Multi-Modal Topic Sentiment Analytics for Twitter

Abstract

Sentiment analysis has proven to be very successful in text applications. Social media is also considered a quite rich source to get data regarding user’s behaviors and preference. Identifying social context would make the sentiment analysis more meaningful to the applications. Due to the limited contextual information in social media, it would be quite challenging to conduct context-aware sentiment analysis with social media. Promising frameworks such as CoreNLP, Text Blob, and Vader have been introduced to identify sentiments in the text. However, it seems to not be adequate to contextual sentiment analysis in social media like Twitter. In this thesis, we present a contextual sentiment framework that is designed to leverage the power of the multiple models in the social context. The framework aims to classify contextual sentiment from the Twitter data as well as to discover hidden trends and topics (context) using topic modeling techniques like Latent Dirichlet Allocation (LDA). We have focused on the mismatch cases among multiple models in which different experts (models) have different opinions on social media sentiments. We have identified the five mismatch types in the social sentiment through the analysis of diverse experiments ( human machine model, and machine-machine model). We have implemented the mismatch detection among the three models (i.e., Vader, Text Blob, and CoreNLP) and automatically corrected them by applying semantic rules to sentiment models. We compared our approach against a traditional single model approach concerning a performance metric (accuracy) and Kappa (evaluating consensus among multi-models) on three benchmarks datasets and our dataset we collected from a health dieting domain. The proposed framework showed notable performance improvement in comparison with the traditional one concerning both evaluation metrics.