Tweet Sentiment Extraction

EasyChair Preprint 7283

Abstract

The classic sentiment analysis problem deals with analysing the overall polarity of a set of responses. Though by only knowing the polarity, an organisation can’t get an idea about why they received such responses. Thus this makes them unable to analyse the responses, which could have possibly helped them in the betterment of the service they were providing. The purpose of this project was to make a question answering model which would extract a phrase out of a given tweet which amplifies a given sentiment (positive/negative/neutral). Using initial training and testing runs which were scored using Jaccard score, we compared the performance between BERT (standard method), RoBERTa, DistilBERT and AlBERT to find out the best performing method on the given Twitter dataset. After, DistilBERT was found out to give the best performance out of the above mentioned methods with an accuracy of 68.92% over BERT’s accuracy of 64.57%, was then further fine-tuned by pre-processing, processing, and post-processing methods to make a final model which gave an accuracy of 73.12%. The project successfully implements a model which can extract phrases out of a given text (in this case tweets), the current accuracy benchmark for which is 73.12%. Further optimization is required to increase the accuracy even more, so that it can replicate BERT’s performance of 85% accuracy which it achieved on the SQUAD dataset.

Keyphrases: BERT, DistilBERT, Extract a phrase, Jaccard score, NLP, Question Answering, Sentiment Analysis

@booklet{EasyChair:7283,
  author    = {Chiranjeev Mishra and Eshit Bansal and A. Helen Victoria},
  title     = {Tweet Sentiment Extraction},
  howpublished = {EasyChair Preprint 7283},
  year      = {EasyChair, 2022}}