Project text intelligent generation method and device

Abstract

The embodiment of the invention provides an intelligent project text generation method and device, and the method achieves the precise screening of positive and negative training corpora through the construction of a multi-dimensional scoring mechanism, and the comprehensive evaluation of the text coverage degree, the technical correlation degree, the fluency and the professional term specification degree. And constructing a complete knowledge graph based on the relation extraction model and the external knowledge base, and fusing the enterprise technical feature matrix and the basic information to generate an enterprise portrait vector. By transforming a decoding layer structure of a pre-training large model, enterprise portrait vectors are used as condition vectors to be embedded, and personalized text generation is achieved. According to the method, the defects of the traditional technology in the aspects of training sample quality evaluation, knowledge graph integrity and enterprise feature fusion are effectively overcome, and the professionality and pertinence of project text generation are remarkably improved.

Description

Technical Field The present application relates to the field of data processing, and specifically relates to a method and device for intelligent generation of project texts. Background Art Existing methods for generating project texts have obvious deficiencies. Traditional methods lack systematic evaluation criteria when screening training corpora, fail to effectively distinguish positive and negative samples, and have insufficient multi-dimensional evaluation of text quality, which affects the training effect of the model.

[0001] In addition, there are bottlenecks in knowledge graph construction and knowledge supplementation in the existing technology. Most systems fail to make full use of external knowledge base resources and it is difficult to construct a complete technical dependency network. The generation of text templates lacks in-depth analysis of technological innovation points, which affects the professionalism of the generated content.

[0002] Existing systems have technical shortcomings in model customization and integration of enterprise characteristics. The utilization of enterprise technology R & D data is insufficient, and the model is difficult to adapt to the specific needs of different enterprises. Solving these problems is of great significance for improving the quality and pertinence of project text generation. Summary of the Invention In view of the problems in the existing technology, the present application provides a method and device for intelligent generation of project texts, which can solve the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and integration of enterprise characteristics, and significantly improve the professionalism and pertinence of project text generation.

[0003] To solve at least one of the above problems, the present application provides the following technical solutions: In a first aspect, the present application provides a method for intelligent generation of project texts, including: Performing text similarity clustering and deduplication on historical declaration documents, using a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculating the text coverage score of the historical declaration documents and the project declaration requirements, the technical correlation score with the key points of technology R & D, the text fluency score, and the professional term standardization score, performing weighted fusion on each score to obtain a comprehensive score, screening high-score samples based on the comprehensive score to construct a positive training corpus, and constructing a negative training corpus for samples with a comprehensive score lower than a preset threshold; Constructing the text content in the positive training corpus as knowledge graph nodes, using a relation extraction model to extract technical dependencies and innovation point associations from the text content to construct knowledge association edges, obtaining supplementary knowledge from external standard libraries and patent libraries and adding it to the knowledge graph, using a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjusting the topological structure of the knowledge graph according to the weight coefficients, and inputting the knowledge graph into a text structure analysis model to generate a standardized text template, and constructing a differential text template library; Obtain technical R & D data from the enterprise R & D management system and perform standardization processing. Feature encode the technical R & D data to obtain an enterprise technical feature matrix. Based on an autoencoder, perform dimensionality reduction processing on the enterprise technical feature matrix, and splice the dimensionality reduction result with the enterprise basic information to obtain an enterprise portrait vector; transform the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, and optimize the parameters of the pre-trained large model using the cross-entropy loss function to construct an enterprise project text generation model.

[0004] Further, it also includes: calculating the content similarity in historical declaration documents using a text clustering algorithm based on semantic distance, constructing a text similarity matrix, marking text pairs with a similarity higher than the first preset threshold as duplicate texts, performing deduplication processing on the duplicate texts based on the locality-sensitive hashing algorithm, using a word vector model to perform word sequence segmentation on the deduplicated historical declaration documents, and representing the word sequence as a multi-dimensional vector in combination with position encoding, and performing sentence-level embedding representation on the multi-dimensional vector based on the attention mechanism; Construct a multi-branch scoring network model. In the first branch, calculate the text coverage score based on the term frequency-inverse document frequency algorithm. In the second branch, calculate the technical relevance score by extracting technical elements based on named entity recognition. In the third branch, calculate the sentence fluency score based on the language model. In the fourth branch, calculate the term specification score based on a preset professional term dictionary. Input the scores of each branch into a fully connected layer for weighted fusion to obtain a comprehensive score.

[0005] Further, it also includes: using an adaptive weight allocation mechanism to calculate the weights of the text coverage score, technical relevance score, fluency score, and term specification score, applying the Softmax function to normalize the weights, performing a linear combination of each score based on the normalized weights to obtain a text comprehensive score, and performing Z-score standardization processing on the text comprehensive score to obtain a standardized score; Extract text content, project classification labels, and declaration result marks from historical declaration documents. Form positive training sample pairs with text content, project classification labels, and declaration result marks whose standardized scores are higher than the second preset threshold to construct a positive training corpus, and form negative training sample pairs with text content, project classification labels, and declaration result marks whose standardized scores are lower than the third preset threshold to construct a negative training corpus.

[0006] Further, it also includes: performing syntactic analysis and dependency parsing on the text content in the positive training corpus, identifying technical feature words, technical field words, and innovation point description words from the parsing results, constructing entity nodes of the knowledge graph with the technical feature words, technical field words, and innovation point description words, using a string matching algorithm to de-duplicate and merge the entity nodes, aggregating similar entity nodes based on text semantic similarity calculation, and assigning a unique identifier to each entity node; Scanning the text content using a relationship extraction model based on the attention mechanism, identifying the technical dependency relationships and innovation point associations between entity nodes, constructing the association edges of the knowledge graph with the technical dependency relationships and innovation point associations, retrieving supplementary knowledge related to the entity nodes from external standard libraries and patent libraries, and constructing new entity nodes and association edges with the supplementary knowledge and adding them to the knowledge graph.

[0007] Further, it also includes: inputting the entity node features and association edge features in the knowledge graph into a graph convolutional network, aggregating and updating the node features through multi-layer graph convolutional operations, calculating the association strength between entity nodes based on the attention mechanism, using the association strength as the weight coefficient of the knowledge association edge, screening the knowledge association edges based on the weight coefficient, removing the association edges with a weight coefficient lower than the fourth preset threshold, and optimizing the topological structure of the knowledge graph; Performing subgraph partitioning on the optimized knowledge graph, extracting the critical path based on the depth-first search algorithm, mapping the critical path to text structure features, performing sequence modeling on the text structure features using a recurrent neural network, generating a standardized text template containing technical features, innovation points, and association relationships, and clustering the standardized text template based on text similarity calculation to construct a differentiated text template library.

[0008] Further, it also includes: retrieving technical R & D data from the enterprise R & D management system, normalizing the fields of the technical R & D data, encoding discrete features using one-hot encoding, standardizing continuous features using min-max normalization, performing feature selection on the technical R & D data based on the principal component analysis method, constructing an enterprise technical feature matrix with the selected features, and filling the missing values in the enterprise technical feature matrix using a multi-layer perceptron; Inputting the enterprise technical feature matrix into a stacked autoencoder, compressing the high-dimensional features into low-dimensional feature representations through the encoder, reconstructing the compressed features through the decoder, optimizing the autoencoder parameters based on the reconstruction error, extracting the low-dimensional features output by the encoder, encoding the enterprise basic information, and combining the low-dimensional features with the enterprise basic information features using the feature splicing method to obtain an enterprise portrait vector.

[0009] Further, it also includes: adding a conditional vector input module in the decoding layer of the pre-trained large model, projecting the enterprise portrait vector into the feature space of the decoding layer by using the multi-head attention mechanism, fusing the projected features with the hidden state of the decoding layer, constructing a template matching matrix based on the text structure features in the differential text template library, performing a tensor multiplication operation on the template matching matrix and the output of the decoding layer to obtain a decoding constraint matrix, and applying the Softmax function to the decoding constraint matrix to generate a word sequence probability distribution; Decoding the word sequence probability distribution by using the greedy decoding strategy, calculating the cross-entropy loss between the decoding result and the standard text in the forward training corpus, introducing a regularization term to constrain the model parameters, using the Adam optimizer to perform gradient update on the parameters of the pre-trained large model, and early stopping the model based on the perplexity metric on the validation set to obtain an enterprise project text generation model.

[0010] In a second aspect, the present application provides an intelligent project text generation device, including: A training library construction module, configured to perform text similarity clustering and deduplication on historical declaration documents, perform embedding representation on the deduplicated historical declaration documents by using a text vectorization model, calculate the text coverage score between the historical declaration documents and the project declaration requirements, the technical relevance score with the key points of technological research and development, the text fluency score, and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, screen high-score samples based on the comprehensive score to construct a forward training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; A knowledge graph construction module, configured to construct the text content in the forward training corpus into knowledge graph nodes, extract technical dependency relationships and innovation point associations from the text content by using a relationship extraction model to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, calculate the weight coefficients of the knowledge association edges by using a graph convolutional network, adjust the topological structure of the knowledge graph according to the weight coefficients, input the knowledge graph into a text structure analysis model to generate a standardized text template, and construct a differential text template library; A project text generation module, configured to obtain technical research and development data from an enterprise research and development management system and perform standardization processing, perform feature encoding on the technical research and development data to obtain an enterprise technical feature matrix, perform dimensionality reduction processing on the enterprise technical feature matrix based on an autoencoder, and perform feature splicing on the dimensionality reduction result and the enterprise basic information to obtain an enterprise portrait vector; transform the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, optimize the parameters of the pre-trained large model by using a cross-entropy loss function, and construct an enterprise project text generation model.

