What drives social sentiment? An entropic measure-based clustering approach towards identifying factors that influence social sentiment polarity

Abstract

Analyzing the public sentiment over social media streams constitutes an extremely demanding task mainly due to the difficulties that are imposed by the wide spectrum of discussion topics that underlie a given collection of posts. This paper addresses the problem of determining the underlying semantic factors that influence the social sentiment polarity in a given corpus of posts through the utilization of an entropic measure-based clustering approach. Extant studies examine the semantic structure of social network data primarily through topic modeling or sentiment analysis methods. The novelty of our approach lies upon the utilization of a semantically-aware clustering procedure that effectively combines topic modeling and sentiment analysis algorithms. Our approach extends the fundamental assumption behind traditional sentiment analysis methods, according to which sentiment can be associated with low level document features such as words, phrases or sentences. We argue that sentiment can be associated with higher level entities such as the semantic axes that span a given volume of posts, thus performing sentiment analysis at the topic level. Our experimentation provides strong evidence that combining topic modeling and sentiment analysis results by a semantically-aware clustering procedure can reveal the distribution of the overall public sentiment on the underlying semantic axes.