Configurable document intelligent template mapping and automatic content generation system

Abstract

The invention discloses a configurable document intelligent template mapping and automatic content generation system, and relates to the technical field of artificial intelligence and natural language processing. The system comprises a data analysis layer, a knowledge base construction engine, a configurable intelligent template arrangement center, a retrieval enhancement and reasoning generation engine and a verification and backfilling module. The method comprises the following steps: converting an image, an audio and a table into a uniform plain text sequence and constructing a vectorized knowledge base; performing logic arrangement on dynamic variables in the document template by utilizing a directed acyclic graph; acquiring a context credential based on a hybrid retrieval strategy, and driving the large language model to generate content; performing fact implication and entity consistency verification through a natural language reasoning model; according to the method, the problems of difficulty in multi-source data fusion, disordered document logic generation and illusion of the generative AI are effectively solved, and high-precision and logic self-consistent automatic document operation is realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and natural language processing technology, and specifically relates to a configurable intelligent document template mapping and automated content generation system. Background Technology

[0002] In professional fields such as public security policing, judicial litigation, administrative approval, and large enterprise auditing, document preparation is a high-frequency, rigorous, and massive core business. This work typically involves the comprehensive processing of massive amounts of case files, on-site investigation records, interrogation transcripts, and various statistical reports. With the advancement of digital transformation, although the electronic storage of basic data has become widespread, the process of transforming multi-source, heterogeneous raw data into standardized legal documents still relies heavily on manual operations. This is not only time-consuming and labor-intensive but also fails to meet the extremely high requirements of timeliness and standardization in modern government and legal work.

[0003] Current document generation technologies mainly fall into two categories: First, traditional rule-based template filling techniques, which use preset placeholders for simple database field mapping. This method is only suitable for structured data and cannot handle semantically complex unstructured text. Second, deep learning-based natural language processing techniques, which utilize OCR (Optical Character Recognition) to extract document text or ASR (Automatic Speech Recognition) to transcribe audio recordings, supplemented by keyword matching for information extraction. In recent years, with the development of Large Language Models (LLM), generative AI-based assisted writing tools have emerged. These tools can generate paragraph text based on prompts and, to some extent, summarize and rewrite text content, improving the flexibility of document drafting.