[0011] In a third aspect, the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the intelligent project text generation method are implemented.

[0012] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the intelligent project text generation method are implemented.

[0013] In a fifth aspect, the present application provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the intelligent project text generation method are implemented.

[0014] As can be seen from the above technical solutions, the present application provides an intelligent project text generation method and device. By constructing a multi-dimensional scoring mechanism and comprehensively evaluating text coverage, technical relevance, fluency, and professional term standardization, accurate screening of positive and negative training corpora is achieved. Based on a relation extraction model and an external knowledge base, a complete knowledge graph is constructed, and an enterprise portrait vector is generated by integrating the enterprise technical feature matrix and basic information. By modifying the structure of the decoding layer of the pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improves the professionalism and pertinence of project text generation. BRIEF DESCRIPTION OF THE DRAWINGS In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0015] Figure 1 It is a schematic flowchart of the intelligent project text generation method in the embodiment of the present application; Figure 2 It is a structural diagram of the intelligent project text generation device in the embodiment of the present application; Figure 3 It is a schematic structural diagram of the electronic device in the embodiment of the present application.

[0016] Reference numerals: 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 implementation manners To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0017] In the technical solutions of the present application, the acquisition, storage, use, processing, etc. of data all comply with the relevant provisions of national laws and regulations.

[0018] Considering the problems existing in the prior art, the present application provides a method and device for intelligent generation of project texts. By constructing a multi-dimensional scoring mechanism, through comprehensive evaluation of text coverage, technical relevance, fluency, and standardization of professional terms, accurate screening of positive and negative training corpora is achieved. Based on a relation extraction model and an external knowledge base, a complete knowledge graph is constructed, and an enterprise portrait vector is generated by fusing the enterprise technical feature matrix and basic information. By modifying the decoding layer structure of a pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improves the professionalism and pertinence of project text generation.

[0019] In order to solve the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improve the professionalism and pertinence of project text generation, the present application provides an embodiment of a method for intelligent generation of project texts. Refer to Figure 1 , and the method for intelligent generation of project texts specifically includes the following content: Step S101: Perform text similarity clustering and deduplication on historical declaration documents, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical relevance score with the key points of technological research and development, the text fluency score, and the standardization score of professional terms, perform weighted fusion on each score to obtain a comprehensive score, and based on the comprehensive score, screen high-score samples to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; Optionally, in this embodiment, an innovative similarity clustering and duplicate removal mechanism is first implemented for the historical declaration document library. Based on the semantic features of the text, an improved K-means clustering algorithm is used to preliminarily group the documents. When calculating the similarity between documents, the TF-IDF weighted cosine similarity method is adopted, with a focus on technical feature words and descriptions of key innovation points. For declaration documents in different technical fields, a domain dictionary is introduced for feature enhancement to improve the accuracy of clustering. On this basis, a multi-level text vectorization processing flow is designed. First, the documents are preprocessed, including punctuation normalization, professional term recognition, and paragraph structure division. A pre-trained BERT model is used as the basic encoder and fine-tuned to adapt to the specific expression of project declaration documents. During the text encoding process, the attention mechanism is used to highlight the importance of technical points and innovation descriptions, generating a 768-dimensional document vector representation.

[0020] This embodiment implements a comprehensive multi-dimensional evaluation mechanism. In terms of text coverage evaluation, a keyword dictionary for project declaration requirements is constructed, including core elements such as necessary technical indicators and assessment key points. The matching degree between the document content and the declaration requirements is calculated through a semantic matching algorithm, considering expressions with similar meanings. At the same time, the structural integrity of the document is analyzed to ensure the comprehensive coverage of important chapters and key content. For technical relevance evaluation, named entity recognition technology is used to identify elements such as technical terms, innovation methods, and performance indicators in the document. The knowledge graph is used to assist in analyzing the association relationships between technical elements to evaluate the rationality and innovation of the technical route. The model particularly focuses on the depth of technical description and the matching degree of application scenarios. In terms of text fluency evaluation, the smoothness of sentences is calculated based on a pre-trained language model, and the normality of sentence structure and semantic coherence are analyzed. The naturalness of the transition between paragraphs is evaluated in a sliding window manner to ensure the professionalism and readability of the text expression.

[0021] This embodiment innovatively designs a professional term standardization evaluation system and a score fusion mechanism. A professional term standard library covering multiple technical fields is established, including a comparison table of standard terms and common error terms. Irregular terms are identified through pattern matching, and term correction suggestions are provided in combination with the context semantics. The evaluation process particularly emphasizes the consistency and accuracy of term usage. In terms of score fusion, the AHP (Analytic Hierarchy Process) is used to determine the weights of each scoring dimension. The weight allocation takes into account the characteristics and review focuses of different types of projects. The weights are normalized through the Softmax function to ensure the stability of the scoring system. A non-linear transformation is introduced during the fusion process to better reflect the mutual influence between scoring indicators.

[0022] In this embodiment, a complete training corpus construction process is designed for sample screening. First, by analyzing historical high-quality project documents, a scoring distribution model is established to achieve adaptive threshold determination. The setting of the threshold comprehensively considers factors such as project type, technical field, and innovation degree, and adopts different screening criteria for different types of projects. On this basis, a bidirectional training corpus is innovatively constructed. Samples above the threshold are used as positive corpus to construct standard examples reflecting best practices. For samples below the threshold, common problems and improvement directions are extracted through error analysis to form a negative corpus. The construction of the bidirectional corpus provides more comprehensive learning materials for subsequent model training.

[0023] Through the above technical innovations in this embodiment, several key problems in the screening and evaluation of project application documents are effectively solved: uneven sample quality, inconsistent evaluation criteria, insufficient feature extraction, etc. In practical applications, this solution can accurately identify high-quality application documents and establish a reliable training corpus. It is particularly suitable for processing application documents with strong professionalism such as technological innovation projects and industrialization projects, significantly improving the efficiency and accuracy of sample screening. The adaptive characteristics of this solution enable it to handle different types of project application requirements and have broad application value. Through a systematic evaluation system and corpus construction mechanism, it provides a high-quality data foundation for subsequent model training and effectively supports the intelligent generation of project application documents.

[0024] Step S102: Construct the text content in the positive training corpus into knowledge graph nodes, use a relation extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, and input the knowledge graph into a text structure analysis model to generate a standardized text template and construct a differential text template library; Optionally, this embodiment implements an innovative knowledge graph construction mechanism based on the positive training corpus. First, a pre-trained language model is used to perform in-depth semantic parsing on the text content to identify key entities including technical features, innovation elements, performance indicators, etc. The logical relationships between entities are extracted through dependency syntactic analysis to establish an initial knowledge node set. For the identified entity nodes, entity alignment is performed using string matching and semantic similarity calculation to ensure the uniqueness and standardization of the nodes. During the entity recognition process, special attention is paid to the core innovation points and key technical indicators in the technical solution, and a knowledge system that conforms to the characteristics of project applications is constructed through a multi-level entity classification system.

[0025] In this embodiment, a dedicated relation extraction model is designed to mine the complex associations between technical entities. The model adopts a bidirectional encoder structure based on BERT, integrating the attention mechanism and graph neural network, which can accurately capture technical dependency relationships and innovation point associations. During the relation extraction process, causal relationships, support relationships, and innovation breakthrough points in the technical route are key identified, while considering the feasibility and innovation of technical solutions. To enhance the integrity of the knowledge graph, this embodiment innovatively designs a knowledge supplementation mechanism. By docking with external standard libraries and patent libraries, standard specifications and patent information related to the project's technical field are retrieved. A semantic matching algorithm is used to screen highly relevant supplementary knowledge, which is then transformed into new knowledge nodes and association edges, enriching the semantic expression of the knowledge graph.

[0026] This embodiment implements a knowledge graph optimization scheme based on graph convolutional networks. The network adopts a multi-layer graph convolutional structure, and the node features are iteratively updated through a message passing mechanism. In each layer of convolution, the node features are updated by aggregating neighbor information, and the weight coefficients reflect the importance of knowledge associations. In particular, evaluation indicators of technical innovation degree and application value are introduced during the weight calculation process, enabling the graph structure to better reflect the technical characteristics of the project. Based on the calculated weight coefficients, the knowledge graph is dynamically pruned to retain high-value knowledge associations and optimize the topological structure of the graph.

[0027] For text template generation, this embodiment designs an innovative structure analysis model. The model is based on the optimized knowledge graph and uses a graph attention network to extract key paths and core subgraphs. Through deep learning methods, the graph structure features are mapped into a text organization framework to generate a standardized template containing key elements such as technical routes, innovation point demonstrations, and effect verifications. During the template generation process, the special requirements of different types of projects are fully considered, and the pertinence of the template is ensured through conditional constraints. Further, clustering analysis is performed on the generated templates, and a differentiated template library is constructed based on dimensions such as project type, technical field, and innovation degree, providing structured guidance for subsequent text generation.

