Document generation method and device based on artificial intelligence

Abstract

The embodiment of the invention provides a document generation method and device based on artificial intelligence, and the method comprises the steps: extracting multi-modal data features of a historical multi-modal file through natural language processing and a computer vision algorithm, and constructing a corresponding knowledge graph according to the multi-modal data features, converting the multi-modal data subjected to feature extraction into high-dimensional vectors by using a vectorization algorithm, and storing the high-dimensional vectors to form a vector library; splitting a current document generation task into a plurality of sub-tasks by adopting a task planning algorithm, and executing the sub-tasks at the same time to obtain a first document; performing iterative optimization of document content according to the document first draft, the cue word and the language model to obtain an optimized first draft, acquiring related information from the vector library by using an attention mechanism-based retrieval algorithm, and fusing the related information and the optimized first draft through a text generation algorithm to obtain a final document; according to the invention, the accuracy and intelligence in the document generation process can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and specifically to a document generation method and device based on artificial intelligence. Background Art

[0002] In today's information age, document generation technology based on artificial intelligence is developing rapidly, but it still faces many challenges and limitations. Although existing document generation methods have made progress in some aspects, they cannot fully meet the complex needs of the modern information environment as a whole.

[0003] Traditional document generation methods mainly rely on manual writing or simple template filling. This method is inefficient when processing large amounts of diverse information and it is difficult to ensure the consistency and timeliness of document quality. With the development of artificial intelligence technology, some automated document generation tools have begun to emerge, but they can usually only process structured data or content in specific fields and lack the ability to process complex, unstructured information.

[0004] At present, there are some document generation systems based on natural language processing technology on the market. These systems can analyze text data, extract key information, and generate simple reports or summaries. However, they are unable to cope with multimodal data (such as images, audio, video, etc.), and it is difficult to fully utilize and integrate various types of information resources. In addition, these systems often lack a deep understanding and semantic grasp of the generated content, and the generated documents may have problems such as logical incoherence, repeated content, or lack of depth.

[0005] In terms of knowledge representation and management, most existing document generation technologies use simple keyword extraction or statistical analysis methods, lacking structured representation and deep understanding of knowledge. This limits the system's ability to generate high-quality documents with depth and breadth. Although some advanced systems have begun to try to use technologies such as knowledge graphs to enhance knowledge representation, these methods still face many challenges in practical applications, such as knowledge updating and cross-domain knowledge integration.

[0006] Document structure generation is another key challenge. Most existing systems rely on preset templates or fixed rules to generate document structures, which lacks flexibility and innovation. This approach is difficult to adapt to the needs of documents of different topics and styles, and the generated document structure may appear rigid or inappropriate. Although some systems have begun to try to use machine learning algorithms to optimize document structures, the effect still needs to be improved.

[0007] In terms of content generation and optimization, existing technologies usually adopt a one-time generation method, lacking the ability of dynamic adjustment and iterative optimization. This static generation method is difficult to ensure the coherence and logic of the document content, especially when dealing with long documents or complex topics. In addition, many systems often perform poorly when fusing and integrating information from different sources, making it difficult to generate truly original and high-quality content.

[0008] Performance and efficiency are also important challenges faced by existing document generation technologies. As data size increases and task complexity increases, many systems face performance bottlenecks when processing large-scale data and complex tasks. This not only affects the speed of document generation, but also limits the application of the system in real-time or near-real-time scenarios. Summary of the invention

[0009] In response to the problems in the prior art, the present application provides a document generation method and device based on artificial intelligence, which can effectively improve the accuracy and intelligence of the document generation process.

[0010] In order to solve at least one of the above problems, the present application provides the following technical solutions:

[0011] In a first aspect, the present application provides a document generation method based on artificial intelligence, comprising:

[0012] The multimodal data features of historical multimodal files are extracted using natural language processing and computer vision algorithms, and a corresponding knowledge graph is constructed based on the multimodal data features. The multimodal data after feature extraction is converted into a high-dimensional vector using a vectorization algorithm, and stored to form a vector library;

[0013] Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0014] Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0015] Furthermore, the method of extracting multimodal data features of historical multimodal files by using natural language processing and computer vision algorithms includes:

[0016] Performing data format normalization preprocessing on the historical multimodal files, performing word segmentation, part-of-speech tagging and syntactic analysis on the text data in the historical multimodal files through a natural language processing model, extracting word vectors and semantic features, and performing image enhancement, size adjustment and color space conversion preprocessing on the image data in the historical multimodal files through a computer vision algorithm;

[0017] The deep learning model is used to extract features from the processed text and image data respectively. The context-related semantic features of the text data are extracted through the attention mechanism. The multi-scale visual features of the image data are extracted using a convolutional neural network. The extracted feature vectors are normalized and dimensionally reduced to obtain multimodal data features with unified representation.

[0018] Furthermore, the multimodal data after feature extraction is converted into a high-dimensional vector using a vectorization algorithm and stored to form a vector library, including:

[0019] The principal component analysis method is used to reduce the dimension of the multimodal data after feature extraction to retain the main feature information. The t-distributed random neighbor embedding algorithm is used to nonlinearly map the reduced-dimensional data to a high-dimensional space to generate a vector representation of uniform dimension;

[0020] Construct an index structure based on hierarchical differentiable hashing, encode the normalized high-dimensional vector into a binary hash code, and use a distributed storage system to store the hash-encoded vector data and metadata index separately, establish associations between data blocks, and form a vector library.

[0021] Furthermore, the article structure outline is generated by using a graph neural network according to the knowledge graph and the vector library, and prompt words are dynamically generated according to a context-aware algorithm, including:

[0022] The graph attention network is used to capture the relationship and semantic information between knowledge graph nodes. The relevant vectors in the vector library are used as input features of the graph neural network. High-order features are extracted through multi-layer graph convolution operations. The node features are aggregated using the graph pooling method to generate a graph-level representation. The recurrent neural network is combined with the graph-level representation to generate the hierarchical relationship of the article structure outline, where the hierarchical relationship includes chapter titles and subtitles.

[0023] A contextual semantic tree is constructed based on the generated article structure outline, and a depth-first traversal is performed on the semantic tree. The self-attention mechanism is applied to each node to extract key information. The traversal sequence is processed using a bidirectional long short-term memory network to capture long-distance dependencies. The processed sequence is input into a pre-trained language model, and a beam search algorithm is used to generate multiple candidate prompt words. A sorting algorithm based on context relevance is applied to determine the prompt word that best suits the current context as the output.

[0024] Furthermore, the determining of the corresponding content of each chapter according to the article structure outline, the prompt words and the preset language model includes:

[0025] The article structure outline is converted into a sequence representation with title and hierarchical information, the prompt words are encoded using a multi-head attention mechanism, and a conditional text generation model based on the transformer architecture is used to take the outline sequence and prompt word encoding as conditional input;

[0026] The corresponding contents of each chapter are determined by the chapter-level decoder and paragraph-level decoder of the conditional text generation model.

[0027] Furthermore, the task planning algorithm is used to split the current document generation task into multiple subtasks and the subtasks are executed simultaneously to obtain a first draft of the document, including:

[0028] According to the task planning algorithm based on directed acyclic graph, a task dependency graph is constructed to represent the relationship between document generation subtasks, the task execution order is determined according to topological sorting, and the subtasks are prioritized through a heuristic algorithm;

[0029] Build a distributed task execution framework and apply parallel computing technology to execute multiple subtasks simultaneously.

[0030] Furthermore, the use of an attention-based retrieval algorithm to obtain relevant information from the vector library, and the fusion of the relevant information with the optimized draft through a text generation algorithm to obtain a final document includes:

[0031] According to the vector retrieval algorithm based on the attention mechanism, the optimized draft is converted into a query vector, the similarity score between the query vector and the vectors in the vector library is calculated, the vector subset with the highest score is selected, and the inverted index structure is used to quickly locate and extract relevant information.