[0004] This reveals the following shortcomings in the practical application of existing technologies: 1. The lack of a unified deep semantic fusion method for multi-source heterogeneous data leads to severe information silos, making it impossible to perform cross-modal evidence verification during document generation; 2. General-purpose large language models are prone to producing illusions when generating serious documents and lack an effective logical consistency verification mechanism, resulting in the accuracy of the generated document content failing to meet the rigor requirements of the law; 3. Existing systems lack the ability to arrange the internal logical dependencies of complex documents, failing to guarantee the self-consistency of the generated content in terms of sequence and causal logic. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the problems in related technologies, the present invention provides a configurable intelligent document template mapping and automated content generation system to overcome the above-mentioned technical problems existing in the existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Data parsing layer: Used to receive heterogeneous data from multiple sources, identify data types through file header features, and process and generate a unified plain text sequence with metadata tags; Knowledge base building engine: used to segment a unified plain text sequence into blocks to obtain overlapping text blocks; to perform vectorization encoding and entity extraction operations on the overlapping text blocks to obtain high-dimensional semantic vectors and inverted indexes, and store them in a vector database to form a global semantic vectorized knowledge base; Configurable intelligent template orchestration center: used to configure variable generation metadata for each dynamically generated variable in the dynamic generation variable area of ​​the target document; construct a directed acyclic graph based on the variable generation metadata, calculate the execution priority sequence of the dynamically generated variables, and generate a template configuration description file; The retrieval enhancement and reasoning generation engine is used to perform hybrid retrieval operations in the global semantic vectorized knowledge base according to the template configuration description file to obtain contextual credentials; based on the contextual credentials, it constructs structured prompt words and inputs them into a large language model to obtain variable-generated content; Verification and backfilling module: Used to perform entity consistency verification and illusion detection on the variable-generated content; if the detection passes, the variable-generated content is backfilled according to the original template style, and an editable target document is output; Otherwise, return to the retrieval enhancement and inference generation engine to obtain new variable generation content until the detection passes; Preferably, the data parsing layer includes: The OCR intelligent recognition subunit is used to extract features from image data; perform sequential text recognition based on the features of image data; for sequential text in low-confidence areas, call the image enhancement algorithm for binarization and noise reduction retry; at the same time, it recognizes red stamps, handwritten signatures and table lines in the image, and outputs structured JSON data containing text content, coordinate position and font attributes. The ASR speech transcription and role separation subunit is used to perform VAD speech activity detection on audio data, remove silent segments, and obtain processed audio data; convert the processed audio data into text; calculate the voiceprint feature distance of different speech segments, classify segments with a distance less than a preset clustering threshold as the same speaker, and output dialogue text with timestamps and role tags. The structured data linearization sub-unit is used to read Excel or CSV files, identify header rows, traverse data rows, convert the content of each cell into a header name, and record the row number index of that cell as metadata. Preferably, the specific logic of the OCR intelligent recognition subunit when processing table lines is as follows: Predict horizontal and vertical line masks in an image using the U-Net semantic segmentation network; Perform Hough transform on the mask to detect lines and obtain the equations of all lines; based on the equations of all lines, calculate the intersection points of the lines and construct the grid topology of the table. Map the center coordinates of the text box recognized by OCR to the grid topology to determine the row and column indices of the text; Preferably, the knowledge base building engine includes: The intelligent segmentation subunit is used to set the basic sliding window size and the overlap step size; based on the basic sliding window size and the overlap step size, the uniform plain text sequence is segmented to obtain overlapping text blocks; during the segmentation process, if the paragraph end character output by the OCR intelligent recognition subunit or the speaker switching mark output by the ASR speech transcription and role separation subunit is detected, the current window is forcibly truncated to generate the overlapping text blocks. A hybrid index construction subunit is used to extract key entities from overlapping text blocks; the key entities are used as inverted index keywords; each text block is vectorized to generate a high-dimensional semantic vector as a vector index; a dual-path retrieval structure combining the inverted index and the vector index is constructed. Preferably, the configurable intelligent template orchestration center includes: The variable dependency graph building unit is used to parse all dynamically generated variables configured by the user and construct a directed acyclic graph based on the context dependencies in the variable generation metadata of the dynamically generated variables. The topology sorting and cycle detection unit is used to execute the Kahn algorithm on a directed graph composed of variables; it counts the in-degree of all nodes and adds variables with an in-degree of 0 to the execution queue; for each dynamically generated variable processed, it decrements the in-degree of its downstream nodes by 1; if there are still nodes that have not entered the queue after the Kahn algorithm ends, it is determined that there is a circular dependency and an error message is output; otherwise, it outputs a linear optimal execution path sequence. Instruction parameterization encapsulation unit, used to embed metadata filters in natural language generation instructions that allow users to dynamically generate variables in the optimal execution path sequence; Preferably, the retrieval enhancement and reasoning generation engine includes: The hybrid retrieval strategy execution subunit is used to calculate the cosine similarity between the query vector of dynamically generated variables and the vectors in the global semantic vectorized knowledge base to obtain a semantic score; calculate the matching degree between the query keywords and the inverted index to obtain a keyword score; perform a weighted summation of the semantic score and the keyword score according to the weighted summation formula to obtain the final score; and select the top-K overlapping text blocks with the highest final scores as contextual evidence. The dynamic prompt word assembly subunit constructs a prompt word template that includes system instructions, known information, historical generation, current task, and constraint conditions based on the highest-scoring Top-K text blocks. The variable generation content unit is used to calculate the total number of tokens for the prompt word template. If the total number of tokens exceeds the context limit of the large language model, text blocks are gradually discarded from low to high according to the final score until the length requirement is met, and the final prompt word template is obtained. The final prompt word template is input into the large language model to obtain the variable generation content. Preferably, the dynamic adjustment mechanism for the weight coefficients of the weighted summation formula in the hybrid retrieval strategy execution subunit includes: Obtain the natural language generation instructions corresponding to the current dynamically generated variables; analyze the intent type of the natural language generation instructions; if the instructions contain semantic verbs, determine that it is an abstract task, set the weight coefficient to favor semantic scoring, and focus on semantic vector retrieval; If the natural language generation instruction contains keyword-type verbs, it is judged as an exact fact task, and the weight coefficient is set to favor keyword scoring, emphasizing keyword inverted index retrieval. If the natural language generation instruction contains entity-constrained nouns, entity filtering is forcibly enabled, and similarity calculation is performed only within the text block containing the corresponding entity; the weight coefficient is set to favor keyword scoring. Preferably, the verification and backfilling module includes: The factual implication calculation subunit is used to take the generated variable content as an assumption and the concatenation of contextual evidence as a premise; it calls the natural language inference model to calculate the implication probability of the premise to the assumption; if the implication probability is lower than a preset threshold, it is judged as a potential illusion, and the system automatically adds a mark that needs to be manually verified after the variable. The entity consistency review subunit is used to extract the entity set in the variable generation content and the entity set in the retrieval credentials; calculate the entity conflict rate; if the generated text contains proper nouns that do not exist in the retrieval credentials and the entity conflict rate is greater than 0, it is judged as an error of fabrication, and the system returns to the dynamic prompt word assembly subunit, and adds a negative constraint that strictly prohibits the fabrication of entities that do not appear, until the variable generation content that passes the test is obtained. The format lossless backfill sub-unit is used to parse Word templates based on the Office Open XML standard; locate the Run node where the dynamically generated variable placeholders are located; write the variable generation content that passes the detection according to the original style attributes of the Run node; if the variable generation content contains table data, it automatically clones the table row structure in the document for dynamic filling. Preferably, the processing procedure for cases where the probability of implicating something is lower than a preset threshold includes: A pre-constructed dataset of judicial document verification containing positive and negative samples was built; the RoBERTa model was fine-tuned and trained to obtain a pre-trained RoBERTa model. During the inference phase, if the implied probability obtained through the RoBERTa model is lower than the implied probability threshold, the prompt word template, the variable generation content, and the prompt "The above variable generation content may contain factual errors. Please check the context and correct them" are concatenated and then input into the large language model for secondary generation. If the implied probability of the second-generation variable generation content is still lower than the implied probability threshold, the second generation result is retained and highlighted to prompt manual intervention. Preferably, the configurable document intelligent template mapping and automated content generation system further includes a privacy desensitization and security audit layer: Before segmenting and vectorizing the unified plain text sequence, sensitive information in the unified plain text sequence is identified using regular matching and NER model. The FPE algorithm is used to replace sensitive information with ciphertext while preserving the original format characteristics. Before backfilling the generated variable content, determine whether to decrypt and restore the ciphertext based on the current user's permission level; if permissions are insufficient, output the masked text. Record the token consumption, retrieved context credential ID, and generated text fingerprint during the generation of all dynamically generated variables, and store them in an immutable audit log blockchain.

[0007] (III) Beneficial Effects The present invention has the following beneficial effects: This invention constructs a multi-source heterogeneous data parsing and normalization layer, combined with OCR, ASR, and multimodal fusion technologies, to uniformly transform data of different formats such as images, audio, and spreadsheets into plain text sequences with metadata tags. It breaks down information silos between different media and significantly improves the recognition accuracy and structuring of the original data through targeted preprocessing algorithms (such as low-confidence retries, table topology restoration, and voiceprint clustering), providing a high-quality data foundation for subsequent intelligent processing.

[0008] This invention utilizes a configurable intelligent template arrangement center to construct a variable dependency graph using a directed acyclic graph. By calculating the execution priority of dynamically generated variables through a topological sorting algorithm, it ensures that the document generation process follows a strict business logic order (e.g., first generating a case description, and then generating liability determination based on the description). This mechanism ensures that the generated document is no longer a messy pile of text, but a logically rigorous and coherent whole document, effectively solving the problem of logical confusion in the writing of complex documents.

[0009] This invention introduces a dual verification mechanism based on retrieval-enhanced generation and natural language reasoning. On the one hand, it accurately locates contextual evidence through a hybrid retrieval strategy, suppressing the divergence of large models. On the other hand, it performs factual implication calculation and entity consistency review on the generated content, constructing a rigorous "anti-illusion" closed loop. This mechanism can automatically identify and correct factual errors and fictitious entities in the generated content, significantly improving the authenticity and accuracy of the document content and meeting the stringent requirements for document quality in professional fields.

