Document review system based on AI intelligent agent

Abstract

The invention discloses a document review system based on an AI agent, and relates to the technical field of document collaborative review, the system comprises a document processing unit used for obtaining documents of various formats and preprocessing the documents by using a format conversion model to obtain preprocessed structured text data, and a data extraction unit used for extracting the structured text data from the document processing unit, the data analysis module is used for analyzing the structured data based on a data extraction model and extracting enhanced metadata and optimized text blocks; the data processing unit is used for processing the enhanced metadata and the optimized text block by utilizing a data conversion model to obtain first-order vector representation; associating the first-order vector representations based on a text perception model to generate associated second-order vector representations; and the text generation unit is used for constructing a vector knowledge base, collecting and storing the second-order vector representation, and generating a reviewed text by using a document generation model. According to the method, repeated work such as data entry, format adjustment and content auditing is automatically completed through the AI.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document collaborative review, in particular to a document review system based on AI agent. BACKGROUND

[0002] In modern enterprise operation, document processing is an indispensable part, especially in the case of needing to process a large number of files and data. These work often includes data entry, format adjustment, content review and other simple but repetitive tasks, which consumes a lot of human resources and time, at the same time, in order to improve work efficiency and quality, the enterprise also needs to analyze the data in the document, and take it as an important part of the internal knowledge base, in order to facilitate the rapid retrieval and use of information, however, the traditional manual operation mode is not only low in efficiency, but also prone to error, it is difficult to meet the high requirements of modern enterprises.

[0003] The accuracy and reliability of the system are highly dependent on the quality and diversity of the training data. If the training samples are not rich enough or representative, it may lead to misjudgment of the system in actual application. Although AI technology can understand the content of the document to a certain extent, its understanding and processing ability is still limited when facing complex or professional field knowledge, which may need manual intervention to ensure the accuracy of the results. It needs high cost and technical support to build and maintain such a complex AI system, which may be a big burden for small and medium-sized enterprises.

[0004] At present, there is no effective solution to the problems in the related art. SUMMARY

[0005] In view of the problems in the related art, the present application proposes a document review system based on AI agent to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows: The present application provides a document review system based on AI agent, comprising: a document processing unit, configured to obtain documents of various formats, and pre-process the documents by using a format conversion model to obtain structured text data after pre-processing; a data extraction unit, configured to parse the structured data based on a data extraction model, and extract enhanced metadata and optimized text blocks; a data processing unit, configured to process the enhanced metadata and the optimized text blocks by using a data conversion model to obtain first-order vector representation, and associate the first-order vector representation based on a text perception model to generate associated second-order vector representation; The text generation unit is configured to build a vector knowledge base, collect and store the second-order vector representation, and generate the reviewed text by using a document generation model; The iterative optimization unit is configured to collect the modification and evaluation of the generated reviewed text by using a loop feedback mechanism, and further iteratively optimize the document generation model.

[0007] Further, the document processing unit comprises: The document classification module is configured to obtain various formats of documents, divide the obtained various documents into text documents and image documents by using a format analyzer; The document recognition module is configured to process the text documents and the image documents based on a text recognition algorithm and an image recognition algorithm respectively, and obtain structured text data.

[0008] Further, the data extraction unit comprises: The data enhancement module is configured to identify the nouns in the structured data by using a pre-trained language model, and perform definition classification to generate enhanced metadata; The data optimization module is configured to perform semantic coherence and semantic coherence integrity on the text block content in the structured data by using an improved pre-trained language model, and generate optimized text blocks.

[0009] Further, the data processing unit comprises: The data conversion module is configured to perform vectorization processing on the enhanced metadata and the optimized text blocks by using a data conversion model, and obtain metadata vectors and text block vectors; The data construction module is configured to construct a document knowledge graph based on the metadata vectors and the text block vectors; The data association module is configured to receive the document knowledge graph by using a graph neural network, and associate the document knowledge graph to generate an associated second-order vector representation.