[0032] A fusion text generation model is constructed to encode the retrieved relevant information and the optimized draft as an input sequence to obtain a final document output by the fusion text generation model.

[0033] In a second aspect, the present application provides a document generation device based on artificial intelligence, comprising:

[0034] A vector library construction module is used to extract multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, and to construct a corresponding knowledge graph based on the multimodal data features, and to convert the multimodal data after feature extraction into a high-dimensional vector using a vectorization algorithm, and store them to form a vector library;

[0035] A first draft determination module is used to generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding contents of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0036] The final draft determination module is used to iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0037] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the artificial intelligence-based document generation method when executing the program.

[0038] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the artificial intelligence-based document generation method.

[0039] In a fifth aspect, the present application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the artificial intelligence-based document generation method.

[0040] It can be seen from the above technical scheme that the present application provides a document generation method and device based on artificial intelligence, which extracts multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and constructs a corresponding knowledge graph based on the multimodal data features, and uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, which is stored to form a vector library; a task planning algorithm is used to split the current document generation task into multiple subtasks and the subtasks are executed simultaneously to obtain a document draft; the document content is iteratively optimized according to the document draft, the prompt words and the language model to obtain an optimized draft, and a retrieval algorithm based on an attention mechanism is used to obtain relevant information from the vector library, and the relevant information is integrated with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the accuracy and intelligence of the document generation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 This is one of the flowcharts of the document generation method based on artificial intelligence in the embodiment of the present application;

[0043] Figure 2 This is a second flow chart of the document generation method based on artificial intelligence in an embodiment of the present application;

[0044] Figure 3 The third flowchart of the document generation method based on artificial intelligence in the embodiment of the present application;

[0045] Figure 4 This is a fourth flowchart of the document generation method based on artificial intelligence in an embodiment of the present application;

[0046] Figure 5 This is a fifth flowchart of the document generation method based on artificial intelligence in an embodiment of the present application;

[0047] Figure 6 This is the sixth flow chart of the document generation method based on artificial intelligence in the embodiment of the present application;

[0048] Figure 7 This is the seventh flow chart of the document generation method based on artificial intelligence in the embodiment of the present application;

[0049] Figure 8 This is a structural diagram of an artificial intelligence-based document generation device in an embodiment of the present application;

[0050] Fig. 9 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application.

[0051] Reference numerals:

[0052] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver program storage unit 9144, antenna 9111, speaker 9131, microphone 9132. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0054] The acquisition, storage, use, and processing of data in the technical solution of this application comply with the relevant provisions of national laws and regulations.

[0055] In view of the problems existing in the prior art, the present application provides a document generation method and device based on artificial intelligence, which extracts multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and constructs a corresponding knowledge graph based on the multimodal data features, and uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, which is stored to form a vector library; a task planning algorithm is used to split the current document generation task into multiple subtasks and the subtasks are executed simultaneously to obtain a document draft; the document content is iteratively optimized according to the document draft, the prompt words and the language model to obtain an optimized draft, and a retrieval algorithm based on an attention mechanism is used to obtain relevant information from the vector library, and the relevant information is integrated with the optimized draft through a text generation algorithm to obtain a final document, thereby effectively improving the intelligence of document generation.

[0056] In order to effectively improve the accuracy and intelligence of the document generation process, the present application provides an embodiment of a document generation method based on artificial intelligence, see Figure 1 , the document generation method based on artificial intelligence specifically includes the following contents:

[0057] Step S101: extracting multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, constructing corresponding knowledge graphs based on the multimodal data features, converting the feature-extracted multimodal data into high-dimensional vectors using a vectorization algorithm, and storing them to form a vector library;

[0058] Optionally, in this embodiment, step S101 is a complex data processing and feature extraction process, which aims to extract valuable information from historical multimodal documents and convert it into a structured form that can be used for subsequent intelligent processing. This step involves multiple technical fields such as natural language processing, computer vision, knowledge graph construction and vectorization, and is the foundation and key to the entire artificial intelligence-based document generation method.

[0059] First, the system needs to process historical multimodal files. These files may contain data in multiple forms such as text, images, audio, and video. For text data, the system uses advanced natural language processing techniques, such as pre-trained models such as BERT or GPT, for word segmentation, part-of-speech tagging, named entity recognition, and semantic analysis. These models are able to understand the contextual semantics of the text and extract key information and topics. For image data, the system uses deep learning models such as convolutional neural networks (CNNs) for feature extraction. This may include the use of pre-trained models such as ResNet or VGG, which are able to identify objects, scenes, and attributes in images. For audio and video data, the system may use specialized deep learning models such as waveform to spectrogram conversion and convolutional neural network combinations for audio processing, or 3D convolutional neural networks for video processing.

[0060] During the feature extraction process, a major technical challenge faced by the system is how to effectively fuse and align features from different modalities. To this end, the system may adopt multimodal fusion techniques, such as attention mechanisms or cross-modal transformer models, to capture the interrelationships and semantic consistency between different modalities. This approach can generate a comprehensive, multi-dimensional feature representation, laying the foundation for subsequent knowledge graph construction and vectorization.

[0061] Based on the extracted multimodal data features, the system then builds a knowledge graph. This process involves entity recognition, relationship extraction, and ontology construction. The system may use graph neural networks or graph embedding techniques to learn the representation of entities and relationships, and use reasoning algorithms to discover implicit knowledge and relationships. The construction of a knowledge graph can not only organize and structure the extracted information, but also reveal the deep semantic associations between the data, providing rich background knowledge and contextual information for subsequent document generation.

[0062] At the same time, the system uses a vectorization algorithm to convert the extracted multimodal data features into high-dimensional vectors. This process may involve dimensionality reduction techniques such as t-SNE or UMAP, as well as encoding techniques such as autoencoders or variational autoencoders. The purpose of vectorization is to convert complex multimodal data into a unified mathematical representation to facilitate subsequent similarity calculations and retrieval. High-dimensional vectors can capture the subtle features and potential semantics of the data, but they also bring the challenge of "dimensionality disaster". To this end, the system may use technologies such as local sensitive hashing (LSH) or approximate nearest neighbor (ANN) to optimize storage and retrieval efficiency.

[0063] Finally, the system stores the generated high-dimensional vectors to form a vector library. This process may involve the design of a distributed storage system and an efficient index structure. The system may use specialized vector retrieval engines such as Elasticsearch or Faiss to support fast retrieval and similarity calculation of large-scale, high-dimensional vectors.

[0064] This step solves the technical problem of how to extract valuable information from complex multimodal historical data and convert it into a structured form that can be used for intelligent processing. Its main technical effects include: First, through multimodal feature extraction, the system can fully understand and represent complex document content. Secondly, the construction of the knowledge graph enables the system to capture and utilize the semantic relationship between data. Furthermore, the establishment of vectorization and vector library provides an efficient foundation for subsequent similarity calculation and information retrieval.

[0065] This method not only improves the efficiency of historical data utilization, but also provides a rich knowledge base and efficient retrieval mechanism for subsequent intelligent document generation. It has broad application prospects in the fields of automatic document generation, intelligent information retrieval, knowledge management, etc., and can significantly improve the intelligent level of information processing and knowledge utilization, and provide users with more accurate and comprehensive information services.

[0066] Step S102: Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0067] Optionally, in this embodiment, step S102 is a complex document generation process, which combines a variety of advanced technologies such as knowledge graphs, vector libraries, graph neural networks, context-aware algorithms, preset language models, and task planning algorithms, aiming to intelligently generate high-quality document drafts. This step not only involves the generation of article structure, but also includes the creation of content and the parallel execution of tasks, and is the core link in the entire document generation method.