[0010] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of a configurable document intelligent template mapping and automated content generation system according to the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0014] To address the technical problems raised in the background section, please refer to [link / reference]. Figure 1 This invention provides a configurable intelligent document template mapping and automated content generation system, comprising: Data parsing layer: Used to receive heterogeneous data from multiple sources, identify data types through file header features, and process and generate a unified plain text sequence with metadata tags; Knowledge base building engine: used to segment a unified plain text sequence into blocks to obtain overlapping text blocks; to perform vectorization encoding and entity extraction operations on the overlapping text blocks to obtain high-dimensional semantic vectors and inverted indexes, and store them in a vector database to form a global semantic vectorized knowledge base; A configurable intelligent template arrangement center is used to configure variable generation metadata for each dynamically generated variable in the target document's dynamically generated variable area; it constructs a directed acyclic graph based on the variable generation metadata, calculates the execution priority sequence of the dynamically generated variables, and generates a template configuration description file; the variable generation metadata includes natural language generation instructions, data source constraints, and contextual dependencies for each dynamically generated variable; the variable generation metadata is a structured configuration file used to fully define all generation parameters of a single dynamically generated variable; the data structure of the variable generation metadata includes at least: generation instructions, i.e., a natural language description defining the task; data type, i.e., specifying the format of the output content (such as "text paragraph", "list item", "amount", "date"); data source constraints, i.e., limiting the scope of the knowledge base to be searched or the metadata filtering conditions; dependency list, i.e., explicitly declaring the set of other variable IDs that must be completed before this variable is generated; and output format constraints, such as rules for word count, tone, and the requirement for key entities to appear. The retrieval enhancement and reasoning generation engine is used to perform hybrid retrieval operations in the global semantic vectorized knowledge base according to the template configuration description file to obtain contextual credentials; based on the contextual credentials, it constructs structured prompt words and inputs them into a large language model to obtain variable-generated content; Verification and backfilling module: Used to perform entity consistency verification and illusion detection on the variable-generated content; if the detection passes, the variable-generated content is backfilled according to the original template style, and an editable target document is output; Otherwise, return to the retrieval enhancement and inference generation engine to obtain new variable generation content until the detection passes; The above embodiments break down data silos by using OCR and ASR combined with multimodal fusion technology, ensure the logical order of document generation through configurable variable dependency graphs, and solve the problems of low efficiency in traditional document writing, difficulty in heterogeneous data fusion, error-prone content generation, and "illusion" by using retrieval-enhanced generation and natural language reasoning verification mechanisms. This results in document writing with high accuracy and logical correctness.

[0015] The data parsing layer in the above embodiments includes: The OCR intelligent recognition subunit is used to extract features from image data; perform sequential text recognition based on the features of image data; for sequential text in low-confidence areas, call the image enhancement algorithm for binarization and noise reduction retry; at the same time, it recognizes red stamps, handwritten signatures and table lines in the image, and outputs structured JSON data containing text content, coordinate position and font attributes. The specific logic of the OCR intelligent recognition subunit in the above embodiment when processing table lines is as follows: Predict horizontal and vertical line masks in an image using the U-Net semantic segmentation network; Perform Hough transform on the mask to detect lines and obtain the equations of all lines; based on the equations of all lines, calculate the intersection points of the lines and construct the grid topology of the table. Map the center coordinates of the text box recognized by OCR to the grid topology to determine the row and column indices of the text; In specific implementation, the aforementioned OCR intelligent recognition subunit function is as follows: After receiving the on-site investigation photos (such as those containing license plates and brake marks) or scanned PDF files uploaded by the user, the system uses the ResNet-50 backbone network to extract high-dimensional feature maps of the images; ResNet-50 contains five stages of convolutional layers, and introduces skip connections through residual blocks, effectively solving the gradient vanishing problem of deep networks, and finally outputs feature vectors through global average pooling layers; ResNet-50 is pre-trained using a large-scale natural scene image dataset to learn general visual features; in this system, it is further fine-tuned using massive amounts of document images, scanned documents, and photographed text images to optimize its adaptability to document layout, fonts, and complex backgrounds; The feature maps are input into the CRNN network for sequential text recognition. The CRNN network extracts feature sequences through convolutional layers, predicts label distribution through a bidirectional LSTM network, and finally outputs text sequences through temporal classification. The CRNN is trained end-to-end using a large number of labeled text line images and their corresponding transcribed texts. Its data usually comes from synthetic text datasets and cropped images of documents and invoices in real-world scenarios to ensure accurate recognition of printed, handwritten, and blurred text. Before use, the confidence threshold for sequential text is set to 0.75. This 0.75 is derived from statistical analysis of large-scale historical document recognition data. When the confidence is below 0.75, the character recognition error rate increases exponentially, requiring enhancement processing. For regions with an average confidence P < 0.75 in the recognition results (e.g., blurry license plates or worn text captured at night), the system automatically triggers an image enhancement process, using adaptive histogram equalization to enhance contrast and Otsu binarization algorithm for processing, followed by secondary recognition. If the confidence of the secondary recognition is still below 0.75, manual review is marked in the output JSON. For scanned documents containing tables (such as handwritten on-site confirmation forms), the U-Net semantic segmentation network is used to predict horizontal and vertical line masks; the Hough transform is performed on the masks to extract the line equations; the intersections of all horizontal and vertical lines are calculated to construct an M×N grid topology; and the center coordinates of the text boxes recognized by OCR are used to... x , yThe unstructured image is mapped to a grid to determine its row and column, thus transforming it into structured HTML table data. For cells spanning multiple rows or columns, they are merged and marked based on the missing grid lines. The U-Net network employs a classic encoder-decoder architecture with skip connections to achieve pixel-level accurate prediction of line masks. The network is initialized using weights pre-trained on public table datasets such as PubTabNet and further fine-tuned using a large number of scanned images of legal documents labeled with table lines to improve robustness in detecting complex, blurry, or incomplete table lines. The ASR speech transcription and role separation subunit is used to perform VAD speech activity detection on audio data, remove silent segments, and obtain processed audio data; convert the processed audio data into text; calculate the voiceprint feature distance of different speech segments, classify segments with a distance less than a preset clustering threshold as the same speaker, and output dialogue text with timestamps and role tags. In specific implementation, the ASR speech transcription and role separation subunit function in the above embodiment is as follows: For recorded interrogation transcripts or alarm phone calls, VAD detection is performed, and silent segments with energy below -40dB are removed to reduce invalid calculations; an x-vector extractor is used to extract a 512-dimensional voiceprint feature vector every 0.5 seconds; unsupervised clustering of the feature vectors is performed using a spectral clustering algorithm; the cosine distance matrix between vectors is calculated to construct a similarity graph; and a clustering stopping threshold of 0.7 is set. This clustering threshold of 0.7 is determined experimentally based on the cosine similarity distribution of a large number of voiceprint samples. When the degree is greater than 0.7, the accuracy rate of identifying the same speaker exceeds 95%; the clustering results divide the audio into "speaker A1" and "speaker B1"; the speech duration and interaction patterns of speakers A1 and B1 are statistically analyzed. In the field of documents, the questioner (such as the police officer) speaks briefly and mostly in question sentences, while the questioned person (such as the party concerned) speaks longer and mostly in declarative sentences; combined with semantic analysis (such as identifying keywords such as "name" and "what is your name"), speaker A1 is automatically marked as "questioner" and B1 is marked as "party concerned", and the output format is a text sequence of "[00:12][questioner]: what was your speed at the time"; Furthermore, when there are contradictions in the same factual descriptions (such as the license plate number of the vehicle involved) from different data sources (such as image OCR and audio ASR), the system will make a decision based on preset credibility priority rules. For example, the rule can be set as: information with higher timestamp accuracy takes priority. Specifically, information confirmed in real time in the interrogation transcript recording (ASR), whose timestamp is accurate to the second, usually has a higher priority than blurry photos (OCR) taken afterward. The system records this decision in the metadata and can add a note "verified by multiple sources" to the generated content. The structured data linearization sub-unit is used to read Excel or CSV files, identify header rows, traverse data rows, convert the content of each cell into a header name, and record the row number index of that cell as metadata. In specific implementation, the structured data linearization subunit function in the above embodiment is as follows: for Excel files such as alcohol test reports or bank statements, the system does not directly store the table file, but instead reads the header row. H =[ h 1, h 2, ..., h n ]( n (Indicate the header type, such as name, detection value, detection time); traverse the data rows. D i =[ d i1 , d i2 ,..., d in Convert each line into a natural language sentence: "In the form 'Alcohol Test Report', the first..." i The line record shows: h 1 is d i1 , h 2 is d i2 ...”; The original line number and sheet name are bound to this text segment as metadata to ensure that the subsequent large model can understand the precise location of the data source; The above embodiments solve the problems of low recognition rate, structural loss and role confusion of traditional single OCR or ASR tools when processing complex case materials by using multi-engine parallel processing and targeted preprocessing algorithms (such as low confidence retries, table topology restoration and voiceprint clustering), and provide a high-quality, structured and metadata-rich unified text stream for subsequent semantic analysis.

