Enterprise project text processing model directional training method and device

Abstract

The embodiment of the invention provides a directional training method and device for an enterprise project text processing model, and the method and device achieve the screening of high-quality training samples through the construction of a project text scoring model and the multi-dimensional evaluation of the text coverage degree, the technical correlation degree and the structural integrity. And analyzing a text structure relationship based on the knowledge graph and the graph convolutional network, and constructing a differentiated text template library. The system performs directional training on a pre-training large model in combination with an enterprise portrait vector, and continuously optimizes model parameters and a knowledge base by adopting an incremental learning strategy. According to the method, the defects of a traditional technology in the aspects of training sample screening, knowledge structure analysis and model personalized customization are effectively overcome, and the practicability and adaptability of a project text processing model are remarkably improved.

Description

Technical Field The present application relates to the field of data processing, and particularly to a method and device for directionally training an enterprise project text processing model. Background Art There are obvious deficiencies in the existing methods for training project text processing models. Traditional methods lack a systematic evaluation mechanism for enterprise historical declaration documents, and cannot effectively screen high-quality training samples, resulting in unsatisfactory model training effects. At the same time, the existing methods have relatively superficial analysis of text structures and technical features.

[0001] In addition, there are bottlenecks in knowledge graph construction and template generation in the existing technology. Most systems fail to fully utilize the reference relationships and semantic associations between texts, and it is difficult to construct a dynamic graph reflecting the knowledge structure. The text templates lack pertinence and differential features, affecting the actual application effect of the model.

[0002] There are technical shortcomings in the directional training and continuous optimization of the existing systems. There is a lack of in-depth analysis and integration of enterprise characteristics, and personalized customization of the model cannot be achieved. Solving these problems is of great significance for improving the practicability and adaptability of the project text processing model. Summary of the Invention In view of the problems in the existing technology, the present application provides a method and device for directionally training an enterprise project text processing model, which can effectively solve the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improve the practicability and adaptability of the project text processing model.

[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 directionally training an enterprise project text processing model, including: Constructing a project text scoring model, inputting enterprise historical declaration documents into the project text scoring model, using a text classifier to identify standardized description segments in the historical declaration documents, calculating a text coverage score, extracting key technical feature words in the historical declaration documents, calculating a technical association score based on a feature word co-occurrence matrix, performing structural semantic analysis on the historical declaration documents to obtain a structural integrity score, performing weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screening high-score samples based on the comprehensive score to construct a training corpus; Construct the text content in the training corpus into knowledge graph nodes, construct knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, calculate the weight coefficients of the knowledge association edges 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; Collect multi-dimensional data of the enterprise to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model, input the test text into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

[0004] Further, it also includes: establishing a text structure specification library, segmenting the enterprise historical declaration documents into text fragments by paragraph, the text classifier constructs text feature vectors based on a bidirectional long short-term memory network, uses a convolutional neural network to classify and identify the text feature vectors, matches and labels the recognition results with the text structure specification library, and calculates the proportion of the labeled fragments in the overall document to obtain a text coverage score; Construct a technical dictionary library, use a conditional random field model to identify technical feature words from the historical declaration documents, construct a feature word co-occurrence matrix based on the distribution positions of the technical feature words in the documents, calculate the association strength between the technical feature words in the feature word co-occurrence matrix based on the cosine similarity between word vectors, perform normalization processing on the association strength to obtain a technical association degree score, and use a deep neural network to perform structural semantic analysis on the historical declaration documents, and calculate the semantic coherence between document paragraphs based on an attention mechanism to obtain a structural integrity score.

[0005] Further, it also includes: using the principal component analysis method to calculate the contribution degrees of the text coverage score, the technical association degree score, and the structural integrity score, using the contribution degrees as weight coefficients, performing a linear weighted combination on the text coverage score, the technical association degree score, and the structural integrity score, and mapping the weighted combination result to a specified score interval based on the maximum-minimum normalization method to obtain a comprehensive score; Establish a hierarchical screening matrix, sort the historical declaration documents in descending order according to the comprehensive score, use a clustering analysis method to calculate the distribution characteristics of the comprehensive score, determine the screening threshold based on the distribution characteristics, use the historical declaration documents higher than the screening threshold as training samples, and perform text vectorization processing on the training samples to construct a training corpus.

[0006] Further, it also includes: performing word segmentation on the text content in the training corpus, extracting technical entity words from the word segmentation results using a named entity recognition model, constructing knowledge graph nodes with the technical entity words, extracting the relationship attributes between the technical entity words using dependency syntactic analysis, adding attribute labels to the knowledge graph nodes based on the syntactic tree structure, and constructing an initial node set of the knowledge graph; Calculating a semantic similarity matrix between the knowledge graph nodes using a text vectorization model, normalizing the semantic similarity matrix, taking the normalized semantic similarity value as the weight of the knowledge association edge, and establishing knowledge association edges between the knowledge graph nodes based on the breadth-first search algorithm to construct an edge set of the knowledge graph.

[0007] Further, it also includes: inputting the knowledge graph into a graph convolutional network, representing the connection relationship of the knowledge association edges using an adjacency matrix, performing a convolution operation on the feature vectors of the knowledge graph nodes, processing the convolution result through a non-linear activation function, iteratively updating the parameters of the graph convolutional network based on the backpropagation algorithm, calculating the weight coefficient of the knowledge association edge, and pruning the edge set of the knowledge graph according to the weight coefficient, retaining the knowledge association edges with weight coefficients higher than a preset threshold; Constructing a text structure analysis model, which uses a hierarchical attention mechanism to analyze the structure of the adjusted knowledge graph, extracts the structure features of the knowledge graph based on a deep neural network, maps the structure features to text layout rules, generates a standardized text template according to the text layout rules, and performs item type annotation on the standardized text template to construct a differentiated text template library.

[0008] Further, it also includes: obtaining technical R & D data from an enterprise R & D management system, performing standardization processing on the technical R & D data using a data cleaning method, performing feature encoding on the technical R & D data to obtain an enterprise technical feature matrix, performing dimensionality reduction processing on the enterprise technical feature matrix based on an autoencoder, and performing feature splicing on the dimensionality reduction result and enterprise basic information to obtain an enterprise portrait vector; Reforming the structure of the decoding layer of the pre-trained large model, embedding the enterprise portrait vector as a conditional vector into the decoding layer, inputting the training corpus into the encoding layer for feature extraction, constructing a decoding constraint matrix based on the differentiated text template library, calculating the error between the model prediction output and the training samples using a cross-entropy loss function, and optimizing the parameters of the pre-trained large model through the backpropagation algorithm to construct an enterprise project text generation model.

[0009] Further, it further includes: inputting the test text into the enterprise project text generation model, decoding the test text using the beam search algorithm, calculating the matching degree score between the generation result and the templates in the differential text template library, constructing a text reconstruction error based on the variational autoencoder, using the weighted sum of the matching degree score and the text reconstruction error as the model evaluation index, and performing gradient update on the network parameters of the enterprise project text generation model; Construct a sample pool dynamic update mechanism, add the generation results higher than the evaluation threshold to the training corpus, extract new text structure features from the generation results based on the text clustering method, integrate the text structure features into the differential text template library, and perform online fine-tuning on the enterprise project text generation model using the incremental learning strategy to update the knowledge representation of the model.

[0010] In a second aspect, the present application provides an enterprise project text processing model directed training device, including: A project analysis module, configured to construct a project text scoring model, input the enterprise historical declaration documents into the project text scoring model, use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association degree score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association degree score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; A content processing module, configured to construct the text content in the training corpus into knowledge graph nodes, construct knowledge association edges based on the citation relationship and semantic similarity between the knowledge graph nodes, 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 model training module, configured to collect enterprise multi-dimensional data to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model, input a test text into the enterprise project text generation model for text generation testing, perform dynamic optimization on the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using the incremental learning strategy.

[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 enterprise project text processing model directed training 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 enterprise project text processing model directed training 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 enterprise project text processing model directed training method are implemented.

[0014] As can be seen from the above technical solutions, the present application provides an enterprise project text processing model directed training method and device. By constructing a project text scoring model, high-quality training samples are screened through multi-dimensional evaluations of text coverage, technical relevance, and structural integrity. Based on the knowledge graph and graph convolutional network, the text structure relationship is analyzed to construct a differentiated text template library. The system combines the enterprise portrait vector to conduct directed training on the pre-trained large model, and adopts an incremental learning strategy to continuously optimize the model parameters and knowledge base. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model. 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, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 It is a schematic flowchart of the enterprise project text processing model directed training method in the embodiments of the present application; Figure 2 It is a structural diagram of the enterprise project text processing model directed training device in the embodiments of the present application; Figure 3 It is a schematic structural diagram of the electronic device in the embodiments of the present application.

