Sentiment Analysis of Tweets Using Bi-Directional Long Short-Term Memory and Word Embeddings

Social media users have grown significantly over the last few years. This surge in the number of users has been ascribed to the increased frequency of attacks on immigrants and other minority groups on various social media platforms. Due to the various forms of freedom of speech exercised by the different social media sites, both provoked and unprovoked attacks on these minorities have increased. This kind of language has been used in cyberbullying, which is one of the major reasons behind not only lower self-esteem issues but also suicides. To address this issue, there is a growing interest in detecting hate speech to uncover and mitigate this occurrence. In this paper, we proposed a sequential machine learning model that can perform Sentiment Analysis to categorize tweets as either positive or negative speech. A Bi-directional long-short-term memory (Bi-LSTM) neural network is implemented because of its feature, the memory state cell. Since two outputs are expected, either positive or negative, a sigmoid activation function is used at the last dense layer. On evaluation of the model, it has an accuracy of 85%.