Comparative Analysis of Different Transformer Based Architectures Used in Sentiment Analysis

In the era of research where topics like Sentiment Analysis, NLP (natural language processing), Artificial Intelli- gence, Transfer learning are buzz words. Sentiment Analysis is a field which have proved its importance in day to day activities which could be related to monetary analysis, person's emotional count etc. With the introduction of Transformer based language models the field of NLP has got a huge attraction of researchers as well as practitioners. [16] [9] [6]. Using transfer learning with these models, have proved to give exceptional accuracy. Some of the State-of-the-art model includes BERT [16], DistilBERT, XLNet and T5. In this paper we have investigated the classification power of sentiments from pre-trained language models, e.g., BERT, XLnet. Specifically, we have simulated and presented number of experimental results which shows what amount of accuracy could be attained with different models. We have used most popular imdb-reviews dataset for doing analysis of the models