[0016] Reference Signs: 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 of 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 regulations of national laws and regulations.

[0018] Considering the problems existing in the prior art, the present application provides a method and device for directional training of an enterprise project text processing model. By constructing a project text scoring model, high-quality training samples are screened through multi-dimensional evaluation of text coverage, technical relevance, and structural integrity. Based on a knowledge graph and a graph convolutional network, the text structure relationship is analyzed, and a differentiated text template library is constructed. The system combines the enterprise portrait vector to perform directional training on the pre-trained large model, and adopts an incremental learning strategy to continuously optimize the model parameters and knowledge base. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model.

[0019] To effectively solve the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improve the practicability and adaptability of the project text processing model, the present application provides an embodiment of a method for directional training of an enterprise project text processing model. Refer to Figure 1 The method for directional training of the enterprise project text processing model specifically includes the following contents: Step S101: Construct a project text scoring model. Input the enterprise's historical declaration documents into the project text scoring model. Use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words from the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; Optionally, in this embodiment, a scoring model specifically for enterprise project declaration documents is first designed. The model adopts a multi-level evaluation architecture, including three dimensions: basic text normality evaluation, technical relevance evaluation, and structural integrity evaluation. When processing the enterprise's historical declaration documents, document preprocessing is first performed, including format unification, encoding conversion, and noise removal, to ensure the quality of the input data.

[0020] In this embodiment, an improved BERT model is used as the basic architecture in the design of the text classifier. The model contains 12 layers of Transformer encoders, with 12 attention heads set in each layer, and the hidden layer dimension is 768. To adapt to the characteristics of project declaration documents, a large number of project declaration-related corpora are added in the pre-training stage to enhance the model's understanding ability of professional terms and industry expressions. The output dimension of the text classifier is set to multiple predefined description categories, such as project background, technical solution, innovation point, etc.

[0021] This embodiment implements an accurate text coverage calculation method. By designing a standardized description template library, the enterprise declaration documents are matched with the templates by paragraph. The calculation formula for the coverage score is: Coverage = Σ(wi·mi) / N, where wi is the weight of the i-th standardized description segment, mi is the matching degree of this segment, and N is the total number of standard templates. This calculation method can accurately reflect the coverage of the necessary content in the document.

[0022] This embodiment innovatively implements a technical feature word extraction mechanism. An improved TextRank algorithm is used, combined with an industry dictionary and a professional term library, to identify key technical words from the document. The algorithm considers multiple factors such as word frequency, position importance, and context relevance, and obtains the importance score of each word through iterative calculation. For newly emerging technical terms, the system dynamically supplements them through context semantic analysis.

[0023] This embodiment designs an efficient method for constructing a co-occurrence matrix of feature words. Using the sliding window technique, the co-occurrence frequency of technical feature words in the document is statistically calculated to construct an N×N co-occurrence matrix (N is the number of feature words). The technical relevance score is obtained through matrix decomposition and word vector calculation: Relevance = tr(W^T·C·W), where W is the word vector matrix, C is the co-occurrence matrix, and tr represents the trace of the matrix.

[0024] In the structural semantic analysis section of this embodiment, a document structure evaluation model based on deep learning is implemented. The model uses a hierarchical LSTM network, which can capture the semantic dependency relationships between different levels of the document. The coherence score between paragraphs is calculated through the attention mechanism: Coherence = Σ(αij·sim(hi,hj)), where αij is the attention weight, and hi and hj are the semantic representations of the paragraphs.

[0025] This embodiment innovatively designs a scoring fusion mechanism. The principal component analysis method is used to determine the weight coefficients of each score, and the weight calculation takes into account the variance contribution rate and correlation of the indicators. The fusion formula is: Score = λ1·Coverage + λ2·Relevance + λ3·Coherence, where λ1, λ2, and λ3 are dynamically adjusted weight coefficients.

[0026] This embodiment implements an intelligent sample screening strategy. Based on the GMM clustering model, the distribution characteristics of the comprehensive score are analyzed, and the screening threshold is adaptively determined. The setting of the threshold takes into account the mean, variance, and distribution skewness of the score to ensure that the selected samples are representative and diverse. For special types of project application documents, the system will appropriately adjust the threshold standard.

[0027] This embodiment designs a mechanism for constructing a training corpus. The selected high-score samples are vectorized, and the Word2Vec and BERT models are used to extract text features. The feature vector contains semantic information at the word level and sentence level, and the dimension is set to 768. Batch normalization is used to maintain the consistency of the features.

[0028] Through the above technical innovations, this embodiment effectively solves multiple key problems in the evaluation of enterprise project application documents: inconsistent evaluation criteria, insufficient understanding of technical content, inaccurate structural evaluation, etc. In practical applications, this solution can accurately evaluate the quality of application documents and provide high-quality samples for subsequent model training. It is particularly suitable for the evaluation of application documents with strong professionalism such as scientific and technological innovation projects and R & D projects, significantly improving the accuracy and efficiency of evaluation. The adaptive characteristics of this solution enable it to handle different types of project application requirements and have a broad application prospect. By analyzing and screening a large number of historical application documents, the system can continuously optimize the evaluation criteria and improve the quality of the sample library.

[0029] Step S102: Construct the text content in the training corpus into knowledge graph nodes, construct knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, calculate the weight coefficients of the knowledge association edges 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; Optionally, in this embodiment, a dedicated knowledge graph node construction mechanism is first implemented. Based on the text content in the training corpus, an improved named entity recognition model is used for entity extraction. This model adopts a BiLSTM-CRF architecture and enhances the recognition ability of professional terms through the dual representation of word vectors and character vectors. The input layer dimension of the model is 768, the number of hidden layer units is 256, and it supports the recognition of various types of entities such as technical solutions, innovation points, and application scenarios.

[0030] This embodiment designs an innovative entity attribute annotation method. The attribute information of entities is extracted through dependency syntactic analysis. A syntactic tree is constructed using Stanford Parser, and modifiers and determiners are extracted from the syntactic structure as entity attributes. For example, for technical solution entities, the system extracts attribute information such as their technical parameters, application conditions, and implementation effects to form a complete entity description.

[0031] This embodiment implements an intelligent knowledge association edge construction mechanism. First, an initial connection is established based on the explicit citation relationships within the document, including literature citations, technical dependencies, etc. Then, the semantic similarity between nodes is calculated through the BERT model: Similarity = cosine(BERT(ni), BERT(nj)), where ni and nj are the text representations of the nodes. Knowledge association edges are established between node pairs with a similarity higher than the threshold.

[0032] For the optimization of the graph structure in this embodiment, an improved graph convolutional network is designed. The network includes three graph convolutional layers, and the feature dimensions of each layer are 256, 128, and 64 in sequence. The feature update formula for nodes is: H(l+1) = σ(D^(-1 / 2)AD^(-1 / 2)H(l)W(l)), where A is the adjacency matrix, D is the degree matrix, H is the node feature, and W is the learnable weight matrix.

[0033] This embodiment innovatively implements an edge weight calculation method. The importance scores between nodes are calculated through an attention mechanism, and the attention scores consider the local structural features and global position information of the nodes. The weight calculation formula is: Weight = softmax(Q·K^T / √d)·V, where Q, K, and V are the query, key, and value matrices respectively, and d is the feature dimension.

[0034] This embodiment designs a dynamic adjustment strategy for the atlas structure. Based on the calculated edge weights, adaptive thresholds are used to prune the edges. The threshold values are dynamically determined by the global connectivity and local clustering degree of the graph, ensuring that the adjusted atlas maintains necessary connections while avoiding excessive density. For important technical routes and innovation points, the system will maintain their key connections.

[0035] In the text structure analysis section of this embodiment, an analysis model based on hierarchical attention is implemented. The model can parse the hierarchical structure of the knowledge graph from bottom to top, capturing the dependencies between different levels through the multi-head attention mechanism. The design of the attention layer takes into account the node types and attribute information to better understand the organization logic of knowledge.

[0036] This embodiment innovatively designs a template generation mechanism. Based on the structural features of the atlas, the organization patterns of the document are extracted, including chapter arrangement, content layout, and logical order. The template generation process combines rules and learning to ensure that the generated templates meet the specification requirements while maintaining flexibility.