[0068] First, the system uses the knowledge graph and vector library as input to generate an article structure outline through a graph neural network. In this process, the knowledge graph provides a structured representation of domain knowledge, while the vector library provides rich semantic information. Graph neural networks, such as graph attention networks (GAT) or graph convolutional networks (GCN), can effectively process and learn graph structure data. It captures complex semantic relationships and knowledge structures by message passing and feature aggregation of nodes and edges in the knowledge graph. At the same time, it can also integrate the semantic information in the vector library to generate an article structure outline that is consistent with the domain knowledge structure and rich in semantic connotations. This method solves the problem of lack of logic and coherence in the structure of traditional document generation, ensuring that the generated outline has a clear hierarchy and reflects the intrinsic connection between topics.

[0069] Next, the system uses a context-aware algorithm to dynamically generate prompt words. The key to this step is to understand the context of the current generation task and generate appropriate prompt words to guide subsequent content generation. The context-aware algorithm may be based on a recurrent neural network (RNN) or a transformer model, which can capture long-distance dependencies and contextual information. By analyzing the generated outline structure and previous content, the algorithm can predict the topics and key concepts that may be involved in the next step, thereby generating relevant prompt words. This method of dynamically generating prompt words solves the problem that static prompt words cannot adapt to different content and contexts, and improves the relevance and coherence of the generated content.

[0070] The system then determines the specific content of each chapter based on the generated article structure outline, dynamic prompt words, and preset language models. The preset language model may be a large-scale pre-trained model based on the transformer architecture, such as the GPT series. In this process, the article structure outline provides the overall framework, the dynamic prompt words provide specific directions, and the language model is responsible for generating smooth and coherent text content. The generation process of the language model may adopt conditional generation or guided generation methods to ensure that the generated content conforms to the outline structure and is closely related to the theme of the prompt words. This method effectively solves the problems of inconsistent content and divergent topics in traditional document generation, and ensures the quality and relevance of the generated content.

[0071] Finally, the system uses a task planning algorithm to split the document generation task into multiple subtasks and execute these subtasks simultaneously to improve efficiency. The task planning algorithm may be based on a directed acyclic graph (DAG) to represent the dependencies between tasks and use heuristic algorithms such as A* to optimize the order of task execution. Parallel execution of subtasks may involve distributed computing techniques such as MapReduce or more modern distributed deep learning frameworks. This approach solves the efficiency problem in large-scale document generation and significantly reduces the generation time while ensuring the coordination and consistency of the content of each part.

[0072] The main technical effects of this step include: First, the article structure outline generated by the graph neural network has strong logic and high topic relevance. Second, the dynamically generated prompt words can accurately grasp the context and guide the generation of more appropriate content. Third, the application of the preset language model ensures the fluency and coherence of the generated content. Finally, task planning and parallel execution greatly improve the efficiency of document generation.

[0073] Step S103: Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0074] Optionally, in this embodiment, step S103 is the final stage of the document generation process, which improves the document quality through iterative optimization and information fusion, and finally generates a high-quality final document. This step involves advanced technologies such as multiple rounds of iteration, attention mechanism retrieval and text generation fusion, aiming to solve the deficiencies that may exist in the first draft of the document and further enrich and optimize the document content.

[0075] First, the system iteratively optimizes the document content based on the first draft of the document, the prompt words, and the preset language model. This process adopts an iterative optimization strategy based on reinforcement learning. The system inputs the first draft of the document into the language model and uses the prompt words as the optimization target. The language model may be a large-scale pre-trained model based on the transformer architecture, such as GPT-3 or BERT. In each round of iteration, the system generates multiple optimized versions and then scores these versions using predefined evaluation indicators (such as relevance, coherence, information richness, etc.). The version with the highest score is selected as the basis for the next round of iteration. This process is repeated until the preset number of iterations or quality threshold is reached. This method solves the problems of unclear expression and incoherent logic that may exist in the first draft of the document, and improves the overall quality of the document through multiple rounds of optimization.

[0076] Next, the system uses an attention-based retrieval algorithm to obtain relevant information from the vector library. This process first converts the optimized draft into a query vector. The conversion may use the same encoding method as the vector library, such as models such as BERT or Sentence-BERT. The system then calculates the similarity between the query vector and all the vectors in the vector library. Metrics such as cosine similarity or dot product may be used here. The attention mechanism plays a key role in this process. It can dynamically adjust the attention to different vectors in the vector library according to different parts of the query vector. For example, for the part of the document that discusses a specific topic, the system will pay more attention to the vectors in the vector library that are related to that topic. This approach solves the problem that traditional retrieval methods are difficult to capture the subtle semantic differences of queries, and improves the accuracy and relevance of retrieval results.

[0077] Finally, the system uses a text generation algorithm to fuse the retrieved relevant information with the optimized draft to obtain the final document. This fusion process uses an advanced conditional text generation model, such as the GPT-3 fine-tuned version or the T5 model. The system uses the optimized draft as the input sequence and the retrieved relevant information as the conditional information. Through the attention mechanism and multi-head self-attention layer, the model can understand the content of the draft and appropriately incorporate new information. During the generation process, the model dynamically adjusts the attention to the draft and new information to ensure that the generated content not only maintains the original structure and theme, but also naturally incorporates new and valuable information. This method solves the problem that simple splicing or replacement is difficult to produce coherent and natural text, ensuring the fluency and information richness of the final document.

[0078] The main technical effects of this step include: First, the quality of the document is significantly improved through iterative optimization, including the clarity of expression, logical coherence and content accuracy. Second, the retrieval algorithm based on the attention mechanism greatly improves the accuracy and efficiency of relevant information acquisition. Finally, the advanced text generation and fusion technology ensures that new information can be naturally and seamlessly integrated into the document, improving the overall quality and information value of the document.

[0079] This approach not only improves the quality and reliability of automatically generated documents, but also provides new possibilities in the fields of intelligent content creation, automatic report generation, personalized document customization, etc. It can significantly reduce the time and effort of manual editing and modification of documents while ensuring the professionalism, accuracy and information richness of the generated content.

[0080] From the above description, it can be seen that the artificial intelligence-based document generation method provided in the embodiment of the present application can extract multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and construct a corresponding knowledge graph based on the multimodal data features, and use a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, and store it to form a vector library; use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a document draft; iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the intelligence of document generation.

[0081] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 2 , and can also include the following:

[0082] Step S201: performing data format normalization preprocessing on the historical multimodal file, performing word segmentation, part-of-speech tagging and syntactic analysis on the text data in the historical multimodal file through a natural language processing model, extracting word vectors and semantic features, and performing image enhancement, size adjustment and color space conversion preprocessing on the image data in the historical multimodal file through a computer vision algorithm;

[0083] Step S202: Use the deep learning model to perform feature extraction on the processed text and image data respectively, extract context-related semantic features from the text data through the attention mechanism, and use the convolutional neural network to extract multi-scale visual features from the image data, normalize and reduce the dimension of the extracted feature vectors to obtain multimodal data features with unified representation.

[0084] Optionally, in this embodiment, steps S201 and S202 are a complex multimodal data preprocessing and feature extraction process, which aims to convert historical files of different formats into a unified data representation that can be used for subsequent processing. These two steps involve multiple technical fields such as natural language processing, computer vision and deep learning, and are the basic links of the entire intelligent document generation method.

[0085] In step S201, the system first performs data format normalization preprocessing on the historical multimodal files. This process aims to solve the inconsistency problems that may exist in files of different sources and formats. The system may use various file format conversion tools and parsing libraries, such as Python's PyPDF2 to process PDF files, docx to process Word documents, and PIL to process image files. This step ensures that all input data is converted into a format that the system can uniformly process, laying the foundation for subsequent feature extraction.