[0010] Further, the construction of the document knowledge graph based on the metadata vectors and the text block vectors comprises: Each metadata vector and each text block vector is converted into a metadata node and a text block node; The relationship edges between the metadata nodes and the text block nodes are established based on a relationship extraction model, and the document knowledge graph is constructed.

[0011] Further, the receiving of the document knowledge graph by using the graph neural network and the association of the document knowledge graph to generate the associated second-order vector representation comprises: The message passing mechanism of the graph neural network is used to receive and associate the adjacent nodes in the document knowledge graph to generate the associated second-order vector representation.

[0012] Further, the relationship extraction model is used to establish a relationship edge between the metadata node and the text block node, and to construct a document knowledge graph, which comprises: The entity recognition model based on the attention mechanism captures the context information in the text, identifies the relationship between the nodes, extracts the relationship between the metadata node and the text block node, and constructs a document knowledge graph. The model is trained using labeled data to optimize the entity recognition model parameters, thereby improving the accuracy and robustness of relationship extraction.

[0013] Further, the message passing mechanism of the graph neural network receives and associates adjacent nodes in the document knowledge graph to generate a second-order vector representation after association, which comprises: The adjacent nodes in the document knowledge graph are aggregated based on the attention mechanism to generate a second-order vector representation after association The adjacent nodes in the document knowledge graph are aggregated using the graph attention network, and the aggregation is performed through the attention mechanism to obtain a second-order vector representation after association.

[0014] Further, the text generation unit comprises: A text storage module is used to store each second-order vector representation as a key-value pair in the constructed vector database, wherein the key-value pair comprises the second-order vector representation and the payload. A text matching module is used to calculate the similarity based on the second-order vector representation using the vector knowledge base, and to retrieve the second-order vector representation with similar similarity through the similarity management mechanism. A text generation module is used to input the second-order vector representation with similar similarity as the key context into the document generation model based on the vector generation mechanism to generate the reviewed text.

[0015] Further, the iterative optimization unit comprises: An iterative optimization module is used to construct a loop feedback channel, store the revisions and evaluations of the reviewed text, and feed back to the document generation model for continuous optimization.

[0016] The beneficial effects of the present application are: 1. The present application automatically completes data entry, format adjustment, content review and other repetitive work through AI, reducing the time of manual operation. The text block vector associated by GNN can quickly identify and extract key information and automatically fill into the specified template, greatly improving the efficiency and accuracy of data entry.

[0017] 2. The present application constructs a document knowledge graph, and the system can integrate data from different sources to ensure the integrity and consistency of the information. The multi-layer aggregation mechanism of GNN can capture the complex relationship between nodes to improve the accuracy of data processing.

[0018] 3、The application can gradually adapt to different fields of professional terms and knowledge by continuously accumulating and updating training data, improving the processing capability in complex scenarios.

[0019] 4、The application can continuously improve performance and effect through continuous learning and optimization, reduce long-term maintenance cost. With the development of technology, the system can be continuously upgraded to maintain its competitiveness in the industry. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 is a principle block diagram of a document review system based on AI agent according to an embodiment of the present application; Figure 2 is a content acquisition interface diagram of a document processing unit of a document review system based on AI agent according to an embodiment of the present application; Figure 3 is a picture operation interface diagram of a document processing unit of a document review system based on AI agent according to an embodiment of the present application; Figure 4 is an excel file processing interface diagram of a document processing unit of a document review system based on AI agent according to an embodiment of the present application; Figure 5 is a PDF operation interface diagram of a document processing unit of a document review system based on AI agent according to an embodiment of the present application; Figure 6 is a file operation interface diagram of a document processing unit of a document review system based on AI agent according to an embodiment of the present application.

