes the aforementioned thirty-six task tags; the industry attribute dimension includes eight main categories; and the user preference dimension includes twelve configuration items such as font preference, font size preference, line spacing preference, margin preference, and heading style preference. The matching process uses a three-level priority: user preference configuration has the highest priority, followed by industry attribute, and then document type. The matching result output is a set of style parameters, including forty-two parameters such as body text font, body text font size, line spacing multiple, paragraph before margin, paragraph after margin, first-level heading font, first-level heading font size, first-level heading line spacing, second-level heading font, second-level heading font size, second-level heading line spacing, header content, footer content, page number format, chart title font, chart title font size, chart border line style, and chart fill color.

[0084] The dynamic layout engine receives a set of style parameters and the generated text content, and executes automatic layout. The layout process uses a constraint-solving algorithm, treating the page as a two-dimensional canvas and text paragraphs, tables, and images as variable-sized rectangles. Constraints include minimum line height, maximum line width, paragraph coherence, table row and column alignment, image aspect ratio maintenance, and reserved areas for headers and footers. The solver uses a simulated annealing algorithm with an initial temperature of 1000, a cooling coefficient of 0.95, and 500 iterations. The objective function is to maximize page utilization and minimize blank areas. The engine supports complex element processing: multi-level numbered lists use a recursive numbering algorithm, with each level's numbering format independently configurable, supporting four numbering systems: Roman numerals, Arabic numerals, Chinese numerals, and English letters; cross-referencing uses an anchor point positioning mechanism, generating a unique identifier for each referential node and inserting a hyperlink at the reference point; footnotes and endnotes use a floating container mechanism, with footnote containers fixed at the bottom of the page and endnote containers concentrated at the end of the document, using a continuously increasing numbering method and supporting automatic renumbering. After the layout is completed, an intermediate format document is generated, which contains all style information and layout coordinates.

[0085] Step S6: Output the final document that meets industry standards and user requirements;

[0086] In the step of outputting a final document that conforms to the specifications, the multi-format export unit converts the intermediate format document into the target format. Supported export formats include Portable Document Format, Rich Text Format, Hypertext Markup Language Format, and native office software formats. Portable Document Format export uses a vector graphics engine to convert all text, tables, and images into path objects, embedding font subsets, setting the compression level to five, and controlling the output file size to within 80% of the source file. Rich Text Format export uses a markup language converter to map style parameters to rich text control words, supporting basic styles such as bold, italics, underline, color, font, font size, and alignment, but not supporting complex layouts such as multi-column layouts or text boxes. Hypertext Markup Language Format export uses semantic tag mapping: first-level headings are mapped to h-tags, second-level headings to h-tags, paragraphs to p-tags, lists to ul or ol tags, tables to table tags, and images to img tags. It supports cascading style sheets and style embedding, with style scope limited to the current document. Exporting to native office software formats utilizes an application programming interface (API) call. This initiates a background office software process, uses object linking and embedding automation interfaces to set document content and styles item by item, saves the file in native format, and then terminates the process, ensuring format compatibility and functional integrity. All exported files are appended with a digital watermark. The watermark content is a hash value of the generated timestamp and user identifier, embedded using finely adjusted character spacing, making it invisible to the naked eye and resistant to printing and scanning attacks.

[0087] During system operation, users access the system through a browser, input the original instructions and material content, and click the generate button. After receiving the input, the system sequentially calls the semantic intent parsing module, the document structure dynamic modeling module, the content generation and filling module, and the format adaptation and output module, and finally returns the generated document file to the user via a download link. The system supports breakpoint resume generation. If the network is interrupted midway, the user can resubmit the same task identifier, and the system can continue execution from the breakpoint to avoid duplicate calculations. The system also supports version comparison. Users can submit two generated documents, and the system automatically compares the differences, outputs addition, deletion, and modification tags and a similarity score. The similarity calculation uses the cosine similarity algorithm, with a threshold set at 0.85. Values ​​higher than this are considered highly similar.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An intelligent office document generation system, characterized in that, include: The user input interface module is used to receive raw instructions and material content input by users through text input, voice input, or mixed mode. The semantic intent parsing module is used to perform multi-level semantic parsing on the original instructions and material content to extract task type, key entities, logical relationships and constraints. The document structure dynamic modeling module is used to match a basic framework from a preset structure template library and dynamically expand it to generate a complete document skeleton based on the task type, key entities, logical relationships and constraints. The content generation and filling module is used to fill in semantically coherent and logically rigorous main text content segment by segment based on the complete document skeleton by calling the domain knowledge base and the context-aware generation model. The format adaptation and output module is used to automatically match the typesetting specifications and style templates according to the target application scenario, perform dynamic layout and format encapsulation, and output the final document.

