Seleção de dados para transferência de aprendizado no contexto de análise de sentimentos em tweets

Abstract

The advent and popularization of social networks have been leading more and more people to feel free to express their opinions on various issues in those environments. This type of attitude generates a growing volume of data, whose analysis is an important tool in the decision-making process of institutions, governments or people, that can assess their performance related to a desired target audience. The computational field of study that aims to meet this objective is called sentiment analysis, which has the polarity classification of texts as one of its most prominent tasks. To meet the need to classify texts as positive or negative, the use of approaches based on supervised machine learning is promising, in which a classifier is trained with a dataset from a given domain whose labels (positive or negative) are known. The idea behind this approach is that this classifier can predict the labels of new data from this same domain. However, labeled data are not always available as the domain of interest can be rare and data scarse, or manually labeling the data can be prohibitive. In this scenario, transfer learning strategies arise, seeking to take advantage of the knowledge acquired in a given source domain to adapt or reuse classifiers for a given target domain. One of the approaches used is based on data selection or enrichment from a source domain - which has been widely proposed in the literature. However, there is a lack of specific studies for instance selection in the challenging scenario of Twitter. This dissertation seeks to investigate data selection techniques for transfer learning in the scenario of sentiment analysis in tweets. For this, experiments are conducted using a set of 22 tweets datasets in English. These experiments propose techniques: (i.) to select source datasets to train classifiers for an unlabeled target dataset, (ii.) to select instances of the union of source datasets to train classifiers for an unlabeled target dataset and (iii.) to select instances of the union of source datasets to train classifiers for a labeled target dataset. With the proposed techniques, it is observed that the size of training set plays a fundamental role in the predictive capability of the classifiers and that using balanced and diverse training sets constitutes a good decision for transfer learning methods based on instance selection and reuse of classifiers.