[0022] In the drawings: 1, document processing unit; 101, document classification module; 102, document recognition module; 2, data extraction unit; 201, data enhancement module; 202, data optimization module; 3, data processing unit; 301, data conversion module; 302, data construction module; 303, data correlation module; 4, text generation unit; 401, text storage module; 402, text matching module; 403, text generation module; 5, iterative optimization unit; 501, iterative optimization module. DETAILED DESCRIPTION

[0023] To further illustrate the embodiments, the present application provides accompanying drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operation principle of the embodiments in conjunction with the related description of the specification. Those of ordinary skill in the art should understand other possible implementations and advantages of the present application by referring to these contents.

[0024] According to an embodiment of the present application, an AI agent-based document review system is provided.

[0025] The present application will be further described in conjunction with the accompanying drawings and specific embodiments. As shown in the drawings, according to an embodiment of the present application, an AI agent-based document review system comprises: Figure 1 A document processing unit 1 is configured to obtain documents of various formats and pre-process the documents using a format conversion model to obtain pre-processed structured text data. A data extraction unit 2 is configured to parse the structured data based on a data extraction model to extract enhanced metadata and optimized text blocks. A data processing unit 3 is configured to process the enhanced metadata and optimized text blocks using a data transformation model, i.e., an embedding model, to obtain first-order vector representations, and to associate the first-order vector representations based on a text perception model, i.e., a context graph perception model, to generate associated second-order vector representations. A text generation unit 4 is configured to build a vector knowledge base, collect and store the second-order vector representations, and generate reviewed text using a document generation model, i.e., an LLM. An iterative optimization unit 5 is configured to collect user corrections and evaluations of the generated reviewed text using a cyclic feedback mechanism, and further iteratively optimize the document generation model.

[0026] In this optional embodiment, the document processing unit comprises: A document classification module 101 is configured to obtain documents of various formats, and divide the obtained documents into text documents and image documents using a format analyzer. A document recognition module 102 is configured to process the text documents and image documents based on a text recognition algorithm and an image recognition algorithm, i.e., OCR, to obtain structured text data.

[0027] ​Specifically, documents of various formats, such as PDF, Word, scanned images, etc., are obtained, and a format analyzer such as Apache Tika, an open-source document processing tool, and a Python code library called filetype, which detects the format of the input document, is used to divide it into text-based documents such as TXT and DOCX, and image-based documents such as PDF scans and JPG. Text-based documents are directly extracted by a text parser such as pdfminer and docx2txt. Image-based documents are processed by OCR technology such as Tesseract and PaddleOCR to generate preliminary structured text.

[0028] Post-processing of unstructured text output by OCR includes: 1. Text cleaning: removing noise characters such as garbled characters and formatting symbols, standardizing encoding, and unifying to UTF-8. 2. Paragraph segmentation: based on rules such as blank lines and indentation, or machine learning models such as LayoutLM, a pre-trained model, to divide logical paragraphs. 3. Language detection: using the langdetect library, a Python library for detecting text language, to filter non-target language content. Standardized structured text data is stored in JSON format.

[0029] As shown in Figure 3 , for the operation of the picture, it can be detected whether the current page has a corresponding picture, which can be used to judge whether the page is loaded, as shown in Table 1, picture detection: by intercepting the pictures in the page, it is determined whether the current screen page has intercepted pictures during running, if the picture detection is consistent, the return value is True, such as success, if no intercepted pictures are detected, the return value is False, such as failure.

[0030] Table 1 Picture detection As shown in Figure 4 , for the processing of excel files, it can read and write files, create, delete, rename sheet tables, etc., as shown in Table 2, cell reading and writing operations, etc.

[0031] Cell reading: used to read the value of a specified cell in the table and return the read value.

[0032] Table 2 Cell reading As shown in Figure 5 , for the processing of FDF files, it can read and write files, read and write page numbers, convert to pictures, extract text, etc.

[0033] As shown in Table 3, get PDF page number: get the page number of a specified PDF file.