[0028] Through the above technical innovations, this embodiment effectively solves multiple key problems in knowledge extraction and structure optimization of project declaration documents, such as incomplete knowledge representation, inaccurate relation extraction, and poor template adaptability. In practical applications, this solution can accurately construct a knowledge graph reflecting the technical characteristics of the project and generate high-quality document templates. It is particularly suitable for processing declaration documents with strong professionalism such as technical innovation projects and industrialization projects, significantly improving the systematicness of knowledge organization and the standardization of document generation. The adaptive characteristics of this solution enable it to handle different types of project declaration requirements and have broad application value. Through a systematic knowledge representation and template generation mechanism, it provides reliable knowledge support and structural guidance for subsequent text generation.

[0029] Step S103: Obtain technology R & D data from the enterprise R & D management system and perform standardization processing. Perform feature encoding on the technology R & D data to obtain an enterprise technology feature matrix. Based on an autoencoder, perform dimensionality reduction processing on the enterprise technology feature matrix, and splice the dimensionality reduction result with enterprise basic information to obtain an enterprise portrait vector. Reconstruct the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, and optimize the parameters of the pre-trained large model using the cross-entropy loss function to construct an enterprise project text generation model.

[0030] Optionally, in this embodiment, a comprehensive technology R & D data processing mechanism is first designed. Key technical indicators, R & D milestones, patent achievements, technical talent reserves, R & D investment, and other multi-dimensional data are obtained from the enterprise R & D management system through an API interface. To ensure data quality, a multi-level data preprocessing process is designed: first, standardize the data format, unify the timestamp format, measurement unit, and technical indicator expression; then design outlier detection rules based on domain knowledge, and use the box plot method combined with expert experience thresholds to identify and process abnormal data; for missing values, use mean filling, nearest neighbor interpolation, or feature completion based on similar enterprises according to the data type. In the feature encoding stage, one-hot encoding is used for discrete features such as technical fields, patent types, and talent levels, and min-max normalization is performed on continuous features such as R & D investment, technical indicators, and achievement conversion rates. Through this systematic feature engineering, a multi-dimensional feature matrix reflecting the enterprise's technological innovation ability and R & D strength is constructed.

[0031] This embodiment implements an innovative feature dimensionality reduction and fusion solution. An improved stacked autoencoder is used to compress and represent the enterprise technology feature matrix. Both the encoder and decoder adopt a four-layer neural network structure, and the dimensions of the intermediate hidden layers are 512, 256, 128, and 64 in sequence. To retain the key information of technical features, both the reconstruction loss and the feature correlation constraint are optimized during the training process. The reconstruction loss ensures that the compressed features can restore the original information, and the correlation constraint ensures that the representations of similar technical features in the hidden space are close. The dimensionality-reduced features are spliced with basic information such as enterprise scale, industry attributes, qualification levels, and development stages to construct a complete enterprise portrait vector. During the feature splicing process, a multi-head attention mechanism is innovatively introduced to dynamically adjust the importance weights of each dimension feature according to the characteristics of different types of projects, improving the expression ability and pertinence of the enterprise portrait.

[0032] In this embodiment, for the transformation of the pre-trained model, a deep conditional text generation mechanism is designed. A conditional control module is added to the original Transformer decoding layer, and the enterprise portrait vector is converted into a conditional vector through linear mapping and LayerNorm. At each decoding time step, the conditional vector is dynamically fused with the self-attention features through a gating mechanism to achieve precise control of the generation process. At the same time, a multi-level decoding constraint matrix is constructed based on the differential text template library, including global structure constraints, local logic constraints, and professional term constraints, to ensure that the generated text conforms to the project declaration specifications. The model training uses a cross-entropy loss function with label smoothing, and an adversarial training strategy is introduced to improve the robustness of the model. The training process is optimized through techniques such as gradient accumulation and dynamic batch size to continuously improve the generation quality of the model.

[0033] Through the above technical innovations, this embodiment effectively solves several key problems in enterprise project text generation: incomplete enterprise feature representation, serious loss of feature dimension reduction information, insufficient pertinence of generated content, non-standard text structure, etc. In practical applications, this solution can generate high-quality documents that meet the project declaration requirements according to the technical characteristics and development status of the enterprise. It is particularly suitable for the generation of declaration materials with strong professionalism such as technological innovation projects and industrialization projects, significantly improving the efficiency and accuracy of document generation. The adaptive characteristics of this solution enable it to meet the project declaration needs of different types of enterprises and have broad application value. Through a systematic feature processing and model optimization mechanism, it provides intelligent support for enterprise project declarations.

[0034] As can be seen from the above description, the project text intelligent generation method provided by the embodiment of the present application can achieve precise screening of positive and negative training corpora through constructing a multi-dimensional scoring mechanism and comprehensive evaluation of text coverage, technical relevance, fluency, and professional term standardization. A complete knowledge graph is constructed based on the relation extraction model and external knowledge base, and the enterprise portrait vector is generated by fusing the enterprise technical feature matrix and basic information. By transforming the structure of the pre-trained large model decoding layer and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improves the professionalism and pertinence of project text generation.

[0035] In an embodiment of the project text intelligent generation method of the present application, the following content may also be specifically included: Step S201: Calculate the content similarity in historical declaration documents using a text clustering algorithm based on semantic distance, construct a text similarity matrix, mark text pairs with similarity higher than the first preset threshold as duplicate texts, perform duplicate removal on the duplicate texts using the locality-sensitive hashing algorithm, segment the word sequences of the historical declaration documents after duplicate removal using a word vector model, represent the word sequences as multi-dimensional vectors by combining positional encoding, and perform sentence-level embedding representation on the multi-dimensional vectors based on the attention mechanism; Step S202: Construct a multi-branch scoring network model. Calculate the text coverage score based on the term frequency-inverse document frequency algorithm in the first branch, calculate the technical relevance score by extracting technical elements based on named entity recognition in the second branch, calculate the sentence fluency score based on a language model in the third branch, calculate the term normality score based on a preset professional term dictionary in the fourth branch, and input the scores of each branch into a fully connected layer for weighted fusion to obtain a comprehensive score.

[0036] Optionally, in this embodiment, an efficient text duplicate removal mechanism is first designed. Use a BERT-based semantic encoder to extract text features from historical declaration documents, and process the semantic representations at the sentence, paragraph, and document levels through a hierarchical encoding strategy. At the sentence level, use a multi-head self-attention mechanism to capture the dependencies between words; at the paragraph level, model the context information through a recurrent neural network; at the document level, apply a hierarchical pooling operation to obtain the global semantic representation. Calculate the cosine similarity between documents based on the extracted features, and construct an N×N similarity matrix (N is the number of documents). Considering the characteristics of technical documents, a weighted strategy is adopted when calculating the similarity, and higher weights are assigned to core paragraphs such as technical solutions, innovation points, and key indicators. To improve the calculation efficiency, document clustering preprocessing is innovatively introduced, and the documents are grouped according to technical fields and topic similarities before calculating the similarity. When the weighted similarity of a document pair exceeds a dynamically adjusted preset threshold, it is marked as a duplicate text pair.

[0037] In the duplicate removal stage of this embodiment, an improved locality-sensitive hashing algorithm is innovatively applied. By designing multiple independent hash function families, each function family is responsible for capturing text features from different angles. In the specific implementation, first project the text feature vectors onto multiple random hyperplanes, and generate binary signatures based on the projection results. To improve the accuracy, a multi-round hashing strategy is adopted, and different combinations of hash functions are used in each round. Quickly identify potential similar document pairs through the Hamming distance of the signatures, greatly reducing the computational complexity. At the same time, a Bloom filter is introduced to optimize the storage structure and reduce hash collisions. For the identified candidate document pairs, perform a secondary exact match to ensure the accuracy of duplicate removal.

[0038] This embodiment realizes a professional text representation method. First, a domain-adaptive word segmentation system is constructed, which includes a technical term library, a named entity library, and a context rule library. The technical term library is continuously expanded by crawling industry standard documents and patent literatures; the named entity library covers professional entities such as technical indicators, method names, and device models; the context rule library contains disambiguation and new word discovery rules. The word segmentation process uses a bidirectional maximum matching algorithm combined with a conditional random field model to effectively handle the boundary recognition problem of professional terms. For the word segmentation results, a pre-trained domain word vector model is used for feature extraction. This model is trained on a large-scale technical document corpus and better expresses the semantic features of professional vocabulary. An innovative position encoding scheme considering the structure of technical documents is designed, which not only encodes the position of words in a sentence but also includes their position information in the document hierarchy.

[0039] This embodiment designs a multi-branch scoring network for document quality evaluation. The first branch uses an improved TF-IDF algorithm to calculate text coverage. An innovative approach is introduced to extract the document topic distribution using a topic model, and the coverage of key content is calculated by combining word frequency statistics and topic relevance. The second branch is based on an improved BiLSTM-CRF model to identify technical elements. The model adopts a multi-task learning framework to simultaneously perform entity recognition and relationship extraction, accurately capturing the integrity and innovation of the technical route. The third branch uses a Transformer-based language model to evaluate text fluency. This model is specifically fine-tuned on technical documents and is more suitable for evaluating the normativity of professional writing. The fourth branch constructs a term specification evaluation module, which includes a term usage specification library, a common error pattern library, and context relevance checking, comprehensively evaluating the accuracy and consistency of term usage. The scoring results of each branch are dynamically weighted through an attention mechanism, and the weight allocation is adaptively adjusted according to the characteristics of different types of projects.

[0040] Through the above technical innovations, this embodiment effectively solves several key problems in the screening of project application documents: inaccurate text duplication recognition, insufficient feature representation, and incomplete quality evaluation, etc. In practical applications, it significantly improves the efficiency and accuracy of document processing, providing high-quality training samples for subsequent intelligent generation. This solution is particularly suitable for processing application documents with strong professionalism such as technological innovation projects and industrialization projects, and has broad application value.

