Dealing with high dimensional sentiment data using gradient boosting machines

Abstract

One of the most common classification tasks that applies on textual information is sentiment analysis, i.e. the prediction of the sentiment of a given document. With the vast use of social media and internet applications such as e-commerce, e-tourism and e-government, numerous comments and opinions are broadcasted per day, thus an automatic way of analyzing them is of great importance. The present paper focuses on sentiment analysis for Greek texts, obtained from Web 2.0 platforms. Greek is a language that lacks an in-depth availability of natural language processing tools in the sense that most of them are not publicly available. The novelty of the article is that instead of utilizing preprocessing tools such as Part-of-Speech taggers, text stemmers and polar-word lexica, it incorporates the translation of the Greek token as provided by the Google Translator® API. Since automatic translation of Greek sentences often results in poor translations where the meaning of the original sentence is severely deteriorated, the translation of each token individually is almost 100 % correct. However, taking the translation of every Greek token poses a significant issue to the outcome of the classification process for practically any classifier, therefore, we introduce the use of a powerful ensemble algorithm that is highly customizable to the particular needs of the application, such as being learned with respect to different loss functions and thus dealing with a large number of dimensions. This algorithm is called Gradient Boosting Machines and experimental results support our claim that it surpasses other, well-known machine learning techniques with a significant improvement for our task.