[0034] Table 3 Obtaining PDF Page Count As shown in Table 4, the PDF information is retrieved: the version, author, creation program, creation tool, creation date, and other information of the specified PDF file are retrieved and returned as a dictionary.

[0035] Table 4 Cell Reading As shown in Table 5, to obtain PDF images: retrieve the images from the target PDF file and save the images to the specified file directory.

[0036] Table 5 Obtaining PDF Information As shown in Table 6, the PDF text retrieval function retrieves the text from a specified page of the target PDF file and returns the text string of the specified page.

[0037] Table 6 Obtaining PDF Images As shown in Table 7, Merge PDF: Merge two PDF files into one PDF file, save it to a specified location, and return the full path of the file.

[0038] Table 7 Merged PDFs As shown in Table 8, splitting a PDF: merges two PDF files into one PDF file and saves it to a specified location. The full path of the split PDF file is a list of elements.

[0039] Table 8 Splitting PDFs As shown in Table 9, PDF encryption: encrypts the unencrypted target PDF file and copies it to generate a new PDF file.

[0040] Table 9 PDF Encryption like Figure 6 As shown in Table 10, to create a file: create a new text file in the specified path; its format can be txt, doc, docx, rtf, etc.

[0041] Table 10 Creating Files As shown in Table 11, create a folder: create a folder based on the specified directory entered.

[0042] Table 11 Create Folder As shown in Table 12, read file: read the content of the specified file and return it in string type.

[0043] Table 12 Read File As shown in Table 13, read file: write content to the specified file.

[0044] Table 13 Read File As shown in Table 14, copy file: copy the file to the specified directory.

[0045] Table 14 Copy File As shown in Table 15, rename file: rename the specified file.

[0046] Table 15 Rename File As shown in Table 16, delete file: copy the file to the specified directory.

[0047] Table 16 Delete File As shown in Table 17, move file: move the file to the specified path.

[0048] Table 17 Move File As shown in Table 18, directory exists: determine whether the path exists.

[0049] Table 18 Directory Exists As shown in Table 19, get file size: return the size of the specified file.

[0050] Table 19 Get File Size In this optional embodiment, the data extraction unit comprises: The data enhancement module 201 is configured to identify the nouns in the structured data by using a pre-trained language model, i.e., BERT, and perform definition classification to generate enhanced metadata. The data optimization module 202 is used to utilize an improved pre-trained language model, namely Sentence-BERT, to perform semantic coherence and semantic completeness on the text block content in structured data, and generate optimized text blocks.

[0051] Specifically, pre-trained BERT models, such as the bert-based-uncased NER task head, are used to identify entities in the text, such as names, organizations, and dates. The identified entities are then classified to generate enhanced metadata. Relation extraction models, such as BERT-based Relation Extraction, are used to identify relationships between entities, generating richer metadata.

[0052] The Sentence-BERT model is used to calculate semantic similarity between text blocks, ensuring semantic coherence between adjacent blocks. Content is supplemented to ensure the completeness of each paragraph. The optimized text blocks undergo final proofreading to ensure grammatical correctness and logical clarity. The enhanced metadata and optimized text blocks are stored in JSON format.

[0053] like Figure 2 As shown, this module is used to retrieve text or table data in browsers or programs. It can retrieve element positions, all items, selected items, tables, etc. As shown in Table 20, element position retrieval: used to obtain the relative position coordinates of pickable controls based on the desktop.

[0054] Table 20 Element Position Acquisition As shown in Table 21, all items are retrieved: all values ​​of the dropdown list are retrieved by picking up the Java dropdown list control.

[0055] Table 21 All Items Acquisition As shown in Table 22, the selection item is obtained by picking up the Java dropdown control and getting the value of the currently selected item.

[0056] Table 22 Accessing Options As shown in Table 23, the table retrieves the value of the currently selected item in the dropdown list by picking up the Java dropdown list control.

[0057] Table 23 Table Retrieval As shown in Table 24, HTML control acquisition: If it is a web page control, obtain the corresponding HTML code of the control.

