Weakly Supervised Joint Entity-Sentiment-Issue Model For Political Opinion Mining

Microblogging has become an important source of opinion-rich data that can be used for understanding public opinion. In this paper, we propose a novel weakly supervised probabilistic topic model, Joint Entity-Sentiment-Issue (JESI), for political opinion mining from Twitter. The model automatically identifies the target entity of the expressed sentiment, the issues discussed and the sentiment towards the issues and entity simultaneously. Unlike other machine learning approaches to opinion mining which require labelled data for training classifiers, JESI requires only a small number of seed words for each entity and issue, and a sentiment lexicon. The model is evaluated on a dataset of tweets collected during the 2016 Australian Federal Election. Experimental results demonstrate that JESI outperforms baselines for sentiment, entity and issue classification, especially achieving higher recall and F1.