Análise adaptativa de fluxos de sentimento

Abstract

In recent years, the sentiment analysis task has attracted the interest of the machine learning researchers. This interest has grown significantly due to the large volume of opinionated content generated and shared via social media. Considering the benefits of to know the sentiment of the population regarding to different topics and entities, the analysis of the generated content by social media, it is a promising and necessary task. Many automatic classification techniques have been used to perform sentiment analysis, however is consensus that the arrival pattern of messages from social media follows the data stream paradigm and the traditional classification techniques are not adequate to address the specific characteristics of the created sentiment stream. Among the challenges to the classification techniques can be highlighted: (1) concept drift (i.e., constant changes in data characteristics, which in this study was approached as sentiment drift), (2) the need of real-time update of the classification model from the most recent messages and (3) the limitation of computing and training resource, which makes the two firsts cited challenges more difficult.We analyze these problems from a proposal of semi-supervised learning. Our algorithm adapts the training set, to the changes in the data, from a self-augmenting training process with the passes of the stream. It uses a small seed of initial training and then classification models are produced in real time using association rules. This strategy keeps the model up-to-date incrementally, so that at any time of the event the model reflects the sentiment that is being transmitted. In order to address the sentiment drift, messages to training are projected on-demand, according to the message content that is being classified. Projection of the training data offers a number of advantages including the ability to quickly detect emerging trends in the information stream. We conducted a case study using the Twitter messages, posted in real time, related to major events in 2010 year. In these experiments the performance of the prediction keeps the same or increases, with the passes of the stream and