Acquiring commonsense knowledge for sentiment analysis using human computation

Marina Boia, Claudiu Cristian Musat, Boi Faltings

Abstract: Many Artificial Intelligence tasks need commonsense knowledge. Extracting this knowledge with statistical methods would require huge amounts of data, so human computation offers a better alternative. We acquire contextual knowledge for sentiment analysis by asking workers to indicate the contexts that influence the polarities of sentiment words. The increased complexity of the task causes some workers to give superficial answers. To increase motivation, we make the task more engaging by packaging it as a game. With the knowledge compiled from only a small set of answers, we already halve the gap between machine and human performance. This proves the strong potential of human computation for acquiring commonsense knowledge.