[0037] For the differential processing of templates in this embodiment, an intelligent classification and annotation system is implemented. The templates are classified according to the project type, technical field, and innovation characteristics, and a multi-dimensional template index is constructed. The system analyzes the project characteristics through a deep learning model and automatically matches the most suitable template type.

[0038] This embodiment designs a dynamic update mechanism for the template library. By continuously analyzing new application documents, new document structure features are extracted to expand the coverage of the template library. The update process adopts an incremental learning method to maintain the stability of existing templates while adapting to new demand changes.

[0039] Through the above technological innovations, this embodiment effectively solves several key problems in the knowledge organization of project application documents, such as non-standard knowledge representation, inaccurate association relationships, and poor template adaptability. In practical applications, this solution can accurately extract and organize the knowledge content in project documents and generate standardized templates that meet the characteristics of different projects. It is particularly suitable for processing application documents with strong professionalism, such as scientific and technological innovation projects and industrialization projects, significantly improving the efficiency of knowledge management and the quality of document generation. The adaptive characteristics of this solution enable it to handle the application requirements of different fields and different types of projects, with broad application value. Through the dynamic optimization of the knowledge graph and the continuous update of templates, it is ensured that the system can adapt to the changing project application requirements and provide more accurate document generation support.

[0040] Step S103: Collect multi-dimensional data of the enterprise to construct an enterprise portrait vector. Input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model. Input the test text into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

[0041] Optionally, in this embodiment, a comprehensive enterprise data collection mechanism is first designed. By docking with the enterprise R & D management system, technical R & D data is obtained, including R & D project information, patent data, technical achievements, etc. At the same time, business data is collected from the enterprise ERP system, including product information, market data, financial indicators, etc. The incremental update strategy is adopted in the collection process to ensure the timeliness and integrity of the data.

[0042] This embodiment realizes an innovative data cleaning method. Special cleaning rules are designed for data from different sources, including format standardization, missing value processing, outlier detection, etc. For technical R & D data, a professional term comparison table is used for standardization to ensure the standardization of technical descriptions. The cleaned data is verified through a quality assessment model to ensure data quality.

[0043] This embodiment designs an efficient feature encoding mechanism. Convert the enterprise technical data into a feature matrix, and the encoding dimensions include multiple dimensions such as technical field, innovation degree, application value, etc. The feature representation adopts a combination of multi-hot encoding and embedding vectors, and the weights of each dimension are determined through feature importance analysis: Feature = Σ(wi·xi), where wi is the feature weight and xi is the feature value.

[0044] This embodiment implements an improved autoencoder model for dimensionality reduction. Both the encoder and the decoder adopt a multi-layer perceptron structure, and the hidden layer dimensions are 512, 256, and 128 in sequence. By adding regularization constraints, over-compression of feature information is prevented. The loss function combined with reconstruction error and KL divergence is used for model training to ensure the representativeness of the features after dimensionality reduction.

[0045] This embodiment innovatively realizes a method for constructing an enterprise portrait vector. Fuse the dimensionality-reduced technical features with the basic enterprise information, and the attention mechanism is used in the fusion process to dynamically adjust the importance of different features. The finally generated enterprise portrait vector has a dimension of 256 and contains comprehensive feature information of the enterprise.

[0046] In this embodiment, a conditional generation architecture is designed for the transformation of the pre-trained model. Based on the original Transformer decoding layer, a conditional injection module is added, and the enterprise portrait vector is embedded as conditional information into the generation process. The attention calculation takes into account the influence of the conditional vector: Attention = softmax((Q + C)·K^T / √d)·V, where C is the projection of the conditional vector.

[0047] This embodiment implements an innovative training strategy. A multi-task learning framework is adopted to optimize both the text generation quality and the template matching degree simultaneously. During the training process, the weights of each task are dynamically adjusted to adaptively balance different objectives according to the performance of the validation set. For special types of project documents, the system will increase the weights of the corresponding training samples.

[0048] This embodiment designs a comprehensive test and evaluation mechanism. The beam search algorithm is used for text generation, and the search width is set to 4. The optimal generation result is selected through a re-scoring mechanism. The evaluation metrics include text fluency, professional accuracy, and template matching degree, and the model optimization is guided by the combined score.

[0049] This embodiment innovatively implements a parameter optimization strategy. Based on the test and evaluation results, the model parameters are updated by combining gradient clipping and learning rate scheduling. The importance of different layers is considered during the optimization process, and a smaller learning rate is used for the key layers to maintain stability.

[0050] This embodiment designs an efficient incremental learning mechanism. A sample pool management system is established, and high-quality generation results are dynamically added to the training corpus. The sliding window strategy is adopted during the update process to maintain the timeliness of the samples. At the same time, new document patterns are identified based on text clustering technology to expand the coverage of the template library.

[0051] This embodiment implements a model knowledge update mechanism. By continuously learning new enterprise project features, the knowledge representation of the model is dynamically adjusted. The progressive fine-tuning strategy is adopted during the update process to avoid the problem of catastrophic forgetting. For newly emerging technical fields or project types, the system will actively expand the corresponding knowledge representation.

[0052] Through the above technological innovations, this embodiment effectively solves multiple key problems in enterprise project text generation, such as incomplete representation of enterprise characteristics, poor model adaptability, and untimely knowledge update. In practical applications, this solution can generate personalized project declaration documents according to the characteristics of enterprises, achieving the intelligence and precision of document generation. It is particularly suitable for the generation of declaration documents with strong professionalism, such as technological innovation projects and industrialization projects, significantly improving the efficiency and quality of document generation. The adaptive characteristics of this solution enable it to continuously optimize and update, adapt to the project declaration needs of different enterprises, and have broad application value. Through the incremental learning and dynamic update mechanism, it ensures that the system can continuously improve its performance and provide more and more accurate document generation services for enterprises.

[0053] As can be seen from the above description, the method for directed training of an enterprise project text processing model provided by the embodiment of this application can screen high-quality training samples through multi-dimensional evaluation of text coverage, technical relevance, and structural integrity by constructing a project text scoring model. Analyze the text structure relationship based on the knowledge graph and graph convolutional network to construct a differentiated text template library. The system combines the enterprise portrait vector to conduct directed training on the pre-trained large model, and adopts an incremental learning strategy to continuously optimize the model parameters and knowledge base. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicality and adaptability of the project text processing model.

[0054] In an embodiment of the method for directed training of an enterprise project text processing model of this application, it may specifically include the following content: Step S201: Establish a text structure specification library, segment the enterprise historical declaration documents into text fragments by paragraphs, the text classifier constructs text feature vectors based on a bidirectional long short-term memory network, uses a convolutional neural network to classify and identify the text feature vectors, matches and labels the recognition results with the text structure specification library, and calculates the proportion of the labeled fragments in the overall document to obtain a text coverage score; Step S202: Construct a technical dictionary library, use a conditional random field model to identify technical feature words from the historical declaration documents, construct a feature word co-occurrence matrix based on the distribution positions of the technical feature words in the documents, calculate the association strength between the technical feature words in the feature word co-occurrence matrix based on the cosine similarity between word vectors, normalize the association strength to obtain a technical relevance score, and use a deep neural network to conduct structural semantic analysis on the historical declaration documents, and calculate the semantic coherence between document paragraphs based on the attention mechanism to obtain a structural integrity score.

[0055] Optionally, in this embodiment, a dedicated text structure specification library construction mechanism is first designed. Based on the analysis of a large number of high-quality project application documents, standardized document structure patterns are extracted, including key sections such as project background, technical solutions, innovation points, implementation plans, etc. The specification library adopts a multi-level structure, including not only organizational specifications at the chapter level, but also expression requirements at the paragraph level and language specifications at the sentence level.

[0056] This embodiment implements an intelligent paragraph segmentation strategy. The paragraph boundaries are identified through natural language processing techniques, considering multiple features such as punctuation marks, keywords, and topic coherence. The sliding window technique is used in the segmentation process, and the window size is dynamically adjusted according to the content complexity to ensure the semantic integrity of the segmentation. For special formats such as tables and chart descriptions, special processing rules are adopted.

[0057] For text feature extraction in this embodiment, a feature vector construction model based on BiLSTM is designed. The model contains a bidirectional LSTM layer with a hidden layer dimension of 512, which captures context dependencies through the fusion of forward and backward states. The calculation formula for the feature vector is: h = [hf; hb], where hf is the forward state and hb is the backward state, ensuring that the feature representation contains complete context information.

