This paper proposes a Text Graph Convolutional Network (Text GCN) for text classification. The method constructs a heterogeneous graph from a corpus, where words and documents are nodes, and edges represent word co-occurrence and document-word relationships. The graph is then processed using a Graph Convolutional Network (GCN) to learn embeddings for both words and documents. The Text GCN is initialized with one-hot representations for words and documents and jointly learns embeddings that are supervised by document class labels. Experimental results on multiple benchmark datasets show that Text GCN outperforms state-of-the-art methods for text classification without using pre-trained word embeddings or external knowledge. The method also learns predictive word and document embeddings. The performance of Text GCN improves as the percentage of training data decreases, indicating its robustness to limited training data. The method is effective in capturing global word co-occurrence information and can learn discriminative document and word embeddings. The Text GCN is tested on several benchmark datasets, including 20NG, Ohsumed, R52, R8, and MR. It achieves strong classification performance and outperforms other methods on these datasets. The method is also effective in learning document embeddings that can be visualized using t-SNE. The results show that Text GCN can capture both document-word and global word-word relationships, and that the GCN model can propagate label information to the entire graph. However, Text GCN does not outperform CNN and LSTM-based models on short text datasets like MR, as these models explicitly model consecutive word sequences. The method is limited by the transductive nature of GCN, which includes test documents in the training process. Future work includes improving classification performance using attention mechanisms and developing unsupervised text GCN frameworks for large-scale unlabeled text data.This paper proposes a Text Graph Convolutional Network (Text GCN) for text classification. The method constructs a heterogeneous graph from a corpus, where words and documents are nodes, and edges represent word co-occurrence and document-word relationships. The graph is then processed using a Graph Convolutional Network (GCN) to learn embeddings for both words and documents. The Text GCN is initialized with one-hot representations for words and documents and jointly learns embeddings that are supervised by document class labels. Experimental results on multiple benchmark datasets show that Text GCN outperforms state-of-the-art methods for text classification without using pre-trained word embeddings or external knowledge. The method also learns predictive word and document embeddings. The performance of Text GCN improves as the percentage of training data decreases, indicating its robustness to limited training data. The method is effective in capturing global word co-occurrence information and can learn discriminative document and word embeddings. The Text GCN is tested on several benchmark datasets, including 20NG, Ohsumed, R52, R8, and MR. It achieves strong classification performance and outperforms other methods on these datasets. The method is also effective in learning document embeddings that can be visualized using t-SNE. The results show that Text GCN can capture both document-word and global word-word relationships, and that the GCN model can propagate label information to the entire graph. However, Text GCN does not outperform CNN and LSTM-based models on short text datasets like MR, as these models explicitly model consecutive word sequences. The method is limited by the transductive nature of GCN, which includes test documents in the training process. Future work includes improving classification performance using attention mechanisms and developing unsupervised text GCN frameworks for large-scale unlabeled text data.