Multi-Task Text Classification using Graph Convolutional Networks for Large-Scale Low Resource Language

  • 2022-05-02 21:44:12
  • Mounika Marreddy, Subba Reddy Oota, Lakshmi Sireesha Vakada, Venkata Charan Chinni, Radhika Mamidi
  • 1


Graph Convolutional Networks (GCN) have achieved state-of-art results onsingle text classification tasks like sentiment analysis, emotion detection,etc. However, the performance is achieved by testing and reporting onresource-rich languages like English. Applying GCN for multi-task textclassification is an unexplored area. Moreover, training a GCN or adopting anEnglish GCN for Indian languages is often limited by data availability, richmorphological variation, syntax, and semantic differences. In this paper, westudy the use of GCN for the Telugu language in single and multi-task settingsfor four natural language processing (NLP) tasks, viz. sentiment analysis (SA),emotion identification (EI), hate-speech (HS), and sarcasm detection (SAR). Inorder to evaluate the performance of GCN with one of the Indian languages,Telugu, we analyze the GCN based models with extensive experiments on fourdownstream tasks. In addition, we created an annotated Telugu dataset, TEL-NLP,for the four NLP tasks. Further, we propose a supervised graph reconstructionmethod, Multi-Task Text GCN (MT-Text GCN) on the Telugu that leverages tosimultaneously (i) learn the low-dimensional word and sentence graph embeddingsfrom word-sentence graph reconstruction using graph autoencoder (GAE) and (ii)perform multi-task text classification using these latent sentence graphembeddings. We argue that our proposed MT-Text GCN achieves significantimprovements on TEL-NLP over existing Telugu pretrained word embeddings, andmultilingual pretrained Transformer models: mBERT, and XLM-R. On TEL-NLP, weachieve a high F1-score for four NLP tasks: SA (0.84), EI (0.55), HS (0.83) andSAR (0.66). Finally, we show our model's quantitative and qualitative analysison the four NLP tasks in Telugu.


Quick Read (beta)

loading the full paper ...