[0016] The knowledge base building engine in the above embodiments includes: The intelligent segmentation subunit is used to set the basic sliding window size and the overlap step size; based on the basic sliding window size and the overlap step size, the uniform plain text sequence is segmented to obtain overlapping text blocks; during the segmentation process, if the paragraph end mark output by the OCR intelligent recognition subunit or the speaker switching mark output by the ASR speech transcription and role separation subunit is detected, the current window is forcibly truncated to prevent semantic cross-paragraph adhesion and generate the overlapping text blocks; In specific implementation, the intelligent segmentation subunit function of the above embodiment is as follows: the basic sliding window size is set to L=512 Tokens, and the overlap step size is S=128 Tokens; when segmenting, the system will pre-read special markers in the text; if "[paragraph end mark]" or "[speaker switch]" is detected at the 400th Token in the window, the current block is forcibly truncated at that point, and the next block starts from that point; this avoids forcibly splicing "police question" and "suspect's answer" into the same semantic vector, causing semantic confusion; A hybrid index construction subunit is used to extract key entities from overlapping text blocks, including names of people, places, organizations, times, and amounts involved in the case; the key entities are used as inverted index keywords; each text block is vectorized and encoded to generate a high-dimensional semantic vector as a vector index; a dual-path retrieval structure combining the inverted index and the vector index is constructed to ensure that semantic similarity can be matched and specific entities can be accurately hit during retrieval; In specific implementation, the hybrid index construction sub-unit function of the above embodiment is as follows: Named entity recognition is performed using the LTP toolkit; personal names, place names, organizations, time, and amount / value are extracted from text blocks; an inverted index table is constructed: Key: "Zhang San" → Value: [Block ID 1, Block ID 5,...]; the existing BERT model is called, which contains a 24-layer Transformer encoder with a hidden layer dimension of 1024. It is trained using Chinese corpus and fine-tuned using supervised contrastive learning with a large amount of professional domain corpus such as legal documents, case records, and judgment documents, so that the vectors generated by the model can better capture the semantic features and entity relationships of legal documents, thereby improving the accuracy and domain adaptability of subsequent retrieval; the text blocks are mapped to 768-dimensional vectors; a vector index is constructed using the HNSW algorithm; each storage unit structure is { "ID": UUID, "Vector": Vector embedding, "Text": Original text, "Keywords": [Entity list], "Metadata": {Source, Time, Page number}}; The above embodiments effectively solve the problems of semantic fragmentation in long documents and low entity retrieval accuracy through intelligent block segmentation strategy and hybrid indexing technology; in particular, the boundary truncation logic prevents confusion between different semantic roles, while the hybrid index takes into account both the breadth of semantic understanding and the accuracy of entity matching, laying the foundation for subsequent accurate retrieval. The above embodiments can be configured with an intelligent template orchestration center including: The variable dependency graph construction unit is used to parse all dynamically generated variables configured by the user and construct a directed acyclic graph based on the contextual dependencies in the variable generation metadata of the dynamically generated variables. That is, if the natural language generation instruction of dynamically generated variable B explicitly references the variable of dynamically generated variable A, then a directed edge from A to B is established to obtain a directed acyclic graph (DAG). In specific implementation, the variable dependency graph construction unit function in the above embodiment is as follows: In the system's template editor, a standard format template of a "Traffic Accident Determination Certificate" is opened; this template has pre-set fixed chapter areas such as "Accident Process" and "Responsibility Determination"; a variable placeholder is inserted in the main text of the "Accident Process" chapter and named "Accident Description Variable"; another variable placeholder is inserted in the main text of the "Responsibility Determination" chapter and named "Responsibility Determination Variable," thus obtaining the dynamically generated variable area defined in the target document template; the system does not completely rely on manual configuration of dependencies, but the variables also rely on the automatic dependency discovery algorithm embedded in the graph construction unit; the automatic dependency discovery algorithm analyzes the semantics of each variable generation instruction and identifies potential logical connections between instructions through existing causal reasoning models; for example, if the instruction for variable A is "describe the accident process" and the instruction for variable B is "analyze the cause of the accident," the algorithm can automatically infer that B depends on the generation result of A, thereby suggesting or automatically establishing a dependency edge from A to B; users can confirm or modify the automatically discovered dependencies, thus achieving a balance between automation and controllability; Furthermore, the instruction for the "Accident Description Variable" is configured as "Describe the accident occurrence in chronological order based on the record and on-site investigation diagram," and it is set to have no dependencies; the instruction for the "Responsibility Determination Variable" is configured as "Determine the responsibility of the parties involved based on the facts in the Accident Description Variable and the alcohol test results, in accordance with the Road Traffic Safety Law," and it is explicitly selected to depend on the "Accident Description Variable" in the dependency settings; after parsing the above configuration, the system automatically constructs a directed acyclic graph, where nodes are variables and edges are dependencies, that is, establishing a directed edge from the "Accident Description Variable" to the "Responsibility Determination Variable"; The topology sorting and cycle detection unit is used to execute the Kahn algorithm on a directed graph composed of variables; it counts the in-degree of all nodes and adds variables with an in-degree of 0 to the execution queue; for each dynamically generated variable processed, it decrements the in-degree of its downstream nodes by 1; if there are still nodes that have not entered the queue after the Kahn algorithm ends, it is determined that there is a circular dependency and an error message is output; otherwise, it outputs a linear optimal execution path sequence. In specific implementation, the topology sorting and cycle detection unit in the above embodiment functions as follows: Execute the Kahn algorithm on the directed acyclic graph (DAG); initialize the in-degree of all nodes: the accident description variable is 0, and the responsibility determination variable is 1; add the accident description variable with an in-degree of 0 to the execution queue Q=[accident description variable]; simulate the execution of the accident description variable, remove its emanating edges, and the in-degree of the responsibility determination variable becomes 0; add the responsibility determination variable to the queue Q=[accident description variable, responsibility determination variable]; output the final execution order; the final execution order is to first generate the accident process, and then generate the responsibility determination. This mechanism ensures that the large model in the "responsibility determination" stage can "see" the standardized case description generated by the system, guaranteeing the rigor of the preceding and following logic. The instruction parameterization encapsulation unit is used to allow users to embed metadata filters in natural language generation instructions that dynamically generate variables in the optimal execution path sequence. Specifically, in the above embodiment, the instruction parameterization encapsulation unit functions as follows: the user inputs "extract transfer records after October 1, 2025" in the instruction for the template variable "recent transfer records"; the system uses regular expressions to parse the time condition ">2025-10-01"; during subsequent retrieval of the vector library, a metadata filter is automatically attached: { "date": { "greater than": "2025-10-01"}}, thereby excluding old data and improving the relevance of the retrieval. The above embodiments creatively solve the problem of logical dependencies and sequential execution of various parts of the document generation by introducing DAG and topological sorting algorithms from graph theory, so that the generated document is no longer a collection of fragments, but a logically rigorous whole; at the same time, the introduction of metadata filters further improves the timeliness and accuracy of information extraction.

