Abstract
Social networks are the hub of social interactions in today’s world and conversations are a treasure trove of sentiments on these platforms. Currently, to the best of our knowledge, there is no system in place in any of the social networking platforms to predict the favorability of a ‘post’ through sentiment analysis of existing discussion on topics throughout the social networks. Having this system will allow the user to carefully reorganize their opinion so that the interpretation is well aligned with the users’ opinion and their audiences’ sentiment. However, such analysis cannot be performed effectively with the use of traditional statistical methods due to the inherent complexity of the problem. In this project, we aim to design and implement a system to support the users in predicting the favorability of a ‘post’. The proposed system consists of two modules. The first one is a sentiment analysis deep learning network using Bidirectional LSTM layers, which predicts the sentiment score of each reply present in the conversation tree collected as a part of the dataset using twitter API. This model is trained using Sentiment140 dataset [1]. The second module leverages the classification results from the first module to perform regression analysis on original tweets. This model is trained on the dataset collected using twitter API. Specifically, the output of the second module is the favorability score of any given tweet, which is based on the predicted percentage of positive responses. Our project also reveals the significant performance gain of using word embedding techniques, e.g., GloVe [2], over the traditional count or TF-IDF based word vector representation. The data for this project has two origins. While the labeled data for sentiment analysis was taken from Sentiment140 website [1], the data for twitter conversation tree was collected using twitter API.