Automatic Unsupervised Polarity Detection on a Twitter Data Stream

In this paper we propose a simple and completely automatic methodology for analyzing sentiment of users in Twitter. Firstly, we built a Twitter corpus by grouping tweets expressing positive and negative polarity through a completely automatic procedure by using only emoticons in tweets. Then, we have built a simple sentiment classifier where an actual stream of tweets from Twitter is processed and its content classified as positive, negative or neutral. The classification is made without the use of any pre-defined polarity lexicon. The lexicon is automatically inferred from the streaming of tweets. Experimental results show that our method reduces human intervention and, consequently, the cost of the whole classification process. We observe that our simple system captures polarity distinctions matching reasonably well the classification done by human judges.