[0017] The retrieval enhancement and inference generation engine in the above embodiments includes: The hybrid retrieval strategy execution subunit is used to calculate the cosine similarity between the query vector of dynamically generated variables and the vectors in the global semantic vectorized knowledge base to obtain a semantic score; calculate the matching degree between the query keywords and the inverted index to obtain a keyword score; perform a weighted summation of the semantic score and the keyword score according to the weighted summation formula to obtain the final score; and select the top-K overlapping text blocks with the highest final scores as contextual evidence. The dynamic adjustment mechanism for the weight coefficients of the weighted summation formula in the hybrid retrieval strategy execution subunit of the above embodiments includes: Obtain the natural language generation instructions corresponding to the current dynamically generated variables; analyze the intent type of the natural language generation instructions; if the instructions contain semantic verbs, determine that it is an abstract task, set the weight coefficient to favor semantic scoring, and focus on semantic vector retrieval; the semantic verbs are such as "summarize, generalize, analyze"; If the natural language generation instruction contains keyword-like verbs, it is determined to be a precise fact task, and the weight coefficient is set to favor keyword scoring, emphasizing keyword inverted index retrieval; keyword-like verbs include "extract," "list," and "statistics." If the natural language generation instruction contains entity-constrained nouns, entity filtering is forcibly enabled, and similarity calculation is performed only within the text block containing the corresponding entity; the weight coefficient is set to favor keyword scoring; the entity-constrained nouns are such as "name, ID number, amount"; In specific implementation, the hybrid retrieval strategy execution subunit function in the above embodiment is as follows: when the system executes the accident description variable, it converts the instruction "describe the accident occurrence process" into a query vector. V q ; Calculate semantic score S sem =Cosine( V q , V doc ); V doc The vector represents a document text block in the global semantic vectorized knowledge base; during calculation, the dot product of the two vectors is first calculated, and then each vector is divided by its respective L2 norm. The resulting cosine value is the semantic similarity score. S sem Its range is [-1, 1]. The closer the value is to 1, the more similar the query instruction is to the text block in semantic space; extract keywords (such as "accident" or "after") from the instruction and calculate the BM25 score. S kw The semantic weight coefficient is not a fixed value, but is derived from a dataset labeled with task types and ideal retrieval results, and then optimized through grid search and cross-validation. For abstract tasks such as 'summarizing, generalizing, and analyzing', the optimized weight range is [0.7, 0.9] (0.8 in this embodiment), which ensures that semantic similarity dominates the retrieval score, making the results more in line with the needs of deep semantic understanding. Similarly, for precise fact tasks such as 'extracting, listing, and statistically analyzing', the optimized weight range is [0.3, 0.5] (0.4 in this embodiment) to strengthen the role of precise keyword matching and ensure the accuracy of fact retrieval. The final score Sfinal is calculated as 0.8*Ssem + 0.2*Skw. For tasks classified as precise fact tasks, the semantic weight coefficient α is set to 0.4 (i.e., keyword weight is 0.6). This is because fact tasks require absolutely accurate data sources, and reducing the semantic weight can reduce noise interference caused by semantic generalization. S final The top 5 highest-ranking text blocks (such as segments in a statement from the party admitting to speeding; OCR recognition results of the length of skid marks at the scene; transcripts of emergency call recordings). The dynamic prompt word assembly subunit constructs a prompt word template that includes system instructions, known information, historical generation, current task, and constraint conditions based on the highest-scoring Top-K text blocks. In specific implementation, the dynamic prompt word assembly subunit function of the above embodiment is as follows: [System Instruction]: You are a professional traffic police document writing assistant; [Known Information]: 1. (Source: Record) The party involved, Zhang San, said: "The speed at the time was about 80 km / h;" 2. (Source: Investigation) The skid marks are 50 meters long, and the road surface is slippery; 3. (Source: Alarm) The person who reported the accident said that he saw two vehicles collide; [Current Task]: Based on the reference materials, describe the accident in chronological order; [Constraints]: Objective and neutral, without speculative words, 300 words; The variable generation content unit is used to calculate the total number of tokens for the prompt word template. If the total number of tokens exceeds the context limit of the large language model, text blocks are gradually discarded from low to high according to the final score until the length requirement is met. Text blocks containing key entities are retained first to obtain the final prompt word template. The prompt word template is input into the large language model to obtain the variable generation content. In specific implementation, the context window adaptive unit function in the above embodiment is as follows: calculate the total number of tokens for the prompt word template; if it exceeds 8192 tokens (model limit), check the Top-5 text blocks; if text block 3... S final If the minimum number of tokens does not contain the specific entity required by the instruction (such as "Zhang San"), then text block 3 is discarded, and the token count is recalculated until the limit is met, ensuring that core information is not lost. The final prompt word template is then input into a large language model (Qwen-72B in this embodiment) to obtain the variable generation content. Qwen-72B is a large language model based on the Transformer decoder architecture, employing rotational position encoding and the SwiGLU activation function. Before use, this model underwent supervised fine-tuning on a massive corpus of legal documents and reinforcement learning based on human feedback to enhance its logical reasoning in the legal field. Regarding text generation capabilities, in the supervised fine-tuning phase, over 100,000 anonymized legal document corpora (including judgments, indictments, contracts, reports, etc., with category distribution matching the target document type) were used. The training epochs lasted 3 epochs, employing the LoRA parameter efficient fine-tuning method to adapt to the professional domain. In the reinforcement learning phase based on human feedback, the model was strengthened using high-quality positive examples extracted from massive legal document assignments and manually labeled corrected negative examples, undergoing iterative optimization within the PPO algorithm framework. The training converged when the factual accuracy and logical coherence scores of the generated content on the retained validation set no longer significantly improved. The above embodiments achieve evidence-based accurate generation through a hybrid retrieval strategy and dynamic prompt word assembly technology; in particular, the dynamic adjustment of weight coefficients enables the system to flexibly respond to two different instruction requirements: "generalization" and "precision"; while the context window adaptive mechanism ensures that the model can always obtain the information with the highest value density under limited computing resources.

