Research On Text Classification Algorithm Based On Word Embedding Model

Text classification task is an important basic research work in natural language processing tasks,which provides basic support and solutions for many language processing related tasks.It represents a fundamental research endeavor in the filed of natural language processing,providing foundational support and solutions for numerous language processing related tasks.However,existing text classification algorithms mainly face three issues that require further research and improvement.The first is how to simplify the model to reduce the computational resources required for text classification tasks.The second is how to effectively construct a model to represent text,thereby further improving classification performance.The third is how to extract important information from the corpus,learn a better word representation vector,and lay a better foundation for downstream tasks.In recent years,convolutional neural networks and graph neural networks have demonstrated significant advantages in data representation and feature extraction,but limited research has applied them to text classification tasks.This thesis applies these methods to text classification tasks to attempt to solve the aforementioned problems:(1)To address the first problem,a simplified convolutional neural network model is proposed in this thesis.The thesis proposes to accelerate the model’s training speed by pretraining on the dataset and simplifying the network structure.The proposed model is compared to several benchmark methods on multiple datasets of different sizes.The results show that the proposed model can significantly reduce training time while maintaining the classification performance,thus further reducing the computational requirements of the model.It also demonstrates that for text classification tasks,a simple work structure can achieve good classification results.(2)To address the second problem,a hierarchical graph neural network model is proposed.The thesis proposes a method of hierarchical representation of input samples to incorporate hierarchical structure into the graph modeling of text,transforming the originally insufficient flat graph structure.Different node aggregation methods are used to merge low-level graph nodes into high-level graph node features,thus extracting semantic information of different levels in the text.The proposed hierarchical graph model is compared to several benchmark methods on multiple mainstream datasets,and the experimental results show that the proposed model effectively improves text classification performance and enhances the model’s ability to extract text features.(3)To address the third problem,a graph attention network-based word embedding model is proposed.The thesis proposes to change the original training task of constructing word pairs to constructing virtual words and the next word,enabling the word embedding model to consider more contextual information,to some extent,addressing the limitation of the original word embedding model that only considers the relationship between words in the corpus.Additionally,a graph attention network is used to model known word sequences,adaptively assigning different weights to word nodes to better form feature vectors of virtual words.The obtained word embedding vectors are used as inputs to the classification algorithm,and the proposed models is compared to several benchmark methods on multiple datasets.The experimental results show that the proposed model can learn good word embedding vectors representations even in smaller corpora,and the improved word embedding model not only enables the classification algorithm to achieve better classification performance,but also accelerates the model training by faster convergence during training.