Comparative Analysis of Various Language Models on Sentiment Analysis for Retail

Understanding what people think about a product/service/movie, etc. has become a crucial part of business as it helps shape businesses and prosper at an exponential rate because no matter what they produce, the success of anything they make entirely depends on its customers/end users as to whether they appreciate it or not. This very task of mining the hidden sentiment within customer reviews is termed sentiment analysis. In this study, we put forward a system that showcases a novel approach for optimizing the analysis of sentiment in textual data using divergent combinations of the miscellaneous hyperparameters obtainable with various Language Models. We experiment on the performance of the Language Models, ALBERT, RoBERTa, and BERT, by varying the method of optimization for a multiclass classification task by pre-processing the customer reviews which gave us our finest model achieving futuristic results. We conducted multiple trials of various Language Models using customer reviews of the Google Play app by varying the optimizer used. We found out that AdamW and Adabound give us the highest accuracies among the three experimented upon.