[0018] The verification and backfilling module in the above embodiment includes: The factual implication calculation subunit is used to take the generated variable content as an assumption and the concatenation of contextual evidence as a premise; it calls the natural language inference model to calculate the implication probability of the premise to the assumption; if the implication probability is lower than a preset threshold, it is judged as a potential illusion, and the system automatically adds a mark that needs to be manually verified after the variable. The processing procedure for implied probability below a preset threshold includes: A pre-constructed dataset of judicial document verification containing positive and negative samples was built; the RoBERTa model was fine-tuned and trained to obtain a pre-trained RoBERTa model. During the inference phase, if the implied probability obtained through the RoBERTa model is lower than the implied probability threshold of 0.6, the prompt word template, the generated variable content, and the prompt "The above generated variable content may contain factual errors; please re-check the context and correct them" are concatenated and input into the large language model for secondary generation. If the implied probability of the secondary generated variable content is still lower than 0.6, the second generation result is retained and highlighted, prompting manual intervention. The implied probability threshold of 0.6 was determined after performance evaluation on an independent judicial document fact verification set. On the verification set, the threshold of 0.6 is sufficient to recall potential hallucinations and avoid wrongful executions. The optimal balance is achieved, resulting in the highest harmonic mean (F1 score) of precision and recall. This threshold can be configured and adjusted according to the rigor requirements of different document types. Furthermore, in addition to the probability threshold, the factual implication calculation subunit is also connected to a knowledge base of legal provisions and business rules. After the generated content passes the probability verification, it will be further analyzed by the rule engine to check whether it conforms to the expression of relevant legal provisions, whether it contains prohibited terms, and whether the numerical logic (such as time sequence and total amount) is self-consistent. Any rule violation will trigger the same correction or manual verification process as the low implication probability, thereby achieving multi-dimensional interception of complex logical fallacies. In specific implementation, the function of the factual implication calculation subunit in the above embodiment is as follows: The variable content "Zhang San's license plate number is Min A88888" generated by the language model is used as the "hypothesis"; the retrieved Top-5 contextual evidence is concatenated as the "premise"; the RoBERTa model is called, and the model outputs the probabilities of three categories: implication, neutral, and contradiction; if the probability of implication (i.e., the implication score) is <0.6, or the probability of contradiction is >0.3, it is judged as high risk; the system triggers a self-correction loop, concatenating the original structured prompt words with the generated erroneous variable content, and the prompt "The above generated content may contain factual errors, please re-check the context and correct it." Next, the data is input into the large language model for secondary generation. If the implication score after secondary generation is still lower than 0.6, the result of the second generation is retained and the following is generated in the document: "Zhang San drives a vehicle (system question: Min A88888, manual verification required)". The RoBERTa model contains a 24-layer Transformer encoder with a hidden layer dimension of 1024 and 16 attention heads. The model is first pre-trained on a massive unsupervised Chinese corpus, and then supervised fine-tuned using approximately 10,000 specially constructed judicial document fact verification data (labeled with "premise-assumption" pairs and their implication relationship labels) to adapt to factual logic reasoning tasks in the judicial field. The entity consistency review subunit is used to extract the entity set in the variable generation content and the entity set in the retrieval credentials; calculate the entity conflict rate; if the generated text contains proper nouns that do not exist in the retrieval credentials and the entity conflict rate is greater than 0, it is judged as an error of fabrication, and the system returns to the dynamic prompt word assembly subunit, and adds a negative constraint that strictly prohibits the fabrication of entities that do not appear, until the variable generation content that passes the test is obtained. In specific implementation, the entity consistency review subunit function in the above embodiment is as follows: extract the set of all entities in the context credentials. E src ={Zhang San, Li Si, Min A12345}; Extract the entity set from the generated variable content. E gen ={Zhang San, Wang Wu, Min A12345}; System detection found E gen The character "Wang Wu" appeared in the text, and E gen The error is not found in the text, indicating it's a classic case of "illusion" (creating something out of nothing); calculate the conflict rate R = | E gen - E src | / | E gen |; If R>0, force regeneration and add a negative constraint to the dynamic prompt word assembly subunit: "Prohibit mentioning names not found in the references"; The format lossless backfill sub-unit is used to parse Word templates based on the Office Open XML standard; locate the Run node where the dynamically generated variable placeholders are located; write the variable generation content that passes the detection according to the original style attributes of the Run node; if the variable generation content contains table data, it automatically clones the table row structure in the document for dynamic filling. In specific implementation, the lossless backfill subunit function of the above embodiment is as follows: parsing the XML structure of the Word template; locating the location of the accident description variables. <w:r>Node; read the node's style attributes (e.g., font: SimSun GB2312, font size: 3, bold: none); write the validated variable content into the node, keeping the style unchanged; if the variable content is a list or table, the system identifies the table rows in the template. <w:tr>And based on the number of generated data rows, clone the data in a loop. <w:tr>Nodes are populated with specific data to achieve dynamic table generation; The above embodiments construct a rigorous "anti-illusion" closed loop by introducing a natural language reasoning model and an entity consistency review mechanism; the generated content is subject to dual quality checks from two dimensions: semantic logic and entity set, ensuring the authenticity and accuracy of the document content; at the same time, the lossless backfilling technology based on XML ensures the professional format of the final delivered document, eliminating the need for manual secondary typesetting.