[0058] This embodiment innovatively implements a CNN classification model. A multi-channel convolutional structure is adopted, including convolutional kernels of 2, 3, and 4 words, and the number of each type of kernel is 128. The most significant features are extracted through the max-pooling layer, and then mapped to a predefined class space through the fully connected layer. The cross-entropy loss function is used for model training, and dropout is used to prevent overfitting.

[0059] This embodiment designs an accurate matching annotation mechanism. The classification results are matched with the specification library in multiple dimensions, including content type, expression method, and logical relationship. The matching process considers synonymous expressions and variant forms, and ensures the accuracy of the matching through semantic similarity calculation. For highly specialized technical descriptions, domain knowledge is used to assist in judgment.

[0060] This embodiment implements a dynamic construction method for the technical dictionary library. First, a basic dictionary is established through expert annotation, and then it is automatically expanded based on the conditional random field model. The feature template of the CRF model contains information such as word form, part of speech, and context, and the recognition accuracy is improved through iterative training. For newly emerging technical terms, the system can automatically include and classify them.

[0061] In this embodiment, a method for analyzing the distribution of feature words is innovatively designed. By constructing a co-occurrence matrix to capture the association patterns between technical feature words, the matrix elements represent the frequency of co-occurrence of feature words within a fixed window. The calculation of the association strength uses an improved cosine similarity: Sim(wi, wj) = cos(vi, vj)·log(1 + cij), where vi and vj are word vectors, and cij is the co-occurrence frequency.

[0062] This embodiment designs a special normalization processing strategy. Min-Max normalization is performed on the association strength, and weight adjustment is considered by taking into account the word frequency. The normalized score reflects the relative association degree between technical feature words, providing a reliable basis for subsequent document evaluation.

[0063] This embodiment implements a deep structural semantic analysis model. Using the Transformer architecture, it contains 6 encoder layers, with 8 attention heads set in each layer. The model can capture long-distance semantic dependencies and calculate the association strength between paragraphs through the self-attention mechanism. The attention scores are normalized by the softmax function to ensure the interpretability of the calculation results.

[0064] This embodiment designs a multi-level calculation method for semantic coherence evaluation. At the paragraph level, the semantic connection degree between adjacent paragraphs is calculated through attention weights. At the global level, a semantic coherence graph is constructed, and the overall structural integrity of the document is analyzed through the connectivity of the graph.

[0065] Through the above technological innovations, this embodiment effectively solves multiple key problems in the evaluation of project application documents: inaccurate judgment of structural standardization, incomplete extraction of technical features, inaccurate evaluation of semantic coherence, etc. In practical applications, this solution can accurately evaluate the quality of each dimension of the document, providing a clear direction for subsequent optimization. It is particularly suitable for the evaluation of application documents with strong professionalism such as technological innovation projects and R & D projects, significantly improving the accuracy and efficiency of the evaluation. The adaptive characteristics of this solution enable it to handle different types of project application requirements, with broad application prospects. Through in-depth analysis of the document structure and content, the system can provide guiding evaluation results to help enterprises improve the quality of application documents.

[0066] In an embodiment of the method for directional training of the enterprise project text processing model of the present application, it may further specifically include the following content: Step S301: Calculate the contribution degrees of the text coverage score, the technical association degree score, and the structural integrity score using the principal component analysis method, use the contribution degrees as weight coefficients, perform a linear weighted combination on the text coverage score, the technical association degree score, and the structural integrity score, and map the weighted combination result to a specified score interval based on the maximum-minimum normalization method to obtain a comprehensive score; Step S302: Establish a hierarchical screening matrix, sort the historical declaration documents in descending order according to the comprehensive score, calculate the distribution characteristics of the comprehensive score using the clustering analysis method, determine the screening threshold based on the distribution characteristics, use the historical declaration documents higher than the screening threshold as training samples, and perform text vectorization processing on the training samples to construct a training corpus.

[0067] Optionally, in this embodiment, a special multi-dimensional score analysis mechanism is first designed. For the three core evaluation dimensions of text coverage, technical relevance, and structural integrity, an improved principal component analysis method is used to calculate the contribution degree of each dimension. In the data preprocessing stage, the scores of the three dimensions are standardized to eliminate the influence of dimensional differences. The principal component analysis process calculates the variance contribution of each dimension through eigenvalue decomposition: PC = V·Λ·V^T, where V is the eigenvector matrix and Λ is the eigenvalue diagonal matrix.

[0068] This embodiment realizes an innovative weight allocation strategy. Based on the variance contribution rate obtained from the principal component analysis, an adaptive weight calculation method is designed. The determination of the weight coefficient not only considers the variance contribution of each dimension but also introduces an expert experience weight for correction. For different types of project declaration documents, the system will dynamically adjust the weight allocation ratio to ensure the rationality of the scoring.

[0069] This embodiment designs a multi-level weighted combination mechanism for score fusion. First, perform linear weighting: Score = w1·Coverage + w2·Relevance + w3·Structure, where w1, w2, and w3 are weight coefficients, corresponding to the contribution weights of text coverage, technical relevance, and structural integrity respectively. The weighting process considers the correlation between each score to avoid duplicate calculation of similar features.

[0070] This embodiment innovatively realizes a score normalization method. An improved maximum-minimum normalization algorithm is used, and a smoothing factor is introduced to prevent the influence of extreme values: Score_norm = (Score - min) / (max - min + ε), where ε is the smoothing factor. The normalization process dynamically adjusts the target score range to ensure the rationality and comparability of the score distribution.

[0071] This embodiment designs an efficient hierarchical screening matrix. The matrix structure includes multiple screening dimensions, and each dimension corresponds to a different evaluation angle. By sorting the historical declaration documents in descending order according to the comprehensive score, a hierarchical document evaluation system is constructed. The design of the screening matrix considers the continuous characteristics of document quality and avoids rigid classification boundaries.

[0072] This embodiment implements an innovative clustering analysis method. The improved K-means algorithm is used to cluster the comprehensive scores, and the number of clusters is dynamically determined by the silhouette coefficient. The algorithm takes into account the distribution density and dispersion degree of the scores, and determines the optimal clustering center through iterative optimization.

[0073] For threshold determination in this embodiment, an adaptive screening mechanism is designed. Based on the clustering results, the distribution characteristics of the scores are analyzed, including statistical indicators such as mean, variance, and skewness. The setting of the threshold comprehensively considers the overall distribution and quality requirements of the scores, and determines a reasonable screening standard in a data-driven manner.

[0074] This embodiment innovatively implements a training sample screening strategy. The documents above the threshold are screened in multiple rounds to ensure the quality and diversity of the samples. The screening process takes into account multiple dimensions such as the distribution of the technical fields of the documents, the degree of innovation, and the integrity, avoiding the concentration of samples in specific types.

[0075] This embodiment designs a comprehensive text vectorization processing scheme. A pre-trained language model is used to extract text features. The model architecture is improved based on BERT, and a domain adaptation layer is added to enhance the understanding ability of professional terms. The vectorization process retains the hierarchical structure information of the documents, facilitating subsequent model training.

[0076] This embodiment implements an efficient corpus construction mechanism. The selected training samples are processed structurally, and a multi-level index system is established. The organization of the corpus takes into account the relevance and reusability of knowledge, supporting flexible retrieval and update operations. Through regular quality assessment and cleaning, the effectiveness of the corpus is maintained.

[0077] Through the above technological innovations, this embodiment effectively solves several key problems in the evaluation and screening of project application documents, such as inconsistent scoring criteria, unstable sample quality, and insufficient feature extraction. In practical applications, this solution can accurately evaluate and screen high-quality application documents, providing a reliable data basis for model training. It is particularly suitable for processing application documents with strong professionalism such as technological innovation projects and industrialization projects, significantly improving the efficiency and quality of sample library construction. The adaptive characteristics of this solution enable it to handle different types of project application requirements, with broad application value. Through a systematic scoring fusion and screening mechanism, the representativeness and effectiveness of the training corpus are ensured, providing high-quality data support for subsequent model training.

[0078] In an embodiment of the method for directional training of the enterprise project text processing model of this application, the following content may also be specifically included: Step S401: performing word segmentation processing on the text content in the training corpus, extracting technical entity words from the word segmentation results using a named entity recognition model, constructing the technical entity words as knowledge graph nodes, extracting relationship attributes between the technical entity words using dependency syntactic analysis, adding attribute labels to the knowledge graph nodes based on a syntactic tree structure, and constructing an initial node set of the knowledge graph; Step S402: Use a text vectorization model to calculate the semantic similarity matrix between the nodes of the knowledge graph, normalize the semantic similarity matrix, use the normalized semantic similarity value as the weight of the knowledge association edge, establish knowledge association edges between the nodes of the knowledge graph based on the breadth-first search algorithm, and construct an edge set of the knowledge graph.