[0086] For text data, the system uses natural language processing models for word segmentation, part-of-speech tagging, and syntactic analysis. The word segmentation process may use tools such as jieba (Chinese) or NLTK (English) to divide the text into the smallest meaningful units based on language characteristics. Part-of-speech tagging uses pre-trained models such as Stanford POS Tagger to assign part-of-speech tags to each word. Syntactic analysis may use dependency syntactic analysis or component syntactic analysis methods, using tools such as SpaCy or Stanford Parser to build a grammatical structure tree for the sentence. These processing steps help the system understand the linguistic structure of the text and prepare for subsequent semantic feature extraction.

[0087] The extraction of word vectors and semantic features is a key step in text processing. The system may use pre-trained word embedding models such as Word2Vec, GloVe, or FastText to map each word to a high-dimensional vector space. For more complex semantic features, the system may use transformer-based models such as BERT or RoBERTa, which can capture the dynamic meaning of words in context.

[0088] For image data, the system performs a series of preprocessing steps. Image enhancement may include operations such as contrast adjustment, noise removal, and sharpening, which help improve image quality and highlight important features. Resizing ensures that all images have a uniform size, which is crucial for subsequent convolutional neural network processing. Color space conversion, such as from RGB to HSV or Lab space, can help the system better capture the color characteristics of the image. These preprocessing steps solve problems such as inconsistent quality, size, and color representation that may exist in the original image data.

[0089] In step S202, the system uses a deep learning model to extract features from the preprocessed text and image data. For text data, the system uses a model with an attention mechanism, such as Transformer or BERT. The attention mechanism allows the model to dynamically focus on different parts of the input sequence, effectively capturing long-distance dependencies and contextual information. This method solves the problem that traditional RNN models are difficult to process long sequences, and improves the accuracy and efficiency of semantic feature extraction.

[0090] For image data, the system uses convolutional neural networks (CNNs) to extract multi-scale visual features. Commonly used CNN architectures such as ResNet, VGG, or Inception networks can automatically learn the hierarchical features of images. By using convolution kernels of different sizes and pooling operations, CNNs can capture multi-scale features from local textures to global structures. This method solves the problem that traditional manually designed features are difficult to adapt to complex image content, greatly improving the effect of visual feature extraction.

[0091] Finally, the system normalizes and reduces the dimensionality of the extracted feature vectors. Normalization may use methods such as L2 regularization to ensure that the scales of different features are consistent. Dimensionality reduction may use techniques such as principal component analysis (PCA) or t-SNE to reduce the dimensionality of the features while retaining key information. This step solves the dimensionality disaster problem that may be caused by high-dimensional features and improves the efficiency of subsequent processing.

[0092] The main technical effects of these two steps include: First, through unified preprocessing, the system can process input data of various formats. Second, the application of deep learning models greatly improves the quality and efficiency of feature extraction. Furthermore, the use of attention mechanism and CNN enables the system to capture more complex and subtle data features. Finally, through normalization and dimensionality reduction, the system obtains multimodal data features with unified representation, laying a solid foundation for subsequent knowledge graph construction and document generation.

[0093] This approach not only improves the efficiency and accuracy of multimodal data processing, but also provides new possibilities in areas such as intelligent document analysis, cross-modal information retrieval, and multimodal content generation.

[0094] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 3 , and can also include the following:

[0095] Step S301: Apply principal component analysis to reduce the dimension of the multimodal data after feature extraction to retain the main feature information. Use the t-distributed random neighbor embedding algorithm to nonlinearly map the reduced-dimensional data to a high-dimensional space to generate a vector representation of uniform dimension;

[0096] Step S302: construct an index structure based on hierarchical derivable hashing, encode the normalized high-dimensional vector into a binary hash code, and use a distributed storage system to store the hash-encoded vector data and metadata index separately, establish associations between data blocks, and form a vector library.

[0097] Optionally, in this embodiment, steps S301 and S302 are a complex data dimension reduction, vectorization and index construction process, which aims to convert high-dimensional multimodal data into a more compact and efficient representation and establish a vector library that can be quickly retrieved. These two steps involve multiple advanced machine learning technologies and are key links in achieving efficient multimodal data management and retrieval.

[0098] In step S301, the system first applies principal component analysis (PCA) to the multimodal data after feature extraction for dimensionality reduction. PCA is a linear dimensionality reduction method that converts variables that may be correlated into a set of linearly independent variables, namely principal components, through orthogonal transformation. In this process, the system calculates the covariance matrix of the data and then solves its eigenvalues ​​and eigenvectors. The eigenvectors are sorted by the size of the corresponding eigenvalues, and the first k eigenvectors are selected to form a projection matrix. By multiplying the original data by this projection matrix, the reduced-dimensional data can be obtained. The advantage of PCA is that it can significantly reduce the data dimension while retaining the main information, solving the "dimensionality disaster" problem in high-dimensional data processing and improving the efficiency of subsequent processing.

[0099] Next, the system uses the t-distributed stochastic neighbor embedding (t-SNE) algorithm to nonlinearly map the reduced data to a high-dimensional space. t-SNE is a nonlinear dimensionality reduction algorithm that is particularly suitable for visualizing high-dimensional data into two-dimensional or three-dimensional space. Its core idea is to convert the similarity between data points in a high-dimensional space into the distance between points in a low-dimensional space. Specifically, t-SNE first calculates the conditional probability distribution between data points in a high-dimensional space, then constructs a similar distribution in a low-dimensional space, and optimizes the low-dimensional representation by minimizing the KL divergence between the two distributions. The use of t-SNE solves the limitations of linear dimensionality reduction methods (such as PCA) in dealing with complex nonlinear structures, and can better preserve the local and global structures of the data.

[0100] The combination of these two steps not only retains the main feature information of the data, but also captures the nonlinear structure of the data, and finally generates a vector representation of uniform dimension. This representation provides an ideal data format for subsequent indexing and retrieval.

[0101] In step S302, the system constructs an index structure based on hierarchical derivable hashing. Derivable hashing is a deep learning based hashing method that directly learns the mapping from raw data to binary hash codes by training a neural network. The hierarchical structure allows the system to represent and index data at different granularities. The advantage of this method is that it can generate semantically sensitive hash codes, that is, semantically similar data points will be mapped to similar hash codes.

[0102] In the specific implementation process, the system may use a multi-layer neural network, with each layer corresponding to a hash bit. The network's loss function usually includes quantization loss (ensuring that the output is close to binary) and similarity preservation loss (ensuring that the hash codes of similar data points are also similar). By optimizing the network parameters through the back-propagation algorithm, the system eventually learns a mapping function from high-dimensional vectors to binary hash codes.

[0103] After encoding the high-dimensional vector into a binary hash code, the system uses a distributed storage system to store the hash-coded vector data and metadata indexes respectively. This may be done using a Hadoop Distributed File System (HDFS) or a distributed key-value storage system such as HBase. The use of distributed storage solves the problem of large-scale data storage and fast access, and improves the scalability and fault tolerance of the system.

[0104] At the same time, the system establishes associations between data blocks to form a complete vector library. This process may involve creating an inverted index, that is, maintaining a list of all data points containing each hash code. This index structure can support efficient similarity searches, such as quickly finding similar hash codes through Hamming distance.

[0105] The main technical effects of these two steps include: First, through the combination of PCA and t-SNE, the system can effectively reduce the data dimension while retaining key structural information. Second, the use of hierarchical derivable hashing enables the system to generate semantically sensitive binary representations, greatly improving retrieval efficiency. Furthermore, the application of distributed storage solves the problem of large-scale data management and improves the scalability of the system. Finally, by establishing associations between data blocks, the system forms an efficient and queryable vector library, laying the foundation for subsequent rapid retrieval and information extraction.

