EFFICIENT ANALYSIS OF SENTIMENTS IN TWITTER DATA BY USING STACKED BI-DIRECTIONAL LSTM

Abstract

With increase in social media popularity and different platforms permitting people to show opinion on diverse subjects, Sentiment Analysis (SA) as well as opinion mining is becoming a subject which attracts attention of researchers. SA has gained much popularity amid people with varying interests as well as motivations. In this paper, Sentiment140 dataset with tweets extracted using Twitter API is used. Pre-processing is performed using tokenisation by stemming and lemmatization. Extraction of features is carried out using Term Frequency-Inverse Document Frequency (TF-IDF), Word to Vector (Word2Vec) and word embedding using BERT. Tweets are categorizedinto positive and negative using Long Short-Term Memory (LSTM), Bi-directional LSTM (Bi-LSTM) and proposed Stacked Bi-LSTM (SBi-LSTM)model.SBi-LSTM offers better performance due to its modest structure with stacked LSTM layers. SBi-LSTM offers improved results based onAccuracy, Recall, Precision and F1-Score.