[0041] In an embodiment of the intelligent generation method for project texts of this application, it may also specifically include the following content: Step S301: Calculate the weights of the text coverage score, technical relevance score, fluency score, and term standardization score using an adaptive weight allocation mechanism, apply the Softmax function to normalize the weights, perform a linear combination of each score based on the normalized weights to obtain a text comprehensive score, and perform Z-score standardization processing on the text comprehensive score to obtain a standardized score; Step S302: Extract the text content, project classification labels, and application result markers from historical application documents. The text content, project classification labels, and application result markers with a standardized score higher than the second preset threshold form positive training sample pairs to construct a positive training corpus, and the text content, project classification labels, and application result markers with a standardized score lower than the third preset threshold form negative training sample pairs to construct a negative training corpus.

[0042] Optionally, in this embodiment, an innovative scoring fusion mechanism is first designed. According to the quality assessment characteristics of project application documents, an adaptive weight allocation model is constructed. This model dynamically calculates the importance weights of each scoring dimension by analyzing the feature distribution of historical high-quality projects. In specific implementation, an improved principal component analysis method is used to extract the feature contribution degree, and the weight calculation formula of the model is: W = softmax(V·λ), where V is the feature vector matrix, λ is the eigenvalue vector, and the Softmax function ensures weight normalization. During the weight allocation process, the model will be dynamically adjusted according to the characteristics of the project type and technical field. For example, for technology innovation projects, it focuses on technical relevance and innovation, and for achievement transformation projects, it pays more attention to technical feasibility and implementation effects. At the same time, to improve the adaptability of the model, a feedback adjustment mechanism is introduced to continuously optimize the weight allocation strategy based on the project review results. In particular, when dealing with new project categories, the model can quickly adjust the weight distribution based on the historical data of similar projects to ensure the rationality of the scoring.

[0043] This embodiment realizes rigorous scoring standardization processing. Based on the calculated weights, a linear combination is performed on the scores of four dimensions: text coverage, technical relevance, fluency, and terminology standardization. To improve the accuracy of scoring, an interaction term between scoring dimensions is innovatively introduced, considering the synergistic effect of each scoring index. For example, a situation where the technical relevance is high but the terminology standardization is low may indicate problems with the improper use of professional terms in the document, and the weight needs to be appropriately reduced in the comprehensive scoring. The Z-score standardization is applied to the combined comprehensive score, and the mean and standard deviation of the local data distribution are calculated using a sliding window mechanism, avoiding the bias that may be brought by global standardization. The window size is dynamically adjusted according to the characteristics of the data stream, ensuring both the timeliness of scoring and the stability of the scoring standard. During the standardization process, the stability of the scoring is ensured through outlier detection based on quartiles and exponential smoothing processing, while retaining the relative differences between samples.

[0044] This embodiment designs a complete corpus construction process. Multidimensional information is extracted from historical application documents, including text content, project classification labels, and application result markers. The data extraction uses improved structured parsing technology, and a deep learning model is used to accurately identify the logical hierarchy and key information of the documents. The model adopts the BiLSTM-CRF architecture and is trained specifically for the characteristics of project application documents, and can effectively identify key paragraphs such as technical solutions, innovation point descriptions, and expected effects. For samples with a standardized score higher than the second preset threshold, a multi-level screening strategy is adopted to ensure the quality of positive samples. First, a preliminary screening is performed based on the scoring level, and then the distribution of samples in the technical field and project type is analyzed. The sample ratio is adjusted through undersampling or oversampling methods to ensure the balance of the corpus. For samples with strong technical innovation, the scoring threshold is appropriately reduced to retain their innovative features. At the same time, a systematic error pattern analysis is performed on samples below the third preset threshold, and typical problem cases are extracted to construct a negative training corpus. The selection of negative samples particularly focuses on covering common error types, such as unclear technical routes, insufficient demonstration of innovation points, and non-standard use of terms.

[0045] Through the above technical innovations, this embodiment effectively solves multiple key problems in the evaluation of project application documents, such as unreasonable scoring weight distribution, inconsistent scoring criteria, and unstable quality of training samples. In practical applications, this solution can accurately evaluate the quality of documents and screen out high-quality training samples. It is particularly suitable for processing application documents with strong professionalism, such as technological innovation projects and industrialization projects, significantly improving the accuracy of evaluation and the quality of the corpus. The adaptive characteristics of this solution enable it to handle different types of project application requirements, providing a reliable data basis for subsequent model training through a systematic evaluation system and corpus construction mechanism. Through continuous model optimization and data accumulation, the evaluation accuracy and adaptability of the system are continuously improved, effectively supporting the intelligent generation of project application documents.

[0046] In an embodiment of the intelligent generation method for project texts in this application, the following content may also be specifically included: Step S401: Perform syntactic analysis and dependency parsing on the text content in the positive training corpus, identify technical feature words, technical field words, and innovation point description words from the parsing results, construct the technical feature words, technical field words, and innovation point description words as entity nodes of the knowledge graph, use a string matching algorithm to de-duplicate and merge the entity nodes, aggregate similar entity nodes based on text semantic similarity calculation, and assign a unique identifier to each entity node; Step S402: Use a relationship extraction model based on the attention mechanism to scan the text content, identify the technical dependency relationships and innovation point associations between entity nodes, construct the technical dependency relationships and innovation point associations as association edges of the knowledge graph, retrieve supplementary knowledge related to entity nodes from external standard libraries and patent libraries, and construct the supplementary knowledge as new entity nodes and association edges and add them to the knowledge graph.

[0047] Optionally, this embodiment first designs a deep text parsing mechanism. For the text content in the positive training corpus, a pre-trained model based on BERT is used for syntactic analysis, and this model is specifically optimized on the technical document corpus. An improved attention mechanism is used during the syntactic analysis process, focusing on the syntactic structure of the technical description paragraphs. In the dependency parsing stage, a bidirectional LSTM-CRF model is used to extract the dependency relationships between words and construct a complete syntactic tree structure. The parsing process particularly focuses on key contents such as technical feature descriptions, method steps, and innovation point demonstrations, and accurately identifies technical terms and key descriptions through multi-level feature extraction. To improve the parsing accuracy, the model integrates the constraint rules of the domain knowledge base and effectively processes professional terms and complex sentence patterns.

[0048] This embodiment realizes an innovative entity recognition and node construction method. From the syntactic parsing results, a conditional random field model is used to identify three types of keywords: technical feature words (such as specific technical parameters and method steps), technical field words (such as application scenarios and technical classifications), and innovation point description words (such as technical advantages and improvement solutions). The entity recognition process introduces a multi-task learning framework to simultaneously perform entity type discrimination and boundary recognition, improving the recognition accuracy of complex terms. The identified entity words are initially sorted, and an improved edit distance algorithm is used for string matching to merge entities with the same meaning. Further, by calculating the semantic similarity between entities, a hierarchical clustering method is used to aggregate similar entities, reducing the redundancy of the knowledge graph. A dynamic threshold is set during the aggregation process, and the aggregation criteria are adaptively adjusted according to the technical field characteristics of the entities. Finally, a globally unique identifier is assigned to each entity node, and an entity index system is established.

[0049] This embodiment designs an efficient relation extraction model. An entity relation recognizer is constructed based on the multi-head attention mechanism, and the model can simultaneously capture local semantic features and global context information. The relation extraction process is divided into two stages: first, directly associated entities are identified, including technical dependency relationships (such as component composition and method step order) and innovation point associations (such as problem solving and effect improvement); then potential associations are discovered through cross-document relation reasoning. The model adopts a gated update mechanism to dynamically adjust the edge weights of the knowledge graph according to the confidence of the relationships. To enrich the semantic expression of the knowledge graph, a knowledge supplementation mechanism is innovatively designed. By docking with external standard libraries and patent libraries, supplementary knowledge related to existing entity nodes is retrieved. The retrieval process combines semantic retrieval and keyword matching to ensure the relevance of the supplementary knowledge. The obtained external knowledge is structured, new entity nodes and associated edges are extracted, and consistency checks are performed to ensure compatibility with the original knowledge.

[0050] This embodiment further optimizes the quality of the knowledge graph. For the newly added nodes and edges, a multi-level quality control mechanism is designed. First, obviously unreasonable associations are removed through rule filtering; then the importance of nodes and the reliability of edges are analyzed based on graph mining algorithms; finally, the accuracy of the knowledge is ensured through the verification rules of the expert knowledge base. To improve the practicality of the graph, a dynamic update mechanism is established, which can continuously optimize and expand the knowledge structure according to new project documents. At the same time, by establishing reverse links and transitive closures of relationships, the reasoning ability of the knowledge graph is enhanced.

[0051] Through the above technical innovations, this embodiment effectively solves several key problems in the knowledge extraction of project application documents: inaccurate entity recognition, incomplete relationship extraction, non-standard knowledge representation, etc. In practical applications, this solution can accurately construct a knowledge graph reflecting the technical characteristics of the project, providing structured knowledge support for subsequent text generation. It is particularly suitable for processing application documents with strong professionalism such as technological innovation projects and industrialization projects, significantly improving the systematicness and integrity of knowledge organization. The adaptive characteristics of this solution enable it to handle the project application requirements in different technical fields, providing a reliable knowledge basis for the intelligent generation of project application documents through continuous knowledge accumulation and optimization.