[0058] Table 24 html control acquisition In this optional embodiment, the data processing unit comprises: a data transformation module 301 configured to vectorize the enhanced metadata and the optimized text block using a data transformation model to obtain a metadata vector and a text block vector.

[0059] a data construction module 302 configured to construct a document knowledge graph based on the metadata vector and the text block vector; a data association module 303 configured to receive the document knowledge graph using a graph neural network (GNN) and associate the document knowledge graph to generate an associated second-order vector representation.

[0060] Specifically, the data transformation model, such as Word2Vec or BERT, is used to convert the metadata and text block into vector representation. The entities and relationships in the enhanced metadata are converted into vectors to generate the metadata vector. The optimized text block is converted into a vector to generate the text block vector.

[0061] Each metadata vector and each text block vector is converted into a node in the graph. A relationship extraction model, such as BERT-based Relation Extraction, is used to establish relationship edges between nodes. All nodes and relationship edges are combined into a document knowledge graph to construct a graph structure containing multiple nodes and edges, representing the relationship between metadata and text blocks in the document.

[0062] The graph neural network (GNN) is used to associate the document knowledge graph. The GNN receives and processes the nodes and edges in the graph through the MessagePassing Mechanism to generate the associated second-order vector representation. The GNN propagates the feature information of the nodes to the adjacent nodes through multiple layers to update the vector representation of the nodes. After multiple layers of propagation, the second-order vector representation of each node is generated to represent the node.

[0063] In this optional embodiment, the document knowledge graph is constructed based on the metadata vector and the text block vector, comprising: each metadata vector and each text block vector is converted into a metadata node and a text block node; based on a relationship extraction model, relationship edges between metadata nodes and text block nodes are established to construct a document knowledge graph.

[0064] Specifically, each metadata vector and each text block vector is converted into a metadata node and a text block node, respectively. It is assumed that the metadata vectors and the text block vectors have been pre-computed and stored in two matrices. A relation extraction model, such as a deep learning-based relation extraction model, is used to establish relation edges between the metadata nodes and the text block nodes. It is assumed that the relation extraction model has been trained and can output the indices of the relation edges.

[0065] In this optional embodiment, the receiving the document knowledge graph and associating the nodes in the document knowledge graph using the graph neural network comprises: The receiving and associating the adjacent nodes in the document knowledge graph using the message passing mechanism of the graph neural network comprises:

[0066] Specifically, the document knowledge graph is received and the nodes are associated. The defined graph neural network layer is used for forward propagation to generate the associated second-order vector representation. The graph neural network can be effectively used to associate the metadata and the text blocks in the document to generate a more rich vector representation.

[0067] In this optional embodiment, the establishing the relation edges between the metadata nodes and the text block nodes based on the relation extraction model to construct the document knowledge graph comprises: The entity recognition model based on the attention mechanism, i.e., BiLSTM-CRF, captures the context information in the text, identifies the relations between the nodes, extracts the relations between the metadata nodes and the text block nodes, and constructs the document knowledge graph. The model is trained using the labeled data to optimize the parameters of the entity recognition model to improve the accuracy and robustness of the relation extraction.

[0068] Specifically, the training data, including the relations between the metadata nodes and the text block nodes, are collected and labeled. The labeling can include different types of relations. A suitable relation extraction model, such as BiLSTM-CRF based on the attention mechanism, is selected. The model can capture the context information in the text and identify the relations between the nodes. The model is trained using the labeled data to optimize the parameters to improve the accuracy and robustness of the relation extraction. The trained model is applied to new documents to extract the relations between the metadata nodes and the text block nodes. According to the extracted relations, the document knowledge graph is constructed. Each node represents a metadata or a text block, and the edge represents the relation between the nodes.