2. The intelligent office document generation system according to claim 1, characterized in that, The document structure dynamic modeling module includes a structure template matching unit, a structure incremental generation unit, and a structure validity verification unit. The structure template matching unit divides the template library according to industry domains and uses weighted similarity calculation to match the basic framework. The structure incremental generation unit dynamically inserts structure nodes based on entity and relationship information. The structure validity verification unit performs triple verification of integrity, hierarchical validity, and reference closure.

3. A method for generating intelligent office documents, applied to the intelligent office document generation system according to any one of claims 1-2, characterized in that, include: Step S1: Receive the user's input of raw instructions and material content through the user input interface module. The raw instructions and material content include text input, voice input, and mixed-mode input. Step S2: Perform multi-level semantic parsing on the original instructions and material content to extract task type, key entities, logical relationships and constraints; Step S3: Based on the task type, key entities, logical relationships and constraints, match the basic framework from the preset structure template library and dynamically expand to generate a complete document skeleton; Step S4: Based on the complete document skeleton, call the domain knowledge base and context-aware generation model to fill in the semantically coherent and logically rigorous main text content segment by segment; Step S5: Automatically match layout specifications and style templates according to the target application scenario, and perform dynamic layout and format encapsulation; Step S6: Output the final document that meets industry standards and user requirements.

4. The intelligent office document generation method according to claim 3, characterized in that: The original instructions and material content undergo multi-level semantic parsing to extract task type, key entities, logical relationships, and constraints, including: Perform word segmentation, part-of-speech tagging, and named entity recognition on the input text to output a tagged word sequence; Construct a syntactic dependency tree structure to identify subject-verb-object attributive-adverbial-complement grammatical components and their interrelationships; Based on the dependency structure, the core arguments and semantic roles of each predicate are labeled, including agent, patient, time, place, and manner; The word sequence, dependency structure, and semantic role annotation results are input into a multilayer perceptron classifier to output document task category labels.

5. The intelligent office document generation method according to claim 3, characterized in that: The task type, key entities, logical relationships, and constraints are matched with a basic framework from a preset structure template library and dynamically expanded to generate a complete document skeleton, including: Retrieve basic frameworks matching the task category from a structure template library categorized by industry sector, including government affairs, finance, law, education, healthcare, manufacturing, scientific research, and the internet. Based on the entity list and relationship graph output by semantic parsing, sub-chapter, list item, table container or appendix area are dynamically inserted on the basic framework; Perform integrity checks, hierarchical nesting validity verification, and reference relationship closure checks on the generated document skeleton.

6. The intelligent office document generation method according to claim 3, characterized in that: The complete document skeleton calls upon the domain knowledge base and context-aware generation model to fill in semantically coherent and logically rigorous main text content segment by segment, including: Based on each content node in the document skeleton, and combined with the domain knowledge graph, plan the core arguments, supporting evidence, and data citation paths that each paragraph should cover; A sequence-to-sequence model based on an attention mechanism is used to transform the semantic planning results into natural language sentences; During the generation process, the topic distribution, term usage frequency, and referential chains of the generated content are tracked in real time, and semantic coherence is ensured through implicit state caching and backtracking mechanisms.

7. The intelligent office document generation method according to claim 3, characterized in that: The automatic matching of layout specifications and style templates based on the target application scenario to perform dynamic layout and format encapsulation includes: Based on document type, industry, and user preferences, select font, font size, line spacing, indentation, header / footer, and chart style parameter sets from the style rule library; Automatic pagination, element positioning, and complex structure numbering are applied to text content, tables, and images based on the selected style parameters. It supports the automatic numbering and jump link generation of multi-level numbered lists, cross-references, footnotes, and endnotes.

8. The intelligent office document generation method according to claim 3, characterized in that: The voice input is achieved through an end-to-end acoustic modeling structure with a sampling frequency of 16000 Hz, a frame length of 25 milliseconds, a frame shift of 10 milliseconds, and feature extraction using Mel frequency cepstral coefficients combined with first-order and second-order differences to form a 45-dimensional feature vector.

9. The intelligent office document generation method according to claim 4, characterized in that: The named entity recognition uses a conditional random field model, and the recognition scope includes personal names, place names, organization names, dates, times, currencies, percentages, product models, and legal clause numbers.

10. The intelligent office document generation method according to claim 6, characterized in that: The sequence-to-sequence model uses a six-layer bidirectional transformer as the encoder and a six-layer unidirectional transformer as the decoder. The vocabulary size is 50,000, the embedding dimension is 768, and the number of attention heads is 12. An adversarial discriminator is introduced during the training process.

Images