BERT-Based Hybrid RNN Model for Multi-class Text Classification to Study the Effect of Pre-trained Word Embeddings
S Shreyashree, Pramod Sunagar, S Rajarajeswari, Anita Kanavalli
Abstract
Due to the Covid-19 pandemic which started in the year 2020, many nations had imposed lockdown to curb the spread of this virus. People have been sharing their experiences and perspectives on social media on the lockdown situation. This has given rise to increased number of tweets or posts on social media. Multi-class text classification, a method of classifying a text into one of the pre-defined categories, is one of the effective ways to analyze such data that is implemented in this paper. A Covid-19 dataset is used in this work consisting of fifteen pre-defined categories. This paper presents a multi-layered hybrid model, LSTM followed by GRU, to integrate the benefits of both the techniques. The advantages of word embeddings techniques like GloVe and BERT have been implemented and found that, for three epochs, the transfer learning based pre-trained BERT-hybrid model performs one percent better than GloVe-hybrid model but the state-of-the-art, fine-tuned BERT-base model outperforms the BERT-hybrid model by three percent, in terms of validation loss. It is expected that, over a larger number of epochs, the hybrid model might outperform the fine-tuned model.