[0069] In this optional embodiment, the receiving and associating the adjacent nodes in the document knowledge graph using the message passing mechanism of the graph neural network comprises: The adjacent nodes in the document knowledge graph are aggregated based on the attention mechanism to generate the associated second-order vector representation By using a graph attention network, or GAT, adjacent nodes in the document knowledge graph are aggregated and then aggregated through an attention mechanism to obtain a second-order vector representation after generating associations.

[0070] Specifically, pre-trained models such as BERT and Doc2Vec are used to extract semantic vectors from documents / entities. Linear projection is then applied to the node features to generate learnable intermediate representations. ; ; In the formula, Indicates the hidden layer dimension. d Indicates the input feature dimension. Indicates the neural network's first... Layer, or first The weight matrix of each module, Representation layer index, =1 represents the first layer. Indicates the neural network's first... The intermediate feature vector after linear transformation of the layer Indicates the neural network's first... The output feature vector of layer -1.

[0071] For nodes v and his neighbors u ∈ N ( v ), calculate attention score: ; In the formula, Represents the attention parameter vector. softmax (⋅) indicates Softmax A function used to normalize the input. LeakyReLU (⋅) denotes the activation function of a linear unit with leakage correction. This indicates a dimension concatenation operation. a T This represents a learnable weight vector. Indicates the first Layers and v Related features Indicates the first Layers and u Features of association.

[0072] In this optional embodiment, the text generation unit includes: The text storage module 401 is used to store each second-order vector representation as a key-value pair in the constructed vector database, wherein the key-value pair contains the second-order vector representation and the payload.

[0073] The text matching module 402 is configured to perform similarity calculation based on the second-order vector representation by using the vector knowledge base, and retrieve the second-order vector representation matched in similarity by a similarity management mechanism, i.e., TOP-K. The text generation module 403 is configured to input the second-order vector representation matched in similarity as key context into a document generation model, i.e., LLM, to generate the reviewed text based on a vector generation mechanism, i.e., RAG.

[0074] Specifically, the search method of the vector database, such as Milvus, is used to input the query vector q ∈ R 512 and return the K most similar vectors. When TOP-K=10, 10 most similar second-order vector representations and their payloads are returned.

[0075] The cosine similarity is used to measure the similarity between vectors. The calculation formula is as follows: ; In the formula, similarity denotes the similarity score between the query vector and the vector in the database, q denotes the query vector, c denotes the vector in the database, denotes the L2 norm of the vector.

[0076] The retrieval results are sorted in descending order according to the similarity score, and the top K are selected as the final output.

[0077] The vector knowledge base is used to perform similarity calculation based on the second-order vector representation, and the second-order vector representation matched in similarity is retrieved by the TOP-K mechanism. These matched vector representations will be input as key context into the document generation model. By adjusting the TOP-K parameter, the number of retrieval results can be controlled, thereby affecting the diversity and accuracy of the generated document. For example, setting TOP-K=5 can ensure that the generated document is based on the 5 most relevant context fragments.

[0078] The RAG, i.e., Retrieval-Augmented Generation model, combines the retrieval and generation steps. 1, Retrieval stage: use the TOP-K retrieval module to obtain the K most similar second-order vector representations and their payloads. 2, Generation stage: concatenate the K retrieved vector representations into a long vector and input it as additional context into the generation model, such as T5 or BART. The generation model generates the final reviewed text based on these context information.

[0079] LLM models can combine vector knowledge bases to quickly search and answer internal systems, processes, historical documents, etc. Employees do not need to spend a lot of time searching for information in folders or systems, but can obtain accurate answers through natural language questions. Support for document summarization, report generation, email writing, etc. Help employees work more efficiently with text. In a multilingual environment, the model can provide high-quality translation and cross-language communication support. The model can analyze and interpret data, and convert complex information into understandable conclusions and recommendations. Combined with business data and knowledge base, it provides trend analysis, comparative analysis and other intelligent support to help managers make decisions. Through the dialog interaction of vector knowledge base and LLM model, employees can quickly understand the company's regulations, standards and operation processes. Reduce training and communication costs, and help new employees get started faster and improve overall collaboration efficiency. LLM can provide customized knowledge and work suggestions according to the needs of different positions and employees.