[0052] In an embodiment of the intelligent text generation method for project texts in this application, the following content may also be specifically included: Step S501: Input the entity node features and associated edge features in the knowledge graph into a graph convolutional network, aggregate and update the node features through multi-layer graph convolutional operations, calculate the association strength between entity nodes based on the attention mechanism, use the association strength as the weight coefficient of the knowledge association edge, screen the knowledge association edge based on the weight coefficient, remove the association edges with weight coefficients lower than the fourth preset threshold, and optimize the topological structure of the knowledge graph; Step S502: Perform subgraph partitioning on the optimized knowledge graph, extract the critical path based on the depth-first search algorithm, map the critical path to text structure features, perform sequence modeling on the text structure features using a recurrent neural network, generate a standardized text template containing technical features, innovation points, and association relationships, and cluster the standardized text template based on text similarity calculation to construct a differentiated text template library.

[0053] Optionally, this embodiment first designs an innovative graph convolutional processing mechanism. For the entity nodes and associated edges in the knowledge graph, a multi-channel feature representation is constructed. The entity node features include multiple dimensions such as technical attributes, domain classification, and innovation degree, and each dimension is encoded through an independent embedding layer; the associated edge features include information such as relationship type, directionality, and initial weight. These features are integrated into a unified representation through a feature fusion network and used as the input of the graph convolutional network. The graph convolutional network adopts a multi-layer structure, and the feature dimension of each layer is [512, 256, 128]. The information loss of the deep network is avoided through residual connections. In each layer of convolutional operation, the node features are updated by aggregating neighbor information, and the aggregation process adopts an attention weighting mechanism, and the calculation formula is: h_i^(l+1) = σ(Σ_j α_ij W^l h_j^l), where h_i is the node feature, α_ij is the attention weight, W is the learnable transformation matrix, and σ is the activation function.

[0054] This embodiment realizes an accurate method for calculating association strength. Based on the multi-head attention mechanism, the association importance between nodes is evaluated from different perspectives. Each attention head focuses on a different feature subspace, and the attention scores are calculated through dot product similarity. At the same time, positional encoding is introduced to consider the structural position information of nodes in the knowledge graph. The calculation of association strength also takes into account the integrity of the technical path, and higher weights are given to associations that form a complete technical solution. For the edges related to innovation points, special attention is paid to their support relationship with technical features, and the rationality of the association is ensured through interpretability analysis. The calculated association strength is used as the weight coefficient of the edge, and an adaptive screening threshold is set. The determination of the threshold is based on the global characteristics of the graph, including topological metrics such as average degree and clustering coefficient, and is dynamically adjusted considering the characteristics of different technical fields.

[0055] This embodiment designs a systematic graph structure optimization strategy. First, edges are screened based on the weight coefficient, and weak association edges below the threshold are removed. To avoid splitting important technical paths, connectivity checks are performed during the edge deletion process, and key connections are retained if necessary. The optimization process adopts an iterative strategy, and the quality metrics of the graph, including structural integrity, knowledge coverage, and reasoning reliability, are evaluated after each round of optimization. At the same time, the optimized graph structure is analyzed through the community discovery algorithm to ensure the rationality of the division of technical modules. To improve the usability of the graph, a bidirectional indexing mechanism is also established to support rapid positioning of relevant knowledge from different dimensions.

[0056] This embodiment realizes an intelligent subgraph partitioning method. An improved Louvain algorithm is used for community detection, and the algorithm adds technical relevance constraints on the basis of modularity optimization. The partitioned subgraphs represent relatively independent technical units, including complete technical features, innovation points, and dependencies. For each subgraph, the key technical paths are extracted through the depth-first search algorithm. The importance weights of nodes and the association strength of edges are considered during the search process, and paths that form the core technical solution are preferentially selected. The extracted paths are normalized and converted into structured feature sequences.

[0057] This embodiment innovatively designs a text template generation mechanism. A bidirectional LSTM network is used to model the structural feature sequences. The hidden layer state dimension of the network is 256, and the information flow is controlled through a gating mechanism. Special attention is paid to the logic and integrity of technical descriptions during the sequence modeling process, and key technical points and innovation highlights are emphasized through the attention mechanism. The model output is converted into a text template through a decoder. The template contains preset placeholders for subsequent personalized filling. To improve the applicability of the template, a template clustering mechanism is designed. Hierarchical clustering algorithm is used for clustering, and the similarity is calculated based on the structural and semantic features of the templates. The clustering results are used to construct a differentiated template library, providing matching template options for different types of projects.

[0058] In this embodiment, aiming at the practicability of the template library, a complete indexing and retrieval mechanism is implemented. A multi-dimensional tag system is established, including attributes such as technical field, innovation type, and document structure. Fast retrieval is achieved through an inverted index, supporting fuzzy matching and combined queries. At the same time, a template scoring mechanism is designed to dynamically evaluate and optimize templates according to historical usage effects. To maintain the timeliness of the template library, a continuous update mechanism is established, which can extract templates from new high-quality documents and integrate them into the library.

[0059] Through the above technological innovations, this embodiment effectively solves several key problems in the generation of project application documents: insufficient optimization of knowledge structure, inaccurate extraction of text frameworks, and insufficient adaptability of templates. In practical applications, this solution can accurately extract the framework features of project documents and generate high-quality text templates. It is particularly suitable for processing application documents with strong professionalism such as technological innovation projects and industrialization projects, significantly improving the standardization and efficiency of document generation. The adaptive characteristics of this solution enable it to meet the application requirements of different types of projects, and through continuous optimization and accumulation, it provides reliable structured support for the intelligent generation of project application documents.

[0060] In an embodiment of the intelligent generation method for project texts of this application, the following content may also be specifically included: Step S601: Retrieve technical R & D data from the enterprise R & D management system, normalize the fields of the technical R & D data, encode discrete features using one-hot encoding, standardize continuous features using min-max normalization, perform feature selection on the technical R & D data based on the principal component analysis method, construct the selected features into an enterprise technical feature matrix, and use a multi-layer perceptron to fill in the missing values in the enterprise technical feature matrix; Step S602: Input the enterprise technical feature matrix into a stacked autoencoder, compress the high-dimensional features into low-dimensional feature representations through the encoder, reconstruct the compressed features through the decoder, optimize the autoencoder parameters based on the reconstruction error, extract the low-dimensional features output by the encoder, perform feature encoding on the enterprise basic information, and combine the low-dimensional features with the enterprise basic information features using the feature splicing method to obtain an enterprise portrait vector.

[0061] Optionally, in this embodiment, a comprehensive technical R & D data processing mechanism is first designed. By calling the enterprise R & D management system through an interface, multi-dimensional data including R & D project information, technical achievement data, talent team composition, R & D investment statistics, etc. is obtained. To ensure data quality, a multi-level data standardization processing process is implemented. At the field level, the naming specifications of different data sources are unified, and the time format, measurement unit, and enumeration value definition are standardized. For technical indicator data, a special parsing rule is designed to accurately extract numerical features and unit information. In the data preprocessing stage, an outlier detection method based on the interquartile range is adopted, and a reasonable filtering threshold is set in combination with domain knowledge to ensure the validity of the data.

[0062] This embodiment implements an innovative feature engineering solution. For discrete features such as technology field classification, project status, and achievement type, an improved one-hot encoding method is adopted. Considering the high-cardinality problem of some features, the feature hashing technology is introduced to map high-dimensional categorical features to a lower-dimensional space while maintaining the distinctiveness of the features. For continuous features such as R & D investment, technical indicators, and team size, min-max normalization is used, and the normalization formula is: x_norm = (x - min) / (max - min + ε), where ε is a smoothing factor to avoid division-by-zero errors. In particular, for features with a long-tailed distribution, logarithmic transformation is first performed and then normalization is carried out to improve the distribution characteristics of the data.

[0063] This embodiment designs an efficient feature selection mechanism. The data is dimensionally reduced based on an improved principal component analysis method, and an innovative feature importance weight is introduced. When calculating the contribution degree of features, the business importance related to technological innovation is considered, and higher weights are assigned to key technical indicators. The cumulative variance contribution rate method is used to select the principal components, and cross-validation is combined to determine the optimal feature dimension. Special attention is paid to the correlation between technical features during the selection process, and information redundancy is avoided through correlation analysis. The finally selected features are constructed into an enterprise technical feature matrix, and each row of the matrix represents the feature vector of a technical dimension.

[0064] This embodiment implements an innovative missing value processing method. A multi-layer perceptron is used to construct a missing value prediction model. The model structure includes three hidden layers with dimensions [256, 128, 64] in sequence. The input layer receives the complete feature vector and marks the missing positions through a masking mechanism. The model is trained using complete samples to construct a training set, and missing situations are simulated by randomly masking some feature values. The loss function combines the mean square error and a regularization term to ensure the rationality of the predicted values. To improve the filling accuracy, the model also considers the logical relationship between features and ensures that the prediction results meet the business rules through a constraint layer.

[0065] In this embodiment, a deep feature compression scheme is designed. The constructed stacked autoencoder adopts a symmetric structure. Both the encoder and the decoder contain four layers, and the dimension configuration is [512, 256, 128, 64]. In the encoder, batch normalization and the LeakyReLU activation function are used in each layer to improve the expressive ability of the model. To maintain the discriminability of features, a contrastive learning loss is introduced in the latent space, which promotes the feature representations of similar enterprises to be close and those of different enterprises to be far apart. The original features are reconstructed through the decoder, and the reconstruction error is used as the main objective for model optimization. The dynamic learning rate strategy is adopted during the training process, and the learning rate is adaptively adjusted according to the performance of the validation set.