[0106] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 4 , and can also include the following:

[0107] Step S401: Use the graph attention network to capture the relationship and semantic information between knowledge graph nodes, use the relevant vectors in the vector library as the input features of the graph neural network, extract high-order features through multi-layer graph convolution operations, use the graph pooling method to aggregate the node features, generate a graph-level representation, and use a recurrent neural network combined with the graph-level representation to generate a hierarchical relationship of the article structure outline, where the hierarchical relationship includes chapter titles and subtitles;

[0108] Step S402: construct a context semantic tree based on the generated article structure outline, perform depth-first traversal on the semantic tree, apply a self-attention mechanism at each node to extract key information, use a bidirectional long short-term memory network to process the traversal sequence, capture long-distance dependencies, input the processed sequence into a pre-trained language model, use a beam search algorithm to generate multiple candidate prompt words, and apply a sorting algorithm based on context relevance to determine the prompt word that best suits the current context as the output.

[0109] Optionally, in this embodiment, steps S401 and S402 are a complex knowledge graph processing and document structure generation process, which aims to use advanced graph neural network technology and natural language processing methods to extract structured information from vectorized knowledge representation and generate the outline of the article and corresponding prompt words. These two steps involve multiple advanced technologies such as graph attention network, graph convolution network, recurrent neural network and pre-trained language model, and are the core links for realizing intelligent document generation.

[0110] In step S401, the system first uses the graph attention network (GAT) to capture the relationship and semantic information between knowledge graph nodes. The core idea of ​​GAT is to assign different attention weights to each node, so that the network can adaptively focus on important neighbor nodes. This method solves the limitations of traditional graph convolutional networks in processing heterogeneous graph structures and can better capture the complex relationships between nodes.

[0111] Next, the system uses the relevant vectors in the vector library as input features of the graph neural network and extracts high-order features through multi-layer graph convolution operations. The graph convolution operation is similar to the traditional convolutional neural network, but acts on graph structured data. It updates the representation of the central node by aggregating the information of neighboring nodes, thereby capturing the local structural information of the node in the graph. The use of multi-layer graph convolution enables the network to capture node relationships at longer distances and extract more abstract and advanced features.

[0112] Graph pooling methods are used to aggregate node features and generate graph-level representations. Commonly used graph pooling methods include average pooling, maximum pooling, or more complex differentiable pooling. This step compresses node-level features into a unified graph representation, providing a global perspective for subsequent article structure generation.

[0113] Finally, the system uses a recurrent neural network (RNN) combined with a graph-level representation to generate the hierarchical relationship of the article structure outline. RNN is particularly suitable for processing sequence data and can capture the temporal dependencies between elements. By using the graph-level representation as the initial state or additional input of the RNN, the system can consider global semantic information when generating chapter titles and subtitles, ensuring that the generated structure is consistent with the content of the knowledge graph and has good hierarchy and coherence.

[0114] In step S402, the system constructs a contextual semantic tree based on the generated article structure outline. The semantic tree may be constructed in a recursive manner, with each node representing a chapter or subheading, and the relationship between nodes reflects the hierarchical structure of the article. This tree structure provides a clear context for subsequent information extraction and generation.

[0115] The system performs a depth-first traversal of the semantic tree, which can maintain the hierarchical structure of the article. At each node, the system applies the self-attention mechanism to extract key information. The self-attention mechanism allows the model to dynamically pay attention to other related nodes when processing each node, effectively capturing the long-distance dependencies between nodes.

[0116] Next, the system uses a bidirectional long short-term memory network (Bi-LSTM) to process the traversal sequence. The advantage of Bi-LSTM is that it can consider both forward and backward context information at the same time and better capture long-distance dependencies. This is crucial for understanding the entire document structure and generating coherent content.

[0117] The processed sequence is fed into a pre-trained language model. Pre-trained language models (such as BERT or GPT) have strong language understanding and generation capabilities, and can generate fluent and coherent text based on context. The system uses a beam search algorithm to generate multiple candidate prompt words. Beam search is a heuristic search algorithm that retains multiple best candidates at each step, balancing generation quality and diversity.

[0118] Finally, the system applies a context-based ranking algorithm to determine the most appropriate prompt word for the current context. This ranking process may consider factors such as the semantic similarity between the candidate prompt word and the current node content, and the consistency with the overall document theme. In this way, the system can select the most appropriate prompt word to guide the subsequent document content generation.

[0119] The main technical effects of these two steps include: First, through the graph neural network and attention mechanism, the system can effectively extract structured and semantic information from the knowledge graph. Second, the RNN and semantic tree-based method enables the system to generate an article outline with a clear hierarchical structure. Furthermore, the depth-first traversal combined with the use of Bi-LSTM ensures that the system can fully consider contextual information. Finally, the combination of pre-trained language model and beam search improves the quality and diversity of generated prompt words.

[0120] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 5 , and can also include the following:

[0121] Step S501: convert the article structure outline into a sequence representation with title and level information, use a multi-head attention mechanism to encode the prompt words, and use a conditional text generation model based on the transformer architecture to take the outline sequence and prompt word encoding as conditional input;

[0122] Step S502: Determine the corresponding contents of each chapter through the chapter-level decoder and paragraph-level decoder of the conditional text generation model.

[0123] Optionally, in this embodiment, steps S501 and S502 constitute an intelligent text generation system based on advanced natural language processing technology, which aims to transform the structured article outline and prompt words into complete and coherent article content. These two steps make full use of the advantages of the Transformer architecture and multi-level decoder, realize the step-by-step generation process from high-level structure to specific content, and are the key links to achieve high-quality automatic article generation.

[0124] In step S501, the system first converts the article structure outline into a sequence representation with titles and hierarchical information. This process may involve flattening the tree-structured outline into a linear sequence while retaining the hierarchical information of each title. For example, special markers may be used to indicate the beginning and end of different levels, or a hierarchical prefix may be added to each title. This sequence representation enables structured information to be effectively processed by the neural network model.

[0125] Next, the system encodes the cue words using a multi-head attention mechanism. Multi-head attention is one of the core components of the Transformer architecture, which allows the model to simultaneously focus on different parts of the input sequence and different representation subspaces. For each cue word, the multi-head attention mechanism calculates multiple sets of attention weights, each focusing on a different aspect of the input, and then combines this information to form a rich representation. This approach can capture the complex relationships and contextual dependencies between cue words, providing richer and more accurate semantic information for subsequent text generation.

[0126] The system uses a conditional text generation model based on the Transformer architecture, taking the outline sequence and prompt word encoding as conditional input. The Transformer model is known for its powerful parallel processing capabilities and ability to capture long-distance dependencies, making it very suitable for tasks that require consideration of global structure and local details. The use of the conditional text generation model allows the system to consider both the outline structure and the semantic information of the prompt words when generating text, ensuring that the generated content conforms to the predetermined structure and is closely related to the topic.

[0127] In step S502, the system determines the corresponding content of each chapter through the chapter-level decoder and paragraph-level decoder of the conditional text generation model. The design of this multi-level decoder reflects the hierarchical and progressive nature of article writing, and can better control the structure and coherence of the generated content.

[0128] The chapter-level decoder first processes the high-level structure of the entire article. It may generate the main topics and key points of each chapter based on the chapter titles in the outline sequence. During this process, the decoder will consider the logical relationship and content distribution between chapters to ensure that the entire article is well structured and the theme is clear.

[0129] The paragraph-level decoder is responsible for generating the specific paragraph content within each chapter. It generates more detailed and coherent text based on the output of the chapter-level decoder and the semantic information of the prompt words. The paragraph-level decoder will continue to pay attention to contextual information during the generation process to ensure natural transitions between paragraphs and coherence of content.