[0080] In this optional embodiment, the iterative optimization unit comprises: The iterative optimization module 501 is used to build a loop feedback channel, store the revisions and evaluations of the reviewed text, and feed back to the document generation model for continuous optimization.

[0081] Specifically, user modifications to generated text, such as grammar corrections, factual corrections, style adjustments, and ratings or classification labels, can effectively provide feedback to the model. These feedbacks not only help identify the bias of text generation, but also refine the optimization direction. Each generated text will record the relevant input parameters, and through the construction of rule-based and machine learning, i.e. ML, classifiers, features and problem labels are extracted from user feedback. These features will be used as training data to continuously update model parameters and improve the model's ability to respond to different problems. According to the user's immediate feedback, the model strategy is constantly adjusted. This strategy optimization enables the model to dynamically choose the most appropriate generation method when facing different problems. Each update of the model will record version information to facilitate backtracking and comparison in future analysis. This ensures the traceability of each improvement and makes the optimization process transparent and controllable. Regularly collect user feedback and process each feedback through the problem tracking system in a timely manner. The feedback system will help evaluate the performance of the model, ensure that user needs are fully addressed, and guide the continuous optimization of the model.

[0082] Direct modification of the generated text by the user, such as grammar correction, fact correction, style adjustment, user-provided ratings or classification labels, input parameters during generation, such as Prompt, knowledge base version, model parameters, use of text difference algorithm, such as Google Diff Match Patch to compare the original text with the corrected text. Construct a rule-based and ML-based classifier, use the extracted features and problem labels as training data, update the model parameters. Use reinforcement learning methods to adjust the model strategy according to user feedback, record the version information of each model update for easy backtracking and comparison. Regularly collect user feedback, evaluate model performance, establish a problem tracking system to ensure that each feedback can be handled in a timely manner, and iteratively optimize the module to effectively collect user feedback, analyze problems, optimize the model, and continuously improve the quality of document generation. is to optimize LLM In summary, with the help of the above technical solutions of the present application, the present application automatically completes repetitive work such as data entry, format adjustment, content review, etc. by AI, reducing manual operation time. The text block vector associated with GNN can quickly identify and extract key information and automatically fill it into the specified template, greatly improving the efficiency and accuracy of data entry. The present application constructs a document knowledge graph, and the system can integrate data from different sources to ensure the integrity and consistency of the information. The multi-layer aggregation mechanism of GNN can capture the complex relationships between nodes and improve the accuracy of data processing. The system supports manual intervention and feedback mechanisms, and for uncertain parts of the AI processing result, professional personnel can perform review and correction to ensure the accuracy of the final result. The present application continuously accumulates and updates training data, and the system can gradually adapt to professional terms and knowledge in different fields to improve its processing capability in complex scenarios. For example, by introducing the knowledge of domain experts, the system can better understand the content of documents in specific fields. Through continuous learning and optimization, the performance and effect can be continuously improved, and the long-term maintenance cost can be reduced. With the development of technology, the system can be continuously upgraded to maintain its competitiveness in the industry.

[0083] The intelligent document processing system based on document knowledge graph and GNN not only significantly improves the efficiency and quality of enterprise document processing, but also builds a powerful internal knowledge base for enterprises, providing long-term sustainable development support.

[0084] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A document review system based on AI intelligent agents, characterized in that, include: The document processing unit is used to acquire documents of various formats and preprocess them using a format conversion model to obtain preprocessed structured text data. The data extraction unit is used to parse structured data based on the data extraction model and extract enhanced metadata and optimized text blocks. The data processing unit is used to process the enhanced metadata and optimized text blocks using a data transformation model to obtain a first-order vector representation; and to associate the first-order vector representation based on a text perception model to generate an associated second-order vector representation. The text generation unit is used to build a vector knowledge base, collect and store second-order vector representations, and generate reviewed text using a document generation model. The iterative optimization unit is used to collect user corrections and evaluations of the generated reviewed text using a cyclical feedback mechanism, and to further iteratively optimize the document generation model.

