Leveraging Graph Neural Networks for Text Classification with Semantic and Structural Insights

: Applications which involve text classification may still need a breakthrough in capturing the latent structure in the text and more complex dependencies which limits its capacity to make correct predictions. This paper presents a new approach to a text classification application in which a hybrid graph representation learning algorithm has been used to demonstrate interactions between latent semantic and structural data in text documents. Text is represented as a graph, where a node represents a sentence and an edge represents the semantic relationship between two nodes. With nodes converted to embeddings generated through Sentence-BERT, it offers contextualized representations for every node. Along with this framework, we also learn low-dimensional representations of the text graphs using graph auto-encoders. Our model thus enhances generalization and has a powerful representation for downstream tasks by minimizing the difference between reconstructed and input graphs. Experimental results demonstrate that our model surpasses traditional meth-ods by successfully integrating semantic and structural information to enhance classification accuracy. This work contributes to the advancement of GNN-based architectures for text retrieval, demonstrating the potential of graphs in natural language processing.