Polarity Detection of Foursquare Tips

In location-based social networks, such as Foursquare, users may post tips with their opinions about visited places. Tips may directly impact the behavior of future visitors, providing valuable feedback to business owners. Sentiment or polarity detection has attracted great attention due to its vast applicability in opinion summarization, ranking or recommendation. However, the automatic detection of polarity of tips faces challenges due to their short sizes and informal content. This paper presents an empirical study of supervised and unsupervised techniques to detect the polarity of Foursquare tips. We evaluate the effectiveness of four methods on two sets of tips, finding that a simpler lexicon-based approach, which does not require costly manual labeling, can be as effective as state-of-the-art supervised methods. We also find that a hybrid approach that combines all considered methods by means of stacking does not significantly outperform the best individual method.