2. The document review system based on AI intelligent agents according to claim 1, characterized in that, The document processing unit includes: The document classification module is used to acquire documents of various formats and use a format analyzer to classify the acquired documents into text documents and image documents. The document recognition module is used to process text documents and image documents based on text recognition algorithms and image recognition algorithms respectively, to obtain structured text data.

3. The document review system based on AI intelligent agents according to claim 1, characterized in that, The data extraction unit includes: The data augmentation module is used to identify nouns in structured data using a pre-trained language model, define and classify them, and generate augmented metadata. The data optimization module is used to improve the semantic coherence and completeness of text blocks in structured data by using an improved pre-trained language model, and generate optimized text blocks.

4. The document review system based on AI intelligent agents according to claim 1, characterized in that, The data processing unit includes: The data transformation module is used to vectorize the enhanced metadata and optimized text blocks using the data transformation model, resulting in metadata vectors and text block vectors. The data construction module is used to build document knowledge graphs based on metadata vectors and text block vectors; The data association module is used to receive document knowledge graphs using graph neural networks, associate the document knowledge graphs, and generate associated second-order vector representations.

5. A document review system based on an AI intelligent agent according to claim 4, characterized in that, The construction of the document knowledge graph based on metadata vectors and text block vectors includes: Transform each metadata vector and each text block vector into metadata nodes and text block nodes; Based on the relation extraction model, relationship edges between metadata nodes and text block nodes are established to construct a document knowledge graph.

6. A document review system based on an AI intelligent agent according to claim 5, characterized in that, The process of receiving a document knowledge graph using a graph neural network, associating the document knowledge graphs, and generating associated second-order vector representations includes: The message passing mechanism of a graph neural network is used to receive and associate adjacent nodes in the document knowledge graph, generating associated second-order vector representations.

7. A document review system based on an AI intelligent agent according to claim 6, characterized in that, The process of establishing relationship edges between metadata nodes and text block nodes based on the relation extraction model, and constructing a document knowledge graph, includes: An attention-based entity recognition model captures contextual information in text, identifies relationships between nodes, extracts relationships between metadata nodes and text block nodes, and constructs a document knowledge graph. By training the model with labeled data and optimizing the parameters of the entity recognition model, the accuracy and robustness of relation extraction can be improved.

8. A document review system based on an AI intelligent agent according to claim 7, characterized in that, The step of using the message passing mechanism of a graph neural network to receive and associate adjacent nodes in the document knowledge graph, and generating associated second-order vector representations, includes: The attention mechanism is used to aggregate adjacent nodes in the document knowledge graph and generate associated second-order vector representations. By using a graph attention network to aggregate adjacent nodes in the document knowledge graph and then using an attention mechanism to aggregate them, a second-order vector representation of the generated associations is obtained.

9. A document review system based on an AI intelligent agent according to claim 1, characterized in that, The text generation unit includes: The text storage module is used to store each second-order vector representation as a key-value pair in the constructed vector database, where the key-value pair contains the second-order vector representation and the payload; The text matching module is used to calculate similarity based on second-order vector representation using a vector knowledge base, and retrieves the second-order vector representation of similarity matching through a similarity management mechanism. The text generation module is used to generate reviewed text by taking the second-order vector representation of similarity matching as the key context and inputting it into the document generation model based on the vector generation mechanism.

10. A document review system based on an AI intelligent agent according to claim 1, characterized in that, The iterative optimization unit includes: The iterative optimization module is used to build a cyclical feedback channel, store the corrections and evaluations of the reviewed text, and feed them back to the document generation model for continuous optimization.

Images