[0066] In this embodiment, a comprehensive feature fusion mechanism is implemented. Feature encoding is performed on the basic enterprise information (such as enterprise scale, industry, development stage, etc.), and a multi-modal feature fusion network is used to integrate it with the technical features. The fusion network includes an attention layer, which can dynamically adjust the importance of each dimension feature according to the characteristics of different types of projects. The dimension balance is considered during the feature splicing process, and the scales of features from different sources are ensured to be similar through feature transformation. The finally obtained enterprise portrait vector comprehensively depicts the enterprise's technological innovation ability and contains information in multiple dimensions such as technical strength, innovation potential, and development status.

[0067] Through the above technological innovations in this embodiment, several key problems in enterprise technical feature extraction are effectively solved: unstable data quality, insufficient feature expression, dimensionality disaster, etc. In practical applications, this scheme can accurately capture the technological innovation features of enterprises and generate high-quality enterprise portraits. It is particularly suitable for professional application scenarios such as technological innovation projects and industrialization projects, significantly improving the accuracy and comprehensiveness of enterprise feature representation. The adaptive characteristics of this scheme enable it to handle enterprises of different scales and types, and provide reliable enterprise feature support for the intelligent generation of project application documents through continuous data accumulation and model optimization.

[0068] In an embodiment of the method for intelligent generation of project text in this application, the following content may also be specifically included: Step S701: Add a conditional vector input module to the decoding layer of the pre-trained large model. Use the multi-head attention mechanism to project the enterprise portrait vector into the feature space of the decoding layer, fuse the projected features with the hidden state of the decoding layer, construct a template matching matrix based on the text structure features in the differential text template library, perform a tensor multiplication operation on the template matching matrix and the output of the decoding layer to obtain a decoding constraint matrix, and apply the Softmax function to the decoding constraint matrix to generate the word sequence probability distribution; Step S702: Decode the word sequence probability distribution using a greedy decoding strategy, calculate the cross-entropy loss between the decoding result and the standard text in the forward training corpus, introduce a regularization term to constrain the model parameters, use the Adam optimizer to update the gradients of the parameters of the pre-trained large model, and perform early stopping on the model based on the perplexity metric on the validation set to obtain an enterprise project text generation model.

[0069] Optionally, in this embodiment, an innovative decoding layer transformation scheme is first designed. For the decoding layer structure of the pre-trained large model, a dedicated conditional control module is added. This module uses a multi-layer perceptron to perform an initial transformation on the enterprise portrait vector, mapping it from the original feature space to the hidden state space of the model. Layer normalization and residual connections are used during the transformation process to ensure the stability of feature conversion. In the multi-head attention mechanism, 8 attention heads are set, and each head independently focuses on different feature dimensions. The attention calculation uses the scaled dot product method, and the calculation formula is: Attention(Q, K, V) = softmax(QK^T / √d)V, where Q, K, and V are the query, key, and value matrices respectively, and d is the feature dimension. In this way, the enterprise portrait information is effectively incorporated into the decoding process to guide the direction of text generation.

[0070] This embodiment implements an accurate feature fusion mechanism. The projected conditional features are dynamically fused with the hidden state of the decoding layer. The fusion process uses a gated update unit, which can adaptively adjust the information flow according to the content relevance. To enhance the generation ability of the model, a multi-level feature fusion strategy is innovatively introduced. Local feature fusion is performed at the shallow layer, focusing on specific technical descriptions; global feature fusion is performed at the deep layer to ensure the overall coherence of the text. The original information is retained through residual connections during the fusion process, avoiding the loss of important features. At the same time, the attention mechanism is used to calculate the importance weights of features in different layers, realizing the selective combination of features.

[0071] This embodiment designs a complete template constraint mechanism. A template matching matrix is constructed based on the differentiated text template library, and each element in the matrix represents the matching degree between the currently generated content and the template structure. The matching calculation considers multiple dimensions: structural similarity, technical relevance, and expression normality. The template matching matrix is combined with the output of the decoding layer through tensor multiplication operation to obtain the decoding probability distribution considering template constraints. A temperature parameter is introduced during the operation process to adjust the smoothness of the distribution, balancing the innovation and normality of generation. The Softmax function is applied to the obtained decoding constraint matrix to generate the final word sequence probability distribution.

[0072] This embodiment implements an efficient decoding optimization strategy. An improved greedy decoding method is adopted to select the word with the highest probability at each time step, and beam search is used to retain multiple candidate paths. A repetition penalty mechanism is introduced during decoding to avoid generating repeated technical descriptions. For the decoding result, the cross-entropy loss is calculated with the standard text in the forward training corpus, and the loss function is smoothed by label smoothing to improve the generalization ability of the model. To prevent overfitting, a variety of regularization constraints are introduced: L2 weight regularization controls the parameter norm, Dropout randomly deactivates neurons to increase the robustness of the model, and feature noise injection improves the generalization ability of the model.

[0073] This embodiment designs a systematic model optimization mechanism. The Adam optimizer is used for parameter update, and the learning rate of the optimizer adopts the cosine annealing strategy to gradually reduce the learning rate in the later stage of training. To improve the training efficiency, gradient accumulation technology is adopted to allow the use of a larger equivalent batch size. During the training process, the model performance is monitored by the perplexity metric on the validation set, and the early stopping mechanism is triggered when the perplexity does not decrease for consecutive epochs. At the same time, the model checkpoint saving and loading functions are implemented, and the model state can be restored from the best performance point.

[0074] This embodiment also optimizes the practical performance of the model. A dynamic model parameter adjustment mechanism is designed according to the declaration requirements of different types of projects. By analyzing the enterprise portrait features, the generation strategy is adaptively adjusted. For example, the document generation of technology innovation-oriented enterprises will pay more attention to the description of innovation points, while the achievement transformation-oriented enterprises will focus more on the demonstration of application effects. To improve the generation efficiency, the batch processing ability of the model is realized, supporting the parallel generation of multiple project documents. At the same time, through model compression and quantization technologies, the deployment complexity of the model is reduced.

[0075] Through the above technological innovations, this embodiment effectively solves several key problems in the generation of project declaration documents: the mismatch between the generated content and enterprise characteristics, the non-standard text structure, the unstable generation quality, etc. In practical applications, this solution can generate project declaration documents that meet the requirements according to enterprise characteristics, and is especially suitable for the generation of declaration materials with strong professionalism such as technology innovation projects and industrialization projects. By combining conditional control and template constraints, the quality and pertinence of the generated text are significantly improved. The adaptive characteristics of this solution enable it to handle different types of project declaration requirements and have broad application value. Through continuous optimization and improvement, the system can provide efficient and accurate project declaration document generation services for enterprises.

[0076] In order to solve the deficiencies of traditional technologies in the evaluation of training sample quality, the integrity of knowledge graphs, and the integration of enterprise characteristics, and significantly improve the professionalism and pertinence of project text generation, this application provides an embodiment of a project text intelligent generation device for implementing all or part of the content of the project text intelligent generation method. See Figure 2 , the project text intelligent generation device specifically includes the following contents: A training library construction module 10, which is used to perform text similarity clustering and deduplication on historical declaration documents, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical relevance score with the key points of technological research and development, the text fluency score, and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, screen high-score samples based on the comprehensive score to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; A knowledge graph construction module 20, which is used to construct the text content in the positive training corpus into knowledge graph nodes, use a relationship extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, input the knowledge graph into a text structure analysis model to generate a standardized text template, and construct a differential text template library; A project text generation module 30, which is used to obtain technical research and development data from an enterprise research and development management system and perform standardized processing, perform feature encoding on the technical research and development data to obtain an enterprise technical feature matrix, perform dimensionality reduction processing on the enterprise technical feature matrix based on an autoencoder, and perform feature splicing on the dimensionality reduction result and enterprise basic information to obtain an enterprise portrait vector; transform the structure of the decoding layer of a pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, optimize the parameters of the pre-trained large model using a cross-entropy loss function, and construct an enterprise project text generation model.

[0077] As can be seen from the above description, the project text intelligent generation device provided by the embodiments of the present application can construct a multi-dimensional scoring mechanism, and through the comprehensive evaluation of text coverage, technical relevance, fluency, and professional term standardization, achieve the accurate screening of positive and negative training corpora. Based on the relation extraction model and the external knowledge base, a complete knowledge graph is constructed, and the enterprise portrait vector is generated by integrating the enterprise technical feature matrix and basic information. By transforming the decoding layer structure of the pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in the evaluation of training sample quality, the integrity of the knowledge graph, and the integration of enterprise characteristics, and significantly improves the professionalism and pertinence of project text generation.

[0078] From a hardware perspective, in order to solve the deficiencies of traditional technologies in the evaluation of training sample quality, the integrity of the knowledge graph, and the integration of enterprise characteristics, and significantly improve the professionalism and pertinence of project text generation, the present application provides an embodiment of an electronic device for implementing all or part of the content in the project text intelligent generation method. The electronic device specifically includes the following: A processor, a memory, a communication interface, and a bus; wherein, the processor, the memory, and the communication interface complete communication with each other through the bus; the communication interface is used to implement information transmission between the project text intelligent generation device and related devices such as the core business system, the user terminal, and the relevant database. The logic controller can be a desktop computer, a tablet computer, a mobile terminal, etc., and this embodiment is not limited thereto. In this embodiment, the logic controller can be implemented with reference to the embodiments of the project text intelligent generation method and the embodiments of the project text intelligent generation device in the embodiments, and the content is incorporated herein, and the repeated parts will not be described again.