[0130] The collaborative work of these two levels of decoders reflects the generation strategy from macro to micro. The chapter-level decoder ensures the rationality of the overall structure, while the paragraph-level decoder ensures the richness and accuracy of the specific content. This hierarchical decoding strategy can effectively deal with the challenges in long text generation, such as maintaining thematic consistency, avoiding content duplication, and maintaining logical coherence.

[0131] The main technical effects of this method include: First, by converting the outline into a sequence representation and combining it with the prompt word encoding, the system can effectively fuse structural information and semantic information to provide rich context for text generation. Secondly, the conditional text generation model based on Transformer has powerful parallel processing capabilities and long-distance dependency capture capabilities, and can generate high-quality and coherent text. Furthermore, the design of the multi-level decoder enables the system to take into account both the overall structure and local details of the article, thereby improving the structured degree and logic of the generated content. Finally, this method greatly improves the quality and efficiency of automatic article generation, making the generated content closer to the level of human writing.

[0132] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 6 , and can also include the following:

[0133] Step S601: construct a task dependency graph according to a task planning algorithm based on a directed acyclic graph to represent the relationship between document generation subtasks, determine the task execution order according to topological sorting, and prioritize the subtasks through a heuristic algorithm;

[0134] Step S602: construct a distributed task execution framework and apply parallel computing technology to execute multiple subtasks simultaneously.

[0135] Optionally, in this embodiment, steps S601 and S602 constitute an efficient document generation task management and execution system, which aims to optimize resource allocation and execution efficiency in the complex document generation process through scientific task planning and advanced parallel computing technology. These two steps make full use of the advantages of graph theory, task scheduling and distributed computing, realize the optimization of the whole process from task dependency analysis to parallel execution, and are the key links to improve the efficiency of large-scale document generation.

[0136] In step S601, the system first constructs a task dependency graph based on a task planning algorithm based on a directed acyclic graph (DAG). This process involves decomposing a complex document generation task into multiple interrelated subtasks and representing the dependencies between them in the form of a graph. In this DAG, each node represents a subtask and the edges represent the dependencies between tasks. For example, the task of "generating chapter summaries" may depend on the task of "analyzing outline structure". The use of DAG ensures that there is no circular dependency between tasks and provides a clear logical framework for subsequent task scheduling.

[0137] Next, the system determines the order of task execution based on topological sorting. Topological sorting is an algorithm that sorts DAG nodes in such a way that for every directed edge (u, v) in the graph, node u appears before node v in the sorting. This sorting method ensures that before any task is executed, all its predecessor dependent tasks have been completed, thus avoiding resource conflicts and deadlock problems.

[0138] The system also prioritizes subtasks using heuristic algorithms. This step takes into account a variety of factors, such as the importance of the task, expected execution time, resource requirements, etc. Commonly used heuristic methods include critical path method, earliest deadline first, etc. The purpose of prioritization is to prioritize the subtasks that are more critical to the overall task completion when resources are limited, thereby improving overall efficiency.

[0139] In step S602, the system constructs a distributed task execution framework and applies parallel computing technology. The design of the distributed framework takes into account the characteristics of the task and the available computing resources, and may adopt a master-slave architecture or a peer-to-peer network structure. In the master-slave architecture, the master node is responsible for task allocation and scheduling, and the slave nodes perform specific tasks. The peer-to-peer network structure allows each node to coordinate autonomously, improving the flexibility and fault tolerance of the system.

[0140] The application of parallel computing technology is the core of this step. Based on the task dependency graph and priority information in step S601, the system identifies a set of tasks that can be executed in parallel. These tasks may include content generation for different chapters, independent data processing or formatting operations, etc. The system uses hardware resources such as multi-core processors, GPU acceleration or distributed clusters to execute multiple subtasks simultaneously.

[0141] During the parallel execution process, the system needs to solve several key problems: first, load balancing, ensuring that the workload of each computing node is relatively balanced, avoiding the situation where some nodes are overloaded while other nodes are idle. Second, data synchronization and communication, ensuring that necessary information can be efficiently exchanged between parallel tasks. Third, fault tolerance mechanism, which can detect and handle errors that may occur during the execution process, such as node failure, network interruption, etc.

[0142] The main technical effects of this method include: First, by representing task dependencies through DAG, the system can clearly grasp the structure of complex document generation tasks, laying the foundation for efficient execution. Second, the combination of topological sorting and priority sorting not only ensures the logical correctness of task execution, but also optimizes resource utilization efficiency. Furthermore, the application of distributed framework and parallel computing technology greatly improves the speed of task execution, especially for large-scale and complex document generation tasks. Finally, this method improves the scalability and robustness of the system, and can better cope with document generation needs of different scales and complexities.

[0143] This technology has broad application prospects in many fields. In enterprise-level document management systems, it can be used to quickly generate a large number of customized reports. In the field of academic research, it can assist researchers in organizing and generating literature reviews more efficiently. In the media industry, it can support the rapid generation of news articles and the parallel processing of multilingual content. In e-commerce platforms, it can be used for the automatic generation and update of large-scale product descriptions.

[0144] In one embodiment of the document generation method based on artificial intelligence of the present application, see Figure 7 , and can also include the following:

[0145] Step S701: Convert the optimized draft into a query vector according to the vector retrieval algorithm based on the attention mechanism, calculate the similarity score between the query vector and the vectors in the vector library, select the vector subset with the highest score and use the inverted index structure to quickly locate and extract relevant information.

[0146] Step S702: Constructing a fusion text generation model to encode the retrieved relevant information and the optimized draft into an input sequence to obtain a final document output by the fusion text generation model.

[0147] Optionally, in this embodiment, steps S701 and S702 constitute an advanced document optimization and generation system that cleverly combines information retrieval technology and advanced natural language processing methods. These two steps are intended to optimize the initially generated document into a more comprehensive, accurate and coherent final document through accurate relevant information retrieval and intelligent text fusion. This process not only improves the quality of the generated document, but also greatly enhances the information richness and reliability of the document.

[0148] In step S701, the system first converts the optimized draft into a query vector using a vector retrieval algorithm based on the attention mechanism. This process involves encoding the draft using a pre-trained language model (such as BERT or Transformer) to generate a high-dimensional vector representation. The attention mechanism plays a key role here, which can identify and highlight important elements in the draft, ensuring that the generated query vector can accurately capture the core semantic information of the document.

[0149] Next, the system calculates the similarity score between the query vector and the vectors in the vector library. This step usually uses a metric such as cosine similarity or dot product. The vectors in the vector library may represent various related documents, paragraphs, or pieces of knowledge. By calculating similarity, the system can quickly identify the most relevant information for the first draft.

[0150] The system then selects the subset of vectors with the highest scores. This process may involve setting a threshold or selecting a fixed number of top-k vectors. The selection criteria not only consider the similarity score, but may also consider the diversity of the vectors to ensure the comprehensiveness of the retrieval results.

[0151] Finally, the system uses an inverted index structure to quickly locate and extract relevant information. An inverted index is an efficient data structure that maps each vector to its corresponding original text content. This structure greatly speeds up the retrieval from vectors to actual text content, enabling the system to achieve real-time retrieval on large-scale data sets.

[0152] In step S702, the system builds a fusion text generation model to encode the retrieved relevant information and the optimized draft into an input sequence. This fusion model may be based on a Transformer architecture, such as GPT or BART. During the encoding process, the system may use special delimiters or positional encodings to distinguish the draft content from the retrieved information, so that the model can understand the roles of different parts.

