Sentiment analysis using deep learning

Abstract

This project seeks to identify the effectiveness of recurrent neural network architectures in identifying sentiment in tweets and classifying emotion in sentences from children’s stories. In a world that increasingly driven by data, possessing the ability to identify emotion in textual data is extremely powerful.

Long Short Term Memory Networks (LSTMs) were purpose built in order to remember information for long periods of time, and therefore work extremely well in evaluating the entire context of the sentence. Through extensive experimentation, we find that Bidirectional LSTMs(BLSTM) offer us the best performance in analysing tweets and predicting emotion in children’s stories. BLSTMs can model the important information about the underlying sentiment and emotion in a text input because they are able to use the left and right context of a sequence of words or phrases.

The experimental results obtained on the SemEval 2016 Task 4 dev-test dataset demonstrates that a merged BLSTM model using Word2vec and GloVe embeddings outperforms all the other models. The experimental results obtained on the test dataset of the children’s story book’s HighAgree sub corpus clearly demonstrate that BLSTM with pre-trained GloVe vectors works better than all of the other models.