[0079] It can be understood 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.

[0080] In actual applications, part of the project text intelligent generation method can be executed on the electronic device side as described above, or all operations can be completed on the client device. Specifically, it can be selected according to the processing capacity of the client device and the limitations of the user usage scenario. The present application does not make any limitations in this regard. If all operations are completed on the client device, the client device may further include a processor.

[0081] The above-mentioned client device may have a communication module (i.e., communication unit), which can communicate with a remote server to realize data transmission with the server. The server may include a server on the task scheduling center side. In other implementation scenarios, it may also include a server of an intermediate platform, such as a server of 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 composed of multiple servers, or a server structure of a distributed device.

[0082] Figure 3 It is a schematic block diagram of the system composition of the electronic device 9600 according to an embodiment of the present application. As Figure 3 shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It should be noted that this Figure 3 is exemplary; other types of structures can also be used to supplement or replace this structure to implement telecommunication functions or other functions.

[0083] In one embodiment, the function of the project text intelligent generation method can be integrated into the central processing unit 9100. Among them, the central processing unit 9100 can be configured to perform the following controls: Step S101: Perform text similarity clustering and deduplication on historical declaration documents, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical correlation score with the key points of technological research and development, the text fluency score, and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, screen high-score samples based on the comprehensive score to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; Step S102: Construct the text content in the positive training corpus into knowledge graph nodes, use a relation extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, input the knowledge graph into a text structure analysis model to generate a standardized text template, and construct a differential text template library; Step S103: Obtain technology R & D data from the enterprise R & D management system and perform standardization processing. Perform feature encoding on the technology R & D data to obtain an enterprise technology feature matrix. Based on an autoencoder, perform dimensionality reduction processing on the enterprise technology feature matrix, and splice the dimensionality reduction result with the enterprise basic information to obtain an enterprise portrait vector. Reconstruct the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, and optimize the parameters of the pre-trained large model using the cross-entropy loss function to construct an enterprise project text generation model.

[0084] As can be seen from the above description, the electronic device provided in the embodiment of the present application realizes accurate screening of positive and negative training corpora through a comprehensive evaluation of text coverage, technical relevance, fluency, and professional term standardization by constructing a multi-dimensional scoring mechanism. Construct a complete knowledge graph based on the relation extraction model and the external knowledge base, and fuse the enterprise technology feature matrix and basic information to generate an enterprise portrait vector. By reconstructing the structure of the decoding layer of the pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improves the professionalism and pertinence of project text generation.

[0085] In another implementation, the project text intelligent generation device can be separately configured from the central processing unit 9100. For example, the project text intelligent generation device can be configured as a chip connected to the central processing unit 9100, and the functions of the project text intelligent generation method are realized through the control of the central processing unit.

[0086] As Figure 3 shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It should be noted that the electronic device 9600 does not necessarily have to include Figure 3 all the components shown in Figure 3 ; in addition, the electronic device 9600 may further include

[0087] As Figure 3 shown, the central processing unit 9100 is sometimes also referred to as a controller or operation control, and may include a microprocessor or other processor devices and / or logic devices. The central processing unit 9100 receives inputs and controls the operations of the various components of the electronic device 9600.

[0088] Among them, the memory 9140 can be, for example, one or more of a buffer, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable devices. It can store the above information related to failures, and can also store programs for executing relevant information. And the central processing unit 9100 can execute the programs stored in the memory 9140 to achieve information storage or processing, etc.

[0089] 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 supply power to the electronic device 9600. The display 9160 is used to display display objects such as images and texts. The display can be, for example, an LCD display, but is not limited thereto.

[0090] The memory 9140 can be a solid-state memory. For example, a read-only memory (ROM), a random access memory (RAM), a SIM card, etc. It can also be a memory that stores information even when powered off, can be selectively erased and has more data. Examples of such a memory are sometimes referred to as EPROMs, etc. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 can include an application / function storage unit 9142, which is used to store application programs and function programs or the processes for operating the electronic device 9600 through the central processing unit 9100.

[0091] The memory 9140 can also include a data storage unit 9143, which is used to store 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 can include various drivers of the electronic device for communication functions and / or for executing other functions of the electronic device (such as a messaging application, an address book application, etc.).

[0092] The communication module 9110 is a transmitter / receiver that transmits and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which can be the same as in the case of a conventional mobile communication terminal.

[0093] Based on different communication technologies, in the same electronic device, multiple communication modules 9110 can be provided, such as a cellular network module, a Bluetooth module, and / or a wireless local area network 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, so as to implement normal telecommunication functions. The audio processor 9130 can include any suitable buffers, decoders, amplifiers, etc. Additionally, the audio processor 9130 is also coupled to a central processor 9100, so that it is possible to record sound on the local device through the microphone 9132 and play the sound stored on the local device through the speaker 9131.

[0094] Embodiments of the present application also provide a computer-readable storage medium that can implement all steps of the project text intelligent generation method in which the execution subject in the above embodiments is a server or a client. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, all steps of the project text intelligent generation method in which the execution subject in the above embodiments is a server or a client are implemented. For example, when the processor executes the computer program, the following steps are implemented: Step S101: Perform text similarity clustering and deduplication on historical declaration documents, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical correlation score with the key points of technological research and development, the text fluency score, and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, screen high-score samples based on the comprehensive score to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; Step S102: Construct the text content in the positive training corpus into knowledge graph nodes, use a relationship extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, input the knowledge graph into a text structure analysis model to generate a standardized text template, and construct a differential text template library; Step S103: Obtain technical R & D data from the enterprise R & D management system and perform standardization processing. Perform feature encoding on the technical R & D data to obtain an enterprise technical feature matrix. Based on an autoencoder, perform dimensionality reduction processing on the enterprise technical feature matrix, and splice the dimensionality reduction result with the enterprise basic information to obtain an enterprise portrait vector. Transform the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, and use the cross-entropy loss function to optimize the parameters of the pre-trained large model to construct an enterprise project text generation model.

[0095] As can be seen from the above description, the computer-readable storage medium provided in the embodiment of the present application realizes the accurate screening of positive and negative training corpora through a comprehensive evaluation of text coverage, technical relevance, fluency, and professional term standardization by constructing a multi-dimensional scoring mechanism. Construct a complete knowledge graph based on the relationship extraction model and external knowledge base, and fuse the enterprise technical feature matrix and basic information to generate an enterprise portrait vector. By transforming the structure of the decoding layer of the pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in training sample quality evaluation, knowledge graph integrity, and enterprise feature integration, and significantly improves the professionalism and pertinence of project text generation.

[0096] The embodiment of the present application also provides a computer program product that can implement all steps in the project text intelligent generation method whose execution subject in the above embodiment is a server or a client. When the computer program / instructions are executed by a processor, the steps of the project text intelligent generation method are realized. For example, the computer program / instructions realize the following steps: Step S101: Perform text similarity clustering and deduplication on historical declaration documents, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical relevance score with the key points of technical R & D, the text fluency score, and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; Step S102: Construct the text content in the positive training corpus into knowledge graph nodes, use a relationship extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, and input the knowledge graph into a text structure analysis model to generate a standardized text template to construct a differential text template library; Step S103: Obtain technology R & D data from the enterprise R & D management system and perform standardization processing. Perform feature encoding on the technology R & D data to obtain an enterprise technology feature matrix. Perform dimensionality reduction processing on the enterprise technology feature matrix based on an autoencoder, and perform feature splicing on the dimensionality reduction result and enterprise basic information to obtain an enterprise portrait vector. Modify the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, and optimize the parameters of the pre-trained large model using the cross-entropy loss function to construct an enterprise project text generation model.

[0097] As can be seen from the above description, the computer program product provided by the embodiments of this application realizes the accurate screening of positive and negative training corpora through a comprehensive evaluation of text coverage, technical relevance, fluency, and professional term standardization by constructing a multi-dimensional scoring mechanism. Construct a complete knowledge graph based on the relation extraction model and external knowledge base, and fuse the enterprise technology feature matrix and basic information to generate an enterprise portrait vector. By modifying the structure of the decoding layer of the pre-trained large model and embedding the enterprise portrait vector as a conditional vector, personalized text generation is realized. This method effectively solves the deficiencies of traditional technologies in the evaluation of training sample quality, the integrity of the knowledge graph, and the integration of enterprise features, and significantly improves the professionalism and pertinence of project text generation.

[0098] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0099] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (apparatuses), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be realized by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate a device for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0100] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction means that implements the functions specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in the function.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in the function.

[0102] Specific embodiments are used in the present invention to illustrate the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present invention.