[0153] The core task of the fusion text generation model is to intelligently integrate the retrieved information with the first draft. This process is not a simple splicing, but requires the model to understand the semantic relationship between the two parts of the content and to filter, integrate and rewrite the information. The model needs to consider multiple aspects during the generation process:

[0154] First, it needs to maintain the core structure and theme of the first draft to ensure the coherence and consistency of the final document. Second, it needs to skillfully integrate the retrieved new information into the document, which may involve expanding existing paragraphs, adding new arguments or examples, or correcting incorrect information. Third, the model needs to adjust the text style to ensure that the newly added content is consistent with the original content in language style. Finally, the model also needs to logically sort out the overall content to ensure that the logical relationship between each part is clear and avoid contradictions or repetitions.

[0155] The main technical effects of this method include: First, through the vector retrieval algorithm, the system can quickly and accurately find the most relevant information from a large-scale knowledge base, greatly improving the information richness and accuracy of the document. Second, the use of the inverted index structure significantly improves the retrieval efficiency and enables the system to run in a real-time environment. Furthermore, the application of the fusion text generation model enables the system to intelligently integrate the first draft and new information to generate more comprehensive and coherent documents. Finally, this method improves the quality of automatic document generation and makes the generated content closer to the level of human experts.

[0156] This technology has broad application prospects in multiple fields. In academic research, it can assist researchers in quickly compiling literature reviews and providing the latest and most relevant research progress. In the news media industry, it can help journalists quickly produce comprehensive and accurate news reports, including background information and latest developments. In the business field, it can be used to generate high-quality market analysis reports, integrating multi-source data and expert insights. In the education field, it can be used to create personalized learning materials, dynamically adjusting content according to students' needs and background knowledge.

[0157] In addition, this technology also provides new possibilities for knowledge management and decision support systems. By intelligently retrieving and integrating relevant information, the system can provide decision makers with more comprehensive and in-depth background information and analysis, thereby supporting more informed decision making. This not only improves the efficiency of information processing and knowledge creation, but also provides strong technical support for the effective use and innovation of organizational knowledge, and promotes the application and development of artificial intelligence in knowledge-intensive fields.

[0158] In order to effectively improve the accuracy and intelligence of the document generation process, the present application provides an embodiment of an artificial intelligence-based document generation device for implementing all or part of the content of the artificial intelligence-based document generation method, see Figure 8 , the document generation device based on artificial intelligence specifically includes the following contents:

[0159] A vector library construction module 10 is used to extract multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, and to construct a corresponding knowledge graph based on the multimodal data features, and to convert the multimodal data after feature extraction into a high-dimensional vector using a vectorization algorithm, and store the vector library;

[0160] A first draft determination module 20 is used to generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding contents of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0161] The final draft determination module 30 is used to iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0162] From the above description, it can be seen that the artificial intelligence-based document generation device provided in the embodiment of the present application can extract multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and construct a corresponding knowledge graph based on the multimodal data features, and use a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, and store it to form a vector library; use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a document draft; iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the intelligence of document generation.

[0163] From a hardware perspective, in order to effectively improve the accuracy and intelligence of the document generation process, the present application provides an embodiment of an electronic device for implementing all or part of the content of the artificial intelligence-based document generation method, and the electronic device specifically includes the following content:

[0164] Processor, memory, communication interface and bus; wherein the processor, memory and communication interface communicate with each other through the bus; the communication interface is used to realize information transmission between the document generation device based on artificial intelligence and related devices such as core business systems, user terminals and related databases; the logic controller can be a desktop computer, a tablet computer and a mobile terminal, etc., but the present embodiment is not limited thereto. In the present embodiment, the logic controller can be implemented with reference to the embodiment of the document generation method based on artificial intelligence and the embodiment of the document generation device based on artificial intelligence in the embodiment, and the contents thereof are incorporated herein, and the repeated parts are not repeated.

[0165] It is understandable that the user terminal may include a smart phone, a tablet electronic device, a network set-top box, a portable computer, a desktop computer, a personal digital assistant (PDA), a vehicle-mounted device, a smart wearable device, etc. Among them, the smart wearable device may include smart glasses, a smart watch, a smart bracelet, etc.

[0166] In practical applications, part of the document generation method based on artificial intelligence can be executed on the electronic device side as described above, or all operations can be completed in the client device. The specific selection can be based on the processing capability of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor.

[0167] The client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and other implementation scenarios may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, or a server cluster consisting of multiple servers, or a server structure of a distributed device.

[0168] Fig. 9 FIG. 9 is a schematic block diagram of the system structure of the electronic device 9600 according to an embodiment of the present application. Fig. 9 As shown, the electronic device 9600 may include a central processor 9100 and a memory 9140; the memory 9140 is coupled to the central processor 9100. It is worth noting that Fig. 9 is exemplary; other types of structures may also be used to supplement or replace this structure to implement telecommunication functions or other functions.

[0169] In one embodiment, the function of the document generation method based on artificial intelligence can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0170] Step S101: extracting multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, constructing corresponding knowledge graphs based on the multimodal data features, converting the feature-extracted multimodal data into high-dimensional vectors using a vectorization algorithm, and storing them to form a vector library;

[0171] Step S102: Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0172] Step S103: Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0173] From the above description, it can be seen that the electronic device provided in the embodiment of the present application extracts multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and constructs a corresponding knowledge graph based on the multimodal data features, and uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, and stores it to form a vector library; uses a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a document draft; iteratively optimizes the document content according to the document draft, the prompt words, and the language model to obtain an optimized draft, and uses an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuses the relevant information with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the intelligence of document generation.

[0174] In another embodiment, the artificial intelligence-based document generation device can be configured separately from the central processing unit 9100. For example, the artificial intelligence-based document generation device can be configured as a chip connected to the central processing unit 9100, and the function of the artificial intelligence-based document generation method can be implemented through the control of the central processing unit.

[0175] like Fig. 9 As shown, the electronic device 9600 may also include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily have to include Fig. 9 In addition, the electronic device 9600 may also include Fig. 9 For components not shown, reference may be made to the prior art.

[0176] like Fig. 9As shown, the central processing unit 9100 is sometimes also referred to as a controller or an operation control, and may include a microprocessor or other processor device and / or logic device. The central processing unit 9100 receives input and controls the operation of various components of the electronic device 9600.

[0177] The memory 9140 may be, for example, one or more of a cache, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory or other suitable devices. The above-mentioned information related to the failure may be stored, and a program for executing the relevant information may also be stored. The CPU 9100 may execute the program stored in the memory 9140 to implement information storage or processing, etc.

[0178] The input unit 9120 provides input to the central processing unit 9100. The input unit 9120 is, for example, a key or a touch input device. The power supply 9170 is used to provide power to the electronic device 9600. The display 9160 is used to display display objects such as images and texts. The display may be, for example, an LCD display, but is not limited thereto.

[0179] The memory 9140 may be a solid-state memory, such as a read-only memory (ROM), a random access memory (RAM), a SIM card, etc. It may also be a memory that saves information even when the power is off, can be selectively erased, and is provided with more data, examples of which are sometimes referred to as EPROMs, etc. The memory 9140 may also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142, which is used to store application programs and function programs or processes for executing the operation of the electronic device 9600 through the central processor 9100.

[0180] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for communication functions of the electronic device and / or for executing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0181] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, which may be the same as the case of a conventional mobile communication terminal.

[0182] Based on different communication technologies, multiple communication modules 9110 may be provided in the same electronic device, such as a cellular network module, a Bluetooth module and / or a wireless LAN module, etc. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide an audio output via the speaker 9131 and receive an audio input from the microphone 9132, thereby realizing a common telecommunication function. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. In addition, the audio processor 9130 is also coupled to the central processor 9100, so that recording can be performed on the local machine through the microphone 9132, and the sound stored on the local machine can be played through the speaker 9131.