[0079] Optionally, this embodiment first designs a special word segmentation processing mechanism. In view of the characteristics of a large number of professional terms and technical terms in project application documents, an improved maximum matching algorithm is used for word segmentation. On the basis of the traditional word segmentation dictionary, a professional technical dictionary is integrated, which contains proper nouns, technical terms and standard vocabulary in various fields. The word segmentation process adopts a two-way scanning strategy to improve the recognition accuracy of complex terms.

[0080] This embodiment implements an innovative named entity recognition model. Using the BiLSTM-CRF architecture, the model input layer uses a dual-channel representation of character vectors and word vectors, and the dimension of each channel is 300. Pre-training is performed on technical document corpora to perform domain adaptation and improve the recognition ability of professional entities. The loss function of the model combines the annotation loss and the transition probability constraint: Loss = -log(P(y|x)) - λ·log(P(T)), where P(y|x) is the annotation probability, P(T) is the transition probability, and λ is the balance coefficient.

[0081] This embodiment designs a knowledge graph node construction strategy for entity transformation. The identified technical entities are classified according to the predefined ontology schema, including technical solutions, innovations, performance indicators and other categories. The node construction process considers the hierarchical relationship of entities and establishes a preliminary node organization structure through parent-child relationship and peer relationship.

[0082] This embodiment innovatively implements a dependency syntactic analysis method. A deep learning syntactic analyzer is used to encode sentence structures through a multi-layer bidirectional Transformer. The analyzer can identify dependency relationships between technical entities, such as part-whole, causal relationships, functional associations, etc. For complex long sentences, a segmentation processing strategy is adopted to ensure the accuracy of the analysis.

[0083] This embodiment designs a complete attribute annotation mechanism. Based on the syntactic tree structure, modifiers and qualifiers are extracted as node attributes. The attribute annotation adopts a multi-level structure, including basic attributes (such as time, quantity) and technical attributes (such as parameters, indicators). By combining rules and learning, the accuracy and integrity of attribute annotation are ensured.

[0084] This embodiment implements an efficient text vectorization model. The BERT model is used as the basic encoder and pre-trained through the masked language model and the next sentence prediction task. A domain adaptation layer is added to the output layer of the model to better capture the semantic features of technical texts. The vector representation of node texts is obtained through average pooling to get a vector with a fixed dimension.

[0085] This embodiment innovatively designs a similarity calculation method. An N×N semantic similarity matrix is constructed (N is the number of nodes), and the matrix elements are calculated by cosine similarity: sim(i,j) = vi·vj / (||vi||·||vj||), where vi and vj are node vectors. To improve the calculation efficiency, the locality-sensitive hashing technique is used to optimize the similarity calculation process.

[0086] This embodiment realizes an adaptive normalization process for the similarity matrix. An improved Softmax function is used for normalization, and a temperature parameter is introduced to adjust the smoothness of the distribution. The normalization process takes into account the importance weights of nodes and appropriately enhances the similarity values of key technical nodes.

[0087] This embodiment designs an intelligent edge construction strategy. Based on the breadth-first search algorithm, the connection relationships between nodes are explored, and the path length and cumulative similarity are considered during the search process. To prevent the graph from being too dense, a dynamic threshold is set to control the generation of edges, and the threshold value is adaptively adjusted according to the degree distribution of nodes.

[0088] This embodiment realizes an optimization mechanism for the edge set. The local structure of the graph is analyzed through the community discovery algorithm to identify tightly connected node groups. While maintaining key connections, redundant edges are pruned to improve the interpretability of the graph. For important technical paths, the connection stability is ensured through path enhancement.

[0089] Through the above technological innovations, this embodiment effectively solves several key problems in the construction of the knowledge graph for project application documents, such as inaccurate identification of technical entities, incomplete relationship extraction, and unreasonable graph structure. In practical applications, this solution can accurately extract and organize technical knowledge in project documents and construct a knowledge graph with a clear structure. It is particularly suitable for knowledge management in highly specialized fields such as technological innovation projects and R & D projects, significantly improving the efficiency and quality of knowledge organization. The adaptive characteristics of this solution enable it to process project documents in different fields and have broad application value. Through in-depth analysis and optimization of the knowledge structure, the system can provide reliable knowledge support for subsequent document generation.

[0090] In an embodiment of the method for directed training of the enterprise project text processing model of the present application, the following content may also be specifically included: Step S501: Input the knowledge graph into a graph convolutional network, represent the connection relationship of the knowledge association edges using an adjacency matrix, perform a convolution operation on the feature vectors of the knowledge graph nodes, process the convolution result through a non-linear activation function, iteratively update the parameters of the graph convolutional network based on the backpropagation algorithm, calculate the weight coefficients of the knowledge association edges, and clip the edge set of the knowledge graph according to the weight coefficients, retaining the knowledge association edges with weight coefficients higher than a preset threshold. Step S502: Construct a text structure analysis model. The text structure analysis model uses a hierarchical attention mechanism to analyze the structure of the adjusted knowledge graph, extracts the structure features of the knowledge graph based on a deep neural network, maps the structure features to text layout rules, generates a standardized text template according to the text layout rules, and performs project type annotation on the standardized text template to construct a differentiated text template library.

[0091] Optionally, this embodiment first designs a dedicated graph convolutional network processing mechanism. The network structure includes three graph convolutional layers, and the feature dimensions of each layer are 512, 256, and 128 respectively. The input layer receives the node feature matrix and the adjacency matrix of the knowledge graph, where the node features include representations in multiple dimensions such as technical descriptions and attribute information. The adjacency matrix is stored in the form of a sparse matrix to improve the calculation efficiency.

[0092] This embodiment implements an innovative graph convolution operation method. The convolution operation adopts dual aggregation in the spatial domain and the feature domain: H(l + 1) = σ(D^(-1 / 2)AD^(-1 / 2)H(l)W(l)), where A is the adjacency matrix, D is the degree matrix, H is the node feature, W is the learnable weight matrix, and σ is the non-linear activation function. This design can consider both the local structure features and the global semantic information of the nodes.

[0093] In this embodiment, an adaptive activation mechanism is designed for the selection of activation functions. The ReLU function is used in the shallow layer to maintain the sparsity of features, and the GELU function is adopted in the deep layer to provide smoother gradients. A batch normalization layer is added to the activation process to reduce internal covariate shift and improve the training stability of the model.

[0094] In this embodiment, a parameter update strategy is innovatively implemented. An improved backpropagation algorithm is adopted, and gradient clipping is introduced to prevent gradient explosion. The loss function design includes a structure preservation term and a feature reconstruction term, and the two objectives are balanced through weighted combination. The dynamic learning rate scheduling is adopted in the training process, and it is adaptively adjusted according to the performance of the validation set.

[0095] In this embodiment, an accurate edge weight calculation method is designed. The importance scores between node pairs are calculated through the attention mechanism, and the attention scores consider the feature similarity and structural correlation of nodes. The softmax function is used for the normalization of weight coefficients to ensure the numerical stability and comparability.

[0096] In this embodiment, an intelligent graph pruning mechanism is implemented. The threshold is set based on the global characteristics of the graph, including topological indexes such as connectivity and clustering coefficient. The important technical paths and innovation chains are maintained during the pruning process, while redundant and noisy connections are removed. For key nodes, the system will appropriately reduce the pruning criteria.

[0097] In this embodiment, a hierarchical attention model is designed for text structure analysis. The model contains multiple attention layers, and each layer is responsible for capturing structural information at different granularities. The multi-head mechanism is adopted for attention calculation, and each head focuses on different feature dimensions to improve the expression ability of the model.

[0098] In this embodiment, a structural feature extraction method is innovatively implemented. The deep neural network adopts a residual connection structure to avoid the problem of network degradation. The hierarchical relationship and logical dependence of nodes are considered during the feature extraction process, and multi-scale structural information is retained through skip connections.

[0099] In this embodiment, an efficient feature mapping mechanism is designed. The extracted structural features are converted into text layout rules, which include multiple dimensions such as chapter organization, paragraph arrangement, and content order. The rule learning and template matching are combined in the mapping process to ensure the operability of the generated rules.

[0100] In this embodiment, an innovative template generation strategy is implemented. A standardized text template is generated based on the layout rules, and the template includes a fixed structure and variable parts. The special requirements of different types of projects are considered during the generation process, and the applicability of the template is ensured through conditional constraints.

