On the combination of "off-the-shelf" sentiment analysis methods.

Sentiment analysis has become an important topic on the Web, especially in social media, with applications in many domains such as the monitoring of businesses and products as well as the analysis of the repercussion of important events. Several methods and techniques have been independently developed for this purpose in the literature. However, recent work has showed that all of them have varying degrees of coverage and prediction accuracy, with no "silver bullet" for all cases and scenarios. In this paper, we tackle this issue by proposing ensemble combination methods aimed at combining the outputs of several state-of-the-art proposals in order to maximize both goals, which sometimes can be conflicting. We focus on combining "off-the-shelf" methods, increasing enormously the applicability of our strategy. We tested our solutions in a very rich experimentation environment, covering thirteen widely used methods and fourteen labeled datasets from many domains, including messages from social networks, movie and product reviews, opinions and comments in news articles. Our experimental results demonstrate that we can be very successful in our goal, meaning that our proposal can produce a real and important impact in the area of sentiment classification research.