Twitter Sentiment Polarity Analysis: A Novel Approach for Improving the Automated Labeling in a Text Corpora

Abstract

The high penetration of Twitter in Chile has favored the use of this social network for companies and brands to get additional information on user opinions and feedback about their products and services. In recent years there have been many studies to determine the polarity of a comment on Twitter mainly considering three classes: positive, negative and neutral. One big difference inherent in the problem of Sentiment Analysis on Twitter as opposed to the Web is the ease with which you can obtain data to perform a supervised training algorithm with its API. To take advantage of this characteristic it is necessary to find a semi-automatic method for obtaining tweets to generate the corpora and avoid the traditional method of manual labeling which is very demanding in time and money. This paper goes deeper into the work of using a semi-automated generated corpora for Twitter sentiment polarity classification, introducing a novel approach of tweet selection for corpora consolidation and the addition of a fourth class of tweets that doesn’t correspond to any of the above. This new class includes tweets that are irrelevant for classification and do not contain much information, a type of posts that Twitter is full of. Experimental evaluations show that the usage of the fourth class of the denominated meaningless tweets together with the tweet filter criteria for corpus generation improves the system accuracy.