[0101] In this embodiment, a multi-dimensional classification system is designed for template annotation. The annotation dimensions include project type, technical field, degree of innovation, etc., and automatic annotation is achieved through a multi-label classification model. The annotation results are used to construct an indexing system for a differentiated template library.

[0102] Through the above technological innovations, this embodiment effectively solves several key problems in the optimization of the project application document structure: inaccurate knowledge association, incomplete structure extraction, poor template adaptability, etc. In practical applications, this solution can accurately analyze and optimize the knowledge graph structure and generate high-quality document 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 handle different types of project application requirements and have broad application value. Through a systematic structure analysis and template generation mechanism, the quality and applicability of the generated documents are ensured, providing strong technical support for enterprise project applications.

[0103] In an embodiment of the method for directional training of the enterprise project text processing model of this application, the following content may also be specifically included: Step S601: Obtain technical R & D data from the enterprise R & D management system, perform standardization processing on the technical R & D data using a data cleaning method, 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; Step S602: Reform 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, input the training corpus into the encoding layer for feature extraction, construct a decoding constraint matrix based on the differentiated text template library, calculate the error between the model prediction output and the training samples using a cross-entropy loss function, and optimize the parameters of the pre-trained large model through backpropagation algorithm to construct an enterprise project text generation model.

[0104] Optionally, this embodiment first designs a comprehensive technical R & D data collection mechanism. Connect to the enterprise R & D management system through an API interface to obtain multi-dimensional data including project information, patent data, technical reports, etc. in real time. A data cache pool is established during the collection process, and an incremental update strategy is used to ensure the timeliness of the data. Special parsing rules are designed for different types of R & D data to ensure the accuracy of data extraction.

[0105] This embodiment implements an innovative data cleaning method. Aiming at the characteristics of technology R & D data, multi-layer filtering rules are designed, including outlier detection, missing value processing, and format standardization. The outlier detection adopts an improved IQR method and sets dynamic thresholds in combination with domain knowledge. For missing data, strategies such as mean filling and nearest neighbor interpolation are adopted according to the data type to ensure data integrity.

[0106] This embodiment designs an adaptive encoding strategy for feature encoding. One-hot encoding is used for discrete features, and normalization is performed on continuous features. The feature encoding formula is: F = [Fd; Fc], where Fd is the discrete feature encoding and Fc is the continuous feature encoding. The importance and relevance of features are considered during the feature selection process, and redundant features are removed through LASSO regularization.

[0107] This embodiment innovatively implements an autoencoder model. A deep structure is adopted, and both the encoder and decoder contain three hidden layers with dimensions of 512, 256, and 128 in sequence. The activation function uses LeakyReLU to avoid gradient vanishing. The loss function combines reconstruction error and KL divergence: Loss = MSE(x,x') + λ·KL(z||N(0,I)), where x is the input feature, x' is the reconstructed feature, and z is the hidden layer representation.

[0108] This embodiment designs a feature fusion mechanism. The dimensionality-reduced technical features are concatenated with the enterprise basic information (such as scale, industry, qualifications, etc.). The attention mechanism is adopted during the fusion process to dynamically adjust the importance weights of different features and generate the final enterprise portrait vector.

[0109] This embodiment implements an innovative decoder layer structure for the pre-trained model transformation. Based on the original Transformer decoder layer, a conditional injection module is added, and the enterprise portrait vector is injected into each decoder layer after linear mapping. The gating mechanism is adopted during the injection process: h = g⊙tanh(Wx + Uc), where g is the gating unit, x is the original feature, and c is the conditional vector.

[0110] This embodiment designs an efficient feature extraction mechanism. The training corpus is input into the encoding layer, and text features are extracted through the multi-head self-attention mechanism. Each attention head independently learns feature representations from different angles, and then a comprehensive representation is obtained through feature fusion. The attention calculation considers positional encoding to better capture sequence information.

[0111] This embodiment innovatively implements a decoding constraint matrix. Constraint conditions are constructed based on a differentiated text template library, including structural constraints and content constraints. The constraint matrix reduces the computational complexity through sparse representation while maintaining the effectiveness of the constraints. The constraint strength is dynamically adjusted during the decoding process to ensure the normality of the generated text.

[0112] In this embodiment, an optimization strategy is designed for model training. An improved cross-entropy loss function is adopted, and label smoothing is introduced to reduce overfitting. The loss calculation takes into account the importance weights of different parts of the content, and higher weights are assigned to the descriptions of key technologies. The gradient norm is controlled by gradient clipping to improve the training stability.

[0113] This embodiment implements an efficient parameter optimization mechanism. The Adam optimizer is used in the backpropagation process, and the learning rate is dynamically adjusted using the cosine annealing strategy. Different learning rates are set for the parameters of different layers to adapt to new tasks while maintaining pre-trained knowledge. The training process is monitored through the performance of the validation set, and the early stopping strategy is adopted to avoid overfitting.

[0114] Through the above technological innovations, this embodiment effectively solves multiple key problems in enterprise project text generation, such as incomplete representation of enterprise features, poor model adaptability, and non-standard generated text. In practical applications, this solution can generate personalized project declaration documents according to the characteristics of enterprises, realizing the intelligence and precision of document generation. It is particularly suitable for the generation of declaration documents with strong professionalism, such as technological innovation projects and industrialization projects, significantly improving the efficiency and quality of document generation. The adaptive characteristics of this solution enable it to handle different types of enterprise requirements and have broad application value. Through in-depth understanding of enterprise features and targeted optimization of the model, the professionalism and pertinence of the generated documents are ensured.

[0115] In an embodiment of the method for targeted training of the enterprise project text processing model of this application, the following content may also be specifically included: Step S701: Input the test text into the enterprise project text generation model, use the beam search algorithm to decode and generate the test text, calculate the matching degree score between the generation result and the template in the differential text template library, construct the text reconstruction error based on the variational autoencoder, and use the weighted sum of the matching degree score and the text reconstruction error as the model evaluation index to update the network parameters of the enterprise project text generation model by gradient; Step S702: Construct a sample pool dynamic update mechanism, add the generation results higher than the evaluation threshold to the training corpus, extract new text structure features from the generation results based on the text clustering method, integrate the text structure features into the differential text template library, and perform online fine-tuning on the enterprise project text generation model using the incremental learning strategy to update the knowledge representation of the model.

[0116] Optionally, in this embodiment, a dedicated test text processing mechanism is first designed. The input test text is preprocessed, including format standardization, professional term recognition, and semantic unit division. The test text is converted into a hidden layer representation by an encoder, with the dimension set to 768, which contains the semantic and structural information of the text. The preprocessing process pays special attention to the retention of key contents such as technical descriptions and innovation points.

[0117] This embodiment implements an innovative beam search strategy. The search width is set to 8, and the optimal candidate sequence is retained at each decoding step. The scoring function of the candidate sequence comprehensively considers the language model probability and the accuracy of professional terms: Score = α·log P(y|x) + β·Term_accuracy, where P(y|x) is the conditional generation probability, Term_accuracy is the accuracy of the use of professional terms, and α and β are balance coefficients.

[0118] For text matching, this embodiment designs a multi-dimensional evaluation mechanism. The similarity between the generated text and each template in the template library is calculated through a vectorization model. The similarity calculation uses an improved cosine metric, introducing position weighting. The matching process considers the hierarchical features of the document structure and assigns higher weights to key parts.

[0119] This embodiment innovatively implements a variational autoencoder. Both the encoder and the decoder adopt the Transformer architecture, and the KL divergence constraint is introduced in the latent space. The calculation of the reconstruction error integrates content consistency and structure preservation: Loss = MSE(x,x') + λ·KL(q(z|x)||p(z)), where x and x' are the original text and the reconstructed text respectively, and z is the latent variable.

[0120] This embodiment designs an accurate evaluation index calculation method. The matching score and the reconstruction error are combined through dynamic weights: Score_final = w1·Match_score + w2·Recon_error, where the weight coefficients are adaptively adjusted according to the performance of the validation set. The evaluation process pays special attention to the accuracy of technical content and the integrity of the document structure.

[0121] This embodiment implements an efficient parameter update strategy. The Adam optimizer is used for gradient update, and the learning rate is dynamically adjusted using the cosine annealing strategy. A gradient accumulation mechanism is introduced in the update process to improve the stability of training. For the generation bias in a specific domain, directional optimization is performed through a domain adaptation layer.