[0019] The configurable document intelligent template mapping and automated content generation system also includes a privacy desensitization and security audit layer: Before segmenting and vectorizing the unified plain text sequence, sensitive information in the unified plain text sequence is identified using regular matching and NER model. The FPE algorithm is used to replace sensitive information with ciphertext while preserving the original format characteristics. Before backfilling the generated variable content, determine whether to decrypt and restore the ciphertext based on the current user's permission level; if permissions are insufficient, output the masked text. Record the token consumption, retrieved context credential ID, and generated text fingerprint during the generation of all dynamically generated variables, and store them in an immutable audit log blockchain; In specific implementation, the privacy desensitization and security audit layer function of the above embodiment is as follows: Before the data is stored in the vector database, sensitive information (such as mobile phone numbers and ID card numbers) in the uniform plain text sequence is identified using regular expression matching and NER model; the sensitive information is replaced with ciphertext using the FPE (Format Preservation Encryption) algorithm, for example, "12025202520" is encrypted as "12026202620", preserving its format feature as a mobile phone number so that the model can still recognize it as a phone number; before the generated variable content is backfilled, the permission level of the current user is checked; if the user's permission is "ordinary clerk", then no decryption is performed. The output is masked text (e.g., "120****2620"); if it is "authorized case handler", it is decrypted and restored; the token consumption, retrieved context credential ID, and generated text fingerprint of all dynamically generated variables are recorded and stored in an immutable audit log blockchain to ensure that every document generation is traceable; the blockchain is an immutable audit log system built on consortium blockchain (or private blockchain) technology; the specific choice between consortium blockchain and private blockchain depends on the actual deployment environment of the government / judicial system; consortium blockchain is more suitable for cross-departmental collaboration, while private blockchain is more suitable for single-institution internal use.

[0020] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0021] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.< / w:tr> < / w:tr> < / w:r>

Claims

1. A configurable intelligent document template mapping and automated content generation system, characterized in that, include: Data parsing layer: Used to receive unstructured images, unstructured audio, and semi-structured spreadsheets, parse them by identifying data types through file header features, and generate a unified plain text sequence with metadata tags through conflict resolution operations based on timestamp priority rules; Knowledge base building engine: used to divide a unified plain text sequence into blocks to obtain text blocks; to perform vectorization encoding and entity extraction operations on the text blocks to obtain high-dimensional semantic vectors and inverted indexes, and store them in a vector database to form a global semantic vectorized knowledge base; Configurable intelligent template arrangement center: used to configure variables and generate metadata for each dynamically generated variable in the dynamic variable area of ​​the target document; A directed acyclic graph is constructed based on the metadata generated from the variables, and the execution priority sequence of dynamically generated variables is calculated to generate a template configuration description file; The variable generation metadata includes the natural language generation instructions, data source constraints, and context dependencies for each dynamically generated variable; The retrieval enhancement and reasoning generation engine is used to perform hybrid retrieval operations in the global semantic vectorized knowledge base according to the template configuration description file to obtain contextual credentials. Based on contextual credentials, structured prompt words are constructed and input into a large language model to obtain variable-generated content; Validation and backfilling module: used to perform entity consistency verification and illusion detection on the variable-generated content; the content that passes the verification is backfilled into the corresponding position in the document template and the format rendering is maintained, and the target document is output; If the verification fails, the content correction mechanism will be triggered.

2. The configurable intelligent document template mapping and automated content generation system according to claim 1, characterized in that, The data parsing layer includes: The OCR intelligent recognition subunit is used to extract features from image data; perform sequential text recognition based on the features of image data; for sequential text in low-confidence areas, call the image enhancement algorithm for binarization and noise reduction retry; at the same time, it recognizes red stamps, handwritten signatures and table lines in the image, and outputs structured JSON data containing text content, coordinate position and font attributes. The ASR speech transcription and role separation subunit is used to perform VAD speech activity detection on audio data, remove silent segments, and obtain processed audio data; convert the processed audio data into text; calculate the voiceprint feature distance of different speech segments, classify segments with a distance less than a preset clustering threshold as the same speaker, and output dialogue text with timestamps and role tags. The structured data linearization sub-unit is used to read Excel or CSV files, identify header rows, traverse data rows, convert the content of each cell into a header name, and record the row number index of that cell as metadata.