Claims

1. An intelligent generation method for project texts, characterized in that, The method includes: Performing text similarity clustering and deduplication on historical declaration documents, using a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculating the text coverage score of the historical declaration documents and the project declaration requirements, the technical correlation score with the key points of technological R & D, the text fluency score, and the professional term standardization score, performing weighted fusion on each score to obtain a comprehensive score, screening high-score samples based on the comprehensive score to construct a positive training corpus, and constructing a negative training corpus for samples with a comprehensive score lower than a preset threshold; Constructing knowledge graph nodes from the text content in the positive training corpus, using a relation extraction model to extract technological dependency relationships and innovation point associations from the text content to construct knowledge association edges, obtaining supplementary knowledge from external standard libraries and patent libraries and adding it to the knowledge graph, using a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjusting the topological structure of the knowledge graph according to the weight coefficients, and inputting the knowledge graph into a text structure analysis model to generate a standardized text template and constructing a differential text template library; Obtaining technological R & D data from the enterprise R & D management system and performing standardized processing, performing feature encoding on the technological R & D data to obtain an enterprise technological feature matrix, performing dimensionality reduction processing on the enterprise technological feature matrix based on an autoencoder, and performing feature splicing on the dimensionality reduction result and the enterprise basic information to obtain an enterprise portrait vector; modifying the structure of the decoding layer of a pre-trained large model, embedding the enterprise portrait vector as a conditional vector into the decoding layer, constructing a decoding constraint matrix based on the differential text template library, and optimizing the parameters of the pre-trained large model using a cross-entropy loss function to construct an enterprise project text generation model.

2. The intelligent generation method of item text according to claim 1, characterized in that The performing text similarity clustering and deduplication on historical declaration documents, using a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculating the text coverage score of the historical declaration documents and the project declaration requirements, the technical correlation score with the key points of technological R & D, the text fluency score, and the professional term standardization score, includes: Calculating the content similarity in historical declaration documents using a text clustering algorithm based on semantic distance, constructing a text similarity matrix, marking text pairs with a similarity higher than a first preset threshold as duplicate texts, performing deduplication processing on the duplicate texts based on a locality-sensitive hashing algorithm, using a word vector model to perform word sequence segmentation on the deduplicated historical declaration documents, and representing the word sequence as a multi-dimensional vector in combination with position encoding, and performing sentence-level embedding representation on the multi-dimensional vector based on an attention mechanism; Constructing a multi-branch scoring network model, calculating the text coverage score based on the term frequency-inverse document frequency algorithm in the first branch, calculating the technical correlation score by extracting technical elements based on named entity recognition in the second branch, calculating the sentence fluency score based on a language model in the third branch, calculating the term standardization score based on a preset professional term dictionary in the fourth branch, and inputting the scores of each branch into a fully connected layer for weighted fusion to obtain a comprehensive score.

3. The intelligent generation method of project text according to claim 1, wherein Performing weighted fusion on the scores to obtain a comprehensive score, screening high-score samples based on the comprehensive score to construct a positive training corpus, and constructing a negative training corpus for samples with a comprehensive score lower than a preset threshold, including: Using an adaptive weight allocation mechanism to calculate the weights of the text coverage score, technical relevance score, fluency score, and term standardization score, applying the Softmax function to normalize the weights, performing a linear combination of the scores based on the normalized weights to obtain a text comprehensive score, and performing Z-score standardization processing on the text comprehensive score to obtain a standardized score; Extracting text content, project classification labels, and application result markers from historical application documents, forming positive training sample pairs consisting of text content, project classification labels, and application result markers with a standardized score higher than a second preset threshold to construct a positive training corpus, and forming negative training sample pairs consisting of text content, project classification labels, and application result markers with a standardized score lower than a third preset threshold to construct a negative training corpus.

4. The intelligent generation method of project text according to claim 1, characterized in that Constructing the text content in the positive training corpus into knowledge graph nodes, using a relation extraction model to extract technical dependency relationships and innovation point associations from the text content to construct knowledge association edges, and obtaining supplementary knowledge from external standard libraries and patent libraries and adding it to the knowledge graph, including: Performing syntactic analysis and dependency syntactic parsing on the text content in the positive training corpus, identifying technical feature words, technical field words, and innovation point description words from the parsing results, constructing the technical feature words, technical field words, and innovation point description words into entity nodes of the knowledge graph, using a string matching algorithm to deduplicate and merge the entity nodes, aggregating similar entity nodes based on text semantic similarity calculation, and assigning a unique identifier to each entity node; Using a relation extraction model based on the attention mechanism to scan the text content, identifying the technical dependency relationships and innovation point associations between entity nodes, constructing the technical dependency relationships and innovation point associations into association edges of the knowledge graph, retrieving supplementary knowledge related to the entity nodes from external standard libraries and patent libraries, and constructing the supplementary knowledge into new entity nodes and association edges and adding them to the knowledge graph.

5. The intelligent generation method of item text according to claim 1, characterized in that Using a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjusting the topological structure of the knowledge graph according to the weight coefficients, and inputting the knowledge graph into a text structure analysis model to generate a standardized text template and constructing a differential text template library, including: Inputting the entity node features and association edge features in the knowledge graph into a graph convolutional network, aggregating and updating the node features through multi-layer graph convolutional operations, calculating the association strength between entity nodes based on the attention mechanism, using the association strength as the weight coefficient of the knowledge association edge, screening the knowledge association edges based on the weight coefficient, removing association edges with a weight coefficient lower than a fourth preset threshold, and optimizing the topological structure of the knowledge graph; Partition the optimized knowledge graph into subgraphs, extract the critical path based on the depth-first search algorithm, map the critical path to text structure features, use a recurrent neural network to perform sequence modeling on the text structure features, generate a standardized text template containing technical features, innovation points, and association relationships, and cluster the standardized text template based on text similarity calculation to construct a differentiated text template library.

6. The intelligent generation method of project text according to claim 1, wherein Obtain technical R & D data from the enterprise R & D management system and perform standardized processing, perform feature encoding on the technical R & D data to obtain an enterprise technical feature matrix, perform dimensionality reduction processing on the enterprise technical feature matrix based on an autoencoder, and splice the dimensionality reduction result with enterprise basic information to obtain an enterprise portrait vector, including: Retrieve technical R & D data from the enterprise R & D management system, normalize the fields of the technical R & D data, encode discrete features using one-hot encoding, standardize continuous features using min-max normalization, perform feature selection on the technical R & D data based on the principal component analysis method, construct the selected features into an enterprise technical feature matrix, and use a multi-layer perceptron to fill in the missing values in the enterprise technical feature matrix; Input the enterprise technical feature matrix into a stacked autoencoder, compress the high-dimensional features into low-dimensional feature representations through the encoder, reconstruct the compressed features through the decoder, optimize the autoencoder parameters based on the reconstruction error, extract the low-dimensional features output by the encoder, perform feature encoding on the enterprise basic information, and use the feature splicing method to combine the low-dimensional features with the enterprise basic information features to obtain an enterprise portrait vector.

7. The intelligent generation method of project text according to claim 1, wherein Transform the structure of the decoding layer of the pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differentiated text template library, optimize the parameters of the pre-trained large model using the cross-entropy loss function, and construct an enterprise project text generation model, including: Add a conditional vector input module to the decoding layer of the pre-trained large model, project the enterprise portrait vector into the feature space of the decoding layer using the multi-head attention mechanism, fuse the projected features with the hidden state of the decoding layer, construct a template matching matrix based on the text structure features in the differentiated text template library, perform a tensor multiplication operation on the template matching matrix and the output of the decoding layer to obtain a decoding constraint matrix, and apply the Softmax function to the decoding constraint matrix to generate a word sequence probability distribution; Decode the word sequence probability distribution using the greedy decoding strategy, calculate the cross-entropy loss between the decoding result and the standard text in the forward training corpus, introduce a regularization term to constrain the model parameters, use the Adam optimizer to perform gradient updates on the parameters of the pre-trained large model, and perform early stopping on the model based on the perplexity metric on the validation set to obtain an enterprise project text generation model.

8. An intelligent device for generating project texts, characterized in that, The device includes: A training library construction module, which is used to cluster and remove duplicates of historical declaration documents by text similarity, use a text vectorization model to perform embedding representation on the deduplicated historical declaration documents, calculate the text coverage score of the historical declaration documents and the project declaration requirements, the technical relevance score with the key points of technological R & D, the text fluency score and the professional term standardization score, perform weighted fusion on each score to obtain a comprehensive score, screen high-score samples based on the comprehensive score to construct a positive training corpus, and construct a negative training corpus for samples with a comprehensive score lower than a preset threshold; A knowledge graph construction module, which is used to construct the text content in the positive training corpus into knowledge graph nodes, use a relationship extraction model to extract technical dependence relationships and innovation point associations from the text content to construct knowledge association edges, obtain supplementary knowledge from external standard libraries and patent libraries and add it to the knowledge graph, use a graph convolutional network to calculate the weight coefficients of the knowledge association edges, adjust the topological structure of the knowledge graph according to the weight coefficients, input the knowledge graph into a text structure analysis model to generate a standardized text template, and construct a differential text template library; A project text generation module, which is used to obtain technical R & D data from an enterprise R & D management system and perform standardized processing, perform feature encoding on the technical R & D data to obtain an enterprise technical feature matrix, perform dimensionality reduction processing on the enterprise technical feature matrix based on an autoencoder, and perform feature splicing on the dimensionality reduction result and enterprise basic information to obtain an enterprise portrait vector; transform the structure of the decoding layer of a pre-trained large model, embed the enterprise portrait vector as a conditional vector into the decoding layer, construct a decoding constraint matrix based on the differential text template library, use a cross-entropy loss function to optimize the parameters of the pre-trained large model, and construct an enterprise project text generation model.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that When the processor executes the program, it implements the steps of the project text intelligent generation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the project text intelligent generation method according to any one of claims 1 to 7.

Images