[0122] This embodiment designs a dynamic update mechanism for sample pool management. A multi-level cache structure is established, including a temporary pool and a permanent pool. The evaluation threshold is set based on the quality distribution of historical samples and dynamically adjusted in an adaptive manner. The screening of high-quality samples takes into account the two dimensions of innovation and typicality.

[0123] This embodiment innovatively implements a text clustering method. A hierarchical clustering algorithm is used to calculate similarity by combining TF-IDF features and deep semantic features. The clustering process takes into account the topic relevance and structural similarity of the text, and controls the clustering granularity through a dynamic threshold.

[0124] This embodiment designs an innovative feature extraction strategy. New text structure features are extracted from the clustering results, and the features include local organization patterns and global framework features. The extraction process uses an attention mechanism to identify key structural elements to ensure the representativeness and reusability of the features.

[0125] This embodiment implements the template library integration mechanism. The newly extracted structural features are merged with the existing templates, and the fusion process adopts a progressive update strategy. The similarity analysis is used to avoid duplicate templates while maintaining the diversity of the template library. The integrated templates are automatically annotated and the index system is updated.

[0126] This embodiment designs an incremental learning strategy for model fine-tuning. The knowledge distillation technology is used to maintain the original knowledge while learning the characteristic patterns in new samples. The fine-tuning process controls the learning rate within a small range to avoid catastrophic forgetting. The model performance is evaluated regularly and the learning strategy is adjusted dynamically according to the effect.

[0127] Through the above technical innovations, this embodiment effectively solves several key problems in the generation of project application documents: unstable generation quality, poor model adaptability, untimely knowledge update, etc. In practical applications, the solution can continuously optimize the document generation effect and maintain the advancement of the system through dynamic learning and updating mechanisms. It is particularly suitable for the generation of highly professional application documents such as technological innovation projects and industrialization projects, and significantly improves the level of intelligent document generation. The adaptive characteristics of this solution enable it to continuously adapt to new project types and changes in requirements, and it has a wide range of application value. Through a systematic evaluation and update mechanism, the continuous optimization of the generation model and the effective accumulation of knowledge are ensured, providing reliable technical support for enterprise project applications.

[0128] In order to effectively solve the deficiencies of traditional technologies in training sample screening, knowledge structure analysis and model customization, and significantly improve the practicality and adaptability of the project text processing model, the present application provides an embodiment of an enterprise project text processing model directional training device for implementing all or part of the content of the enterprise project text processing model directional training method, see Figure 2, the enterprise project text processing model orientation training device specifically includes the following contents: The project analysis module 10 is used to construct a project text scoring model. Input the enterprise historical declaration documents into the project text scoring model. Use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; The content processing module 20 is used to construct knowledge graph nodes from the text content in the training corpus, construct knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, 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; The model training module 30 is used to collect enterprise multi-dimensional data to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for orientation training to construct an enterprise project text generation model, input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

[0129] As can be seen from the above description, the enterprise project text processing model orientation training device provided by the embodiments of the present application can screen high-quality training samples through multi-dimensional evaluation of text coverage, technical association, and structural integrity by constructing a project text scoring model. Analyze the text structure relationship based on the knowledge graph and the graph convolutional network to construct a differential text template library. The system performs orientation training on the pre-trained large model in combination with the enterprise portrait vector, and continuously optimizes the model parameters and the knowledge base using an incremental learning strategy. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model.

[0130] At the hardware level, in order to effectively address the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly enhance the practicality and adaptability of the project text processing model, this application provides an embodiment of an electronic device for implementing all or part of the content in the method for targeted training of the enterprise project text processing model. The electronic device specifically includes the following: A processor, a memory, a communications interface, and a bus; wherein, the processor, the memory, and the communications interface complete mutual communication through the bus; the communications interface is used to implement information transmission between the device for targeted training of the enterprise project text processing model and related devices such as the core business system, the user terminal, and the relevant database, etc. 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 method for targeted training of the enterprise project text processing model and the embodiments of the device for targeted training of the enterprise project text processing model, and the content is incorporated herein, and the repeated parts will not be elaborated.

[0131] 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.

[0132] In practical applications, part of the method for targeted training of the enterprise project text processing model can be executed on the electronic device side as described above, or all operations can be completed in 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, etc. This application does not make any limitations in this regard. If all operations are completed in the client device, the client device may further include a processor.

[0133] The above-mentioned client device may have a communication module (i.e., a communication unit), and can be communicatively connected to a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side, and may also include a server on an intermediate platform in other implementation scenarios, such as a server on a third-party server platform communicatively linked to the task scheduling center server. The server may include a single computer device, or may include a server cluster composed of multiple servers, or a server structure of a distributed device.

[0134] Figure 3 This is a schematic block diagram of the system composition of the electronic device 9600 according to the embodiment of this application. AsFigure 3 As 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 may also be used to supplement or replace this structure to achieve telecommunication functions or other functions.

[0135] In one embodiment, the function of the enterprise project text processing model directed training method may be integrated into the central processing unit 9100. Among them, the central processing unit 9100 may be configured to perform the following controls: Step S101: Construct a project text scoring model, input the enterprise historical declaration documents into the project text scoring model, use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical correlation score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical correlation score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; Step S102: Construct knowledge graph nodes from the text content in the training corpus, construct knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, 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: Collect enterprise multi-dimensional data to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model, input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

[0136] As can be seen from the above description, the electronic device provided by the embodiments of the present application constructs a project text scoring model, screens high-quality training samples through multi-dimensional evaluation of text coverage, technical relevance, and structural integrity. Analyzes the text structure relationship based on the knowledge graph and graph convolutional network, and constructs a differentiated text template library. The system conducts targeted training on the pre-trained large model in combination with the enterprise portrait vector, and continuously optimizes the model parameters and knowledge base using the incremental learning strategy. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model.

[0137] In another embodiment, the device for targeted training of the enterprise project text processing model can be separately configured from the central processing unit 9100. For example, the device for targeted training of the enterprise project text processing model can be configured as a chip connected to the central processing unit 9100, and the functions of the method for targeted training of the enterprise project text processing model are realized through the control of the central processing unit.

[0138] 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

[0139] components not shown in Figure 3 ; reference may be made to the prior art.

[0140] 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 program stored in the memory 9140 to implement information storage or processing, etc.

[0141] 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.

[0142] 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 that is used to store application programs and function programs or the processes for operating the electronic device 9600 by the central processing unit 9100.

[0143] The memory 9140 can also include a data storage unit 9143 that 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 performing other functions of the electronic device (such as a messaging application, an address book application, etc.).

[0144] 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.

[0145] Based on different communication technologies, multiple communication modules 9110 can be provided in the same electronic device, 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 the speaker 9131 and the microphone 9132 via the audio processor 9130 to provide an audio output via the speaker 9131 and receive an audio input from the microphone 9132, thereby implementing normal telecommunication functions. The audio processor 9130 can include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is also coupled to the central processing unit 9100, so that recording can be performed on the local machine via the microphone 9132 and the sound stored on the local machine can be played via the speaker 9131.

[0146] Embodiments of the present application also provide a computer-readable storage medium capable of implementing all steps in the enterprise project text processing model directed training method where the execution subject in the above embodiments is a server or a client. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements all steps of the enterprise project text processing model directed training method where the execution subject in the above embodiments is a server or a client. For example, when the processor executes the computer program, the following steps are implemented: Step S101: Construct a project text scoring model. Input the enterprise historical declaration documents into the project text scoring model. Use a text classifier to identify the normalized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; Step S102: Construct knowledge graph nodes from the text content in the training corpus. Based on the citation relationships and semantic similarities between the knowledge graph nodes, construct knowledge association edges. 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: Collect enterprise multi-dimensional data to construct an enterprise portrait vector. Input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model. Input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and use an incremental learning strategy to continuously update the training corpus and the differential text template library.

[0147] As can be seen from the above description, the computer-readable storage medium provided by the embodiments of the present application screens high-quality training samples through multi-dimensional evaluation of text coverage, technical association, and structural integrity by constructing a project text scoring model. Analyze the text structure relationship based on the knowledge graph and the graph convolutional network to construct a differential text template library. The system performs directed training on the pre-trained large model in combination with the enterprise portrait vector, and uses an incremental learning strategy to continuously optimize the model parameters and the knowledge base. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model.