3. The configurable intelligent document template mapping and automated content generation system according to claim 2, characterized in that, The specific logic of the OCR intelligent recognition subunit when processing table lines is as follows: Predict horizontal and vertical line masks in an image using the U-Net semantic segmentation network; Perform Hough transform on the mask to detect lines and obtain the equations of all lines; based on the equations of all lines, calculate the intersection points of the lines and construct the grid topology of the table. The center coordinates of the text box recognized by OCR are mapped to the grid topology to determine the row and column indices of the text.

4. The configurable intelligent document template mapping and automated content generation system according to claim 2, characterized in that, The knowledge base building engine includes: The intelligent segmentation subunit is used to set the basic sliding window size and the overlap step size; based on the basic sliding window size and the overlap step size, the uniform plain text sequence is segmented to obtain text blocks; during the segmentation process, if the paragraph end mark output by the OCR intelligent recognition subunit or the speaker switching mark output by the ASR speech transcription and role separation subunit is detected, the current window is forcibly truncated to generate the text block. A hybrid index construction subunit is used to extract key entities from text blocks; the key entities are used as inverted index keywords; each text block is vectorized to generate a high-dimensional semantic vector as a vector index; a dual-path retrieval structure combining the inverted index and the vector index is constructed.

5. The configurable intelligent document template mapping and automated content generation system according to claim 4, characterized in that, The configurable intelligent template orchestration center includes: The variable dependency graph building unit is used to parse all dynamically generated variables configured by the user and construct a directed acyclic graph based on the context dependencies in the variable generation metadata of the dynamically generated variables. The topology sorting and cycle detection unit is used to execute the Kahn algorithm on a directed graph composed of variables; it counts the in-degree of all nodes and adds variables with an in-degree of 0 to the execution queue; for each dynamically generated variable processed, it decrements the in-degree of its downstream nodes by 1; if there are still nodes that have not entered the queue after the Kahn algorithm ends, it is determined that there is a circular dependency and an error message is output; otherwise, it outputs a linear optimal execution path sequence. The instruction parameterization encapsulation unit is used to embed metadata filters in natural language generation instructions that allow users to dynamically generate variables in the optimal execution path sequence.

6. The configurable intelligent document template mapping and automated content generation system according to claim 5, characterized in that, The retrieval enhancement and reasoning generation engine includes: The hybrid retrieval strategy execution subunit is used to calculate the cosine similarity between the query vector of dynamically generated variables and the vectors in the global semantic vectorized knowledge base to obtain a semantic score; calculate the matching degree between the query keywords and the inverted index to obtain a keyword score; perform a weighted summation of the semantic score and the keyword score according to the weighted summation formula to obtain the final score; and select the top-K text blocks with the highest final scores as contextual credentials. The dynamic prompt word assembly subunit constructs a prompt word template that includes system instructions, known information, historical generation, current task, and constraint conditions based on the highest-scoring Top-K text blocks. The variable generation content unit is used to calculate the total number of tokens for the prompt word template. If the total number of tokens exceeds the context limit of the large language model, text blocks are gradually discarded from low to high according to the final score until the length requirement is met, and the final prompt word template is obtained. The final prompt word template is input into the large language model to obtain the variable generation content.

7. A configurable intelligent document template mapping and automated content generation system as described in claim 6, characterized in that, The dynamic adjustment mechanism for the weight coefficients of the weighted summation formula in the hybrid retrieval strategy execution subunit includes: Obtain the natural language generation instructions corresponding to the current dynamically generated variables; analyze the intent type of the natural language generation instructions; if the instructions contain semantic verbs, determine that it is an abstract task, set the weight coefficient to favor semantic scoring, and focus on semantic vector retrieval; If the natural language generation instruction contains keyword-type verbs, it is judged as an exact fact task, and the weight coefficient is set to favor keyword scoring, emphasizing keyword inverted index retrieval. If the natural language generation instruction contains entity-constrained nouns, entity filtering is forcibly enabled, and similarity calculation is performed only within the text block containing the corresponding entity; the weight coefficient is set to favor keyword scoring.

8. A configurable intelligent document template mapping and automated content generation system as described in claim 6, characterized in that, The verification and backfilling module includes: The fact-implied calculation subunit is used to take the generated variable content as an assumption and the concatenation of contextual evidence as a premise; calculate the implied probability of the premise on the assumption; if the implied probability is lower than a preset threshold, it is judged as a potential illusion, and the system automatically adds a mark that needs to be manually verified after the variable. The entity consistency review subunit is used to extract the entity set in the variable generation content and the entity set in the retrieval credentials; calculate the entity conflict rate; if the generated text contains proper nouns that do not exist in the retrieval credentials and the entity conflict rate is greater than 0, it is judged as an error of fabrication, and the system returns to the dynamic prompt word assembly subunit, and adds a negative constraint that strictly prohibits the fabrication of entities that do not appear, until the variable generation content that passes the test is obtained. The format lossless backfill sub-unit is used to parse Word templates based on the Office Open XML standard; locate the Run node where the dynamically generated variable placeholders are located; write the variable generation content that passes the detection according to the original style attributes of the Run node; if the variable generation content contains table data, it automatically clones the table row structure in the document for dynamic filling.

9. A configurable intelligent document template mapping and automated content generation system as described in claim 8, characterized in that, The processing procedure for implied probability below a preset threshold includes: If the obtained implied probability is lower than the implied probability threshold, the prompt word template, the variable generation content, and the prompt "The above variable generation content may contain factual errors. Please check the context and correct it again" are concatenated and then input into the large language model for secondary generation. If the implied probability of the second-generation variable generation content is still lower than the implied probability threshold, the second generation result is retained and highlighted to prompt manual intervention.

10. A configurable intelligent document template mapping and automated content generation system as described in claim 1, characterized in that, It also includes a privacy anonymization and security audit layer: Before segmenting and vectorizing the unified plain text sequence, sensitive information in the unified plain text sequence is identified using regular matching and NER model. The FPE algorithm is used to replace sensitive information with ciphertext while preserving the original format characteristics. Before backfilling the generated variable content, determine whether to decrypt and restore the ciphertext based on the current user's permission level; if permissions are insufficient, output the masked text. Record the token consumption, retrieved context credential ID, and generated text fingerprint during the generation of all dynamically generated variables, and store them in an immutable audit log blockchain.

Images