[0183] The embodiments of the present application also provide a computer-readable storage medium capable of implementing all the steps of the method for generating a document based on artificial intelligence in the above embodiments, wherein the execution subject is a server or a client. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, all the steps of the method for generating a document based on artificial intelligence in the above embodiments are implemented. For example, when the processor executes the computer program, the following steps are implemented:

[0184] Step S101: extracting multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, constructing corresponding knowledge graphs based on the multimodal data features, converting the feature-extracted multimodal data into high-dimensional vectors using a vectorization algorithm, and storing them to form a vector library;

[0185] Step S102: Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0186] Step S103: Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0187] From the above description, it can be seen that the computer-readable storage medium provided in the embodiment of the present application extracts multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and constructs a corresponding knowledge graph based on the multimodal data features, and uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, and stores it to form a vector library; uses a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a document draft; iteratively optimizes the document content according to the document draft, the prompt words, and the language model to obtain an optimized draft, and uses an attention-based retrieval algorithm to obtain relevant information from the vector library, and fuses the relevant information with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the intelligence of document generation.

[0188] The embodiments of the present application also provide a computer program product capable of implementing all the steps of the document generation method based on artificial intelligence in the above embodiments, where the execution subject is a server or a client. When the computer program / instruction is executed by a processor, the steps of the document generation method based on artificial intelligence are implemented. For example, the computer program / instruction implements the following steps:

[0189] Step S101: extracting multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, constructing corresponding knowledge graphs based on the multimodal data features, converting the feature-extracted multimodal data into high-dimensional vectors using a vectorization algorithm, and storing them to form a vector library;

[0190] Step S102: Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document;

[0191] Step S103: Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

[0192] From the above description, it can be seen that the computer program product provided in the embodiment of the present application extracts multimodal data features of historical multimodal files by utilizing natural language processing and computer vision algorithms, and constructs a corresponding knowledge graph based on the multimodal data features, and uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector, which is stored to form a vector library; a task planning algorithm is used to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a document draft; the document content is iteratively optimized according to the document draft, the prompt words, and the language model to obtain an optimized draft, and an attention-based retrieval algorithm is used to obtain relevant information from the vector library, and the relevant information is integrated with the optimized draft through a text generation algorithm to obtain the final document, thereby effectively improving the intelligence of document generation.

[0193] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0194] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0195] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0196] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0197] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A document generation method based on artificial intelligence, characterized in that: The method comprises: The multimodal data features of historical multimodal files are extracted using natural language processing and computer vision algorithms, and a corresponding knowledge graph is constructed based on the multimodal data features. The multimodal data after feature extraction is converted into a high-dimensional vector using a vectorization algorithm, and stored to form a vector library; Generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding content of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document; Iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

2. The document generation method based on artificial intelligence according to claim 1, characterized in that: The method of extracting multimodal data features of historical multimodal files by using natural language processing and computer vision algorithms includes: Performing data format normalization preprocessing on the historical multimodal files, performing word segmentation, part-of-speech tagging and syntactic analysis on the text data in the historical multimodal files through a natural language processing model, extracting word vectors and semantic features, and performing image enhancement, size adjustment and color space conversion preprocessing on the image data in the historical multimodal files through a computer vision algorithm; The deep learning model is used to extract features from the processed text and image data respectively. The context-related semantic features of the text data are extracted through the attention mechanism. The multi-scale visual features of the image data are extracted using a convolutional neural network. The extracted feature vectors are normalized and dimensionally reduced to obtain multimodal data features with unified representation.

3. The document generation method based on artificial intelligence according to claim 1, characterized in that: The method uses a vectorization algorithm to convert the multimodal data after feature extraction into a high-dimensional vector and stores it to form a vector library, including: The principal component analysis method is used to reduce the dimension of the multimodal data after feature extraction to retain the main feature information. The t-distributed random neighbor embedding algorithm is used to nonlinearly map the reduced-dimensional data to a high-dimensional space to generate a vector representation of uniform dimension; Construct an index structure based on hierarchical differentiable hashing, encode the normalized high-dimensional vector into a binary hash code, and use a distributed storage system to store the hash-encoded vector data and metadata index separately, establish associations between data blocks, and form a vector library.

4. The document generation method based on artificial intelligence according to claim 1, characterized in that: The generating of the article structure outline by using a graph neural network according to the knowledge graph and the vector library, and the dynamically generating of prompt words according to a context-aware algorithm, include: The graph attention network is used to capture the relationship and semantic information between knowledge graph nodes. The relevant vectors in the vector library are used as input features of the graph neural network. High-order features are extracted through multi-layer graph convolution operations. The node features are aggregated using the graph pooling method to generate a graph-level representation. The recurrent neural network is combined with the graph-level representation to generate the hierarchical relationship of the article structure outline, where the hierarchical relationship includes chapter titles and subtitles. A contextual semantic tree is constructed based on the generated article structure outline, and a depth-first traversal is performed on the semantic tree. The self-attention mechanism is applied to each node to extract key information. The traversal sequence is processed using a bidirectional long short-term memory network to capture long-distance dependencies. The processed sequence is input into a pre-trained language model, and a beam search algorithm is used to generate multiple candidate prompt words. A sorting algorithm based on context relevance is applied to determine the prompt word that best suits the current context as the output.

5. The document generation method based on artificial intelligence according to claim 1, characterized in that: The determining of the corresponding content of each chapter according to the article structure outline, the prompt words and the preset language model includes: The article structure outline is converted into a sequence representation with title and hierarchical information, the prompt words are encoded using a multi-head attention mechanism, and a conditional text generation model based on the transformer architecture is used to take the outline sequence and prompt word encoding as conditional input; The corresponding contents of each chapter are determined by the chapter-level decoder and paragraph-level decoder of the conditional text generation model.

6. The document generation method based on artificial intelligence according to claim 1, characterized in that: The task planning algorithm is used to split the current document generation task into multiple subtasks and the subtasks are executed simultaneously to obtain a first draft of the document, including: According to the task planning algorithm based on directed acyclic graph, a task dependency graph is constructed to represent the relationship between document generation subtasks, the task execution order is determined according to topological sorting, and the subtasks are prioritized through a heuristic algorithm; Build a distributed task execution framework and apply parallel computing technology to execute multiple subtasks simultaneously.

7. The document generation method based on artificial intelligence according to claim 1, characterized in that: The use of the attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and the fusion of the relevant information with the optimized draft through the text generation algorithm to obtain the final document, including: According to the vector retrieval algorithm based on the attention mechanism, the optimized draft is converted into a query vector, the similarity score between the query vector and the vectors in the vector library is calculated, the vector subset with the highest score is selected, and the inverted index structure is used to quickly locate and extract relevant information. A fusion text generation model is constructed to encode the retrieved relevant information and the optimized draft as an input sequence to obtain a final document output by the fusion text generation model.

8. A document generation device based on artificial intelligence, characterized in that: The device comprises: A vector library construction module is used to extract multimodal data features of historical multimodal files using natural language processing and computer vision algorithms, and to construct a corresponding knowledge graph based on the multimodal data features, and to convert the multimodal data after feature extraction into a high-dimensional vector using a vectorization algorithm, and store them to form a vector library; A first draft determination module is used to generate an article structure outline through a graph neural network according to the knowledge graph and the vector library, dynamically generate prompt words according to a context-aware algorithm, determine the corresponding contents of each chapter according to the article structure outline, the prompt words and a preset language model, and use a task planning algorithm to split the current document generation task into multiple subtasks and execute the subtasks simultaneously to obtain a first draft of the document; The final draft determination module is used to iteratively optimize the document content according to the document draft, the prompt words and the language model to obtain an optimized draft, and use an attention mechanism-based retrieval algorithm to obtain relevant information from the vector library, and fuse the relevant information with the optimized draft through a text generation algorithm to obtain a final document.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the document generation method based on artificial intelligence described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the document generation method based on artificial intelligence described in any one of claims 1 to 7 are implemented.

Images