[0148] An embodiment of the present application also provides a computer program product that can implement all steps in the enterprise project text processing model directed training method where the execution subject in the above embodiments is a server or a client. When the computer program / instructions are executed by a processor, the steps of the enterprise project text processing model directed training method are implemented. For example, the computer program / instructions implement the following steps: Step S101: Construct a project text scoring model. Input the enterprise historical declaration documents into the project text scoring model. Use a text classifier to identify the normalized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; Step S102: Construct knowledge graph nodes from the text content in the training corpus. Based on the citation relationships and semantic similarities between the knowledge graph nodes, construct knowledge association edges. 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: Collect enterprise multi-dimensional data to construct an enterprise portrait vector. Input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directed training to construct an enterprise project text generation model. Input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and use an incremental learning strategy to continuously update the training corpus and the differential text template library.

[0149] As can be seen from the above description, the computer program product provided by the embodiment of the present application screens high-quality training samples through multi-dimensional evaluation of text coverage, technical association, and structural integrity by constructing a project text scoring model. Analyze the text structure relationship based on the knowledge graph and the graph convolutional network to construct a differential text template library. The system performs directed training on the pre-trained large model in combination with the enterprise portrait vector, and uses an incremental learning strategy to continuously optimize the model parameters and the knowledge base. This method effectively solves the deficiencies of traditional technologies in training sample screening, knowledge structure analysis, and model personalized customization, and significantly improves the practicability and adaptability of the project text processing model.

[0150] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, an apparatus, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete 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 storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0151] 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 flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0152] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0154] Specific embodiments are applied in the present invention to elaborate on 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 to the present invention.

Claims

1. A method for directed training of an enterprise project text processing model, characterized in that, The method includes: Construct a project text scoring model, input the enterprise's historical declaration documents into the project text scoring model, use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; Construct the text content in the training corpus into knowledge graph nodes, construct knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, 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; Collect multi-dimensional data of the enterprise to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for targeted training to construct an enterprise project text generation model, input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

2. The method for directed training of the enterprise project text processing model according to claim 1, characterized in that, The constructing of the project text scoring model, inputting the enterprise's historical declaration documents into the project text scoring model, using a text classifier to identify the standardized description segments in the historical declaration documents, calculating the text coverage score, extracting the key technical feature words in the historical declaration documents, and calculating the technical association score based on the feature word co-occurrence matrix and performing structural semantic analysis on the historical declaration documents to obtain the structural integrity score includes: Establish a text structure specification library, segment the enterprise's historical declaration documents into text segments by paragraph, the text classifier constructs text feature vectors based on a bidirectional long short-term memory network, uses a convolutional neural network to classify and identify the text feature vectors, matches and labels the recognition results with the text structure specification library, and calculates the proportion of the labeled segments in the overall document to obtain the text coverage score; Construct a technical dictionary library, use a conditional random field model to identify technical feature words from the historical declaration documents, construct the distribution positions of the technical feature words in the document into a feature word co-occurrence matrix, calculate the association strength between the technical feature words in the feature word co-occurrence matrix based on the cosine similarity between word vectors, perform normalization processing on the association strength to obtain the technical association score, use a deep neural network to perform structural semantic analysis on the historical declaration documents, and calculate the semantic coherence between document paragraphs based on the attention mechanism to obtain the structural integrity score.

3. The method for directed training of the enterprise project text processing model according to claim 1, wherein, The weighted fusion of the text coverage score, the technical relevance score, and the structural integrity score to obtain a comprehensive score, and screening high-score samples based on the comprehensive score to construct a training corpus includes: Using the principal component analysis method to calculate the contribution degrees of the text coverage score, the technical relevance score, and the structural integrity score, taking the contribution degrees as weight coefficients, performing a linear weighted combination on the text coverage score, the technical relevance score, and the structural integrity score, and mapping the weighted combination result to a specified score interval based on the maximum-minimum normalization method to obtain a comprehensive score; Establishing a hierarchical screening matrix, sorting the historical application documents in descending order according to the comprehensive score, using the clustering analysis method to calculate the distribution characteristics of the comprehensive score, determining a screening threshold based on the distribution characteristics, taking the historical application documents higher than the screening threshold as training samples, and performing text vectorization processing on the training samples to construct a training corpus.

4. The method for directional training of the enterprise project text processing model according to claim 1, characterized in that Constructing the text content in the training corpus into knowledge graph nodes, and constructing knowledge association edges based on the citation relationships and semantic similarities between the knowledge graph nodes, includes: Performing word segmentation processing on the text content in the training corpus, using a named entity recognition model to extract technical entity words from the word segmentation results, constructing the technical entity words into knowledge graph nodes, using dependency syntax analysis to extract the relationship attributes between the technical entity words, and adding attribute labels to the knowledge graph nodes based on the syntax tree structure to construct an initial node set of the knowledge graph; Using a text vectorization model to calculate the semantic similarity matrix between the knowledge graph nodes, normalizing the semantic similarity matrix, taking the normalized semantic similarity value as the weight of the knowledge association edge, and establishing knowledge association edges between the knowledge graph nodes based on the breadth-first search algorithm to construct an edge set of the knowledge graph.

5. The method for directional training of the enterprise project text processing model according to claim 1, wherein 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 construct a differential text template library, includes: Inputting the knowledge graph into a graph convolutional network, using an adjacency matrix to represent the connection relationships of the knowledge association edges, performing a convolution operation on the feature vectors of the knowledge graph nodes, processing the convolution result through a non-linear activation function, iteratively updating the parameters of the graph convolutional network based on the backpropagation algorithm, calculating the weight coefficients of the knowledge association edges, and pruning the edge set of the knowledge graph according to the weight coefficients, retaining the knowledge association edges with weight coefficients higher than a preset threshold; Constructing a text structure analysis model, the text structure analysis model using a hierarchical attention mechanism to perform structure analysis on the adjusted knowledge graph, extracting the structure features of the knowledge graph based on a deep neural network, mapping the structure features to text layout rules, generating a standardized text template according to the text layout rules, and performing item type annotation on the standardized text template to construct a differential text template library.

6. The method for directional training of the enterprise project text processing model according to claim 1, characterized in that Collect multi-dimensional data of the enterprise to construct an enterprise portrait vector, and input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for directional training to construct an enterprise project text generation model, including: Obtain technical R & D data from the enterprise R & D management system, perform standardization processing on the technical R & D data by using a data cleaning method, 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 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, input the training corpus into the encoding layer for feature extraction, construct a decoding constraint matrix based on the differential text template library, calculate the error between the model prediction output and the training samples by using a cross-entropy loss function, and optimize the parameters of the pre-trained large model through a backpropagation algorithm to construct an enterprise project text generation model.

7. The method for directed training of the enterprise project text processing model according to claim 1, wherein Input the test text into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library by using an incremental learning strategy, including: Input the test text into the enterprise project text generation model, perform decoding generation on the test text by using a beam search algorithm, calculate the matching degree score between the generation result and the templates in the differential text template library, construct a text reconstruction error based on a variational autoencoder, and use the weighted sum of the matching degree score and the text reconstruction error as a model evaluation index to perform gradient update on the network parameters of the enterprise project text generation model; Construct a sample pool dynamic update mechanism, add the generation results higher than the evaluation threshold to the training corpus, extract new text structure features from the generation results based on a text clustering method, integrate the text structure features into the differential text template library, and perform online fine-tuning on the enterprise project text generation model by using an incremental learning strategy to update the knowledge representation of the model.

8. An enterprise project text processing model orientation training device, characterized in that, The device includes: A project analysis module, which is used to construct a project text scoring model, input the enterprise historical declaration documents into the project text scoring model, use a text classifier to identify the standardized description segments in the historical declaration documents, calculate the text coverage score, extract the key technical feature words in the historical declaration documents, calculate the technical association score based on the feature word co-occurrence matrix, perform structural semantic analysis on the historical declaration documents to obtain the structural integrity score, perform weighted fusion on the text coverage score, the technical association score, and the structural integrity score to obtain a comprehensive score, and screen high-score samples based on the comprehensive score to construct a training corpus; A content processing module, which is used to construct knowledge graph nodes from the text content in the training corpus, construct knowledge association edges based on the reference relationships and semantic similarities between the knowledge graph nodes, calculate the weight coefficients of the knowledge association edges 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 model training module, which is used to collect multi-dimensional enterprise data to construct an enterprise portrait vector, input the enterprise portrait vector, the training corpus, and the differential text template library into a pre-trained large model for targeted training to construct an enterprise project text generation model, input test texts into the enterprise project text generation model for text generation testing, dynamically optimize the network parameters of the enterprise project text generation model based on the test results, and continuously update the training corpus and the differential text template library using an incremental learning strategy.

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 method for targeted training of the enterprise project text processing model 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 method for targeted training of the enterprise project text processing model according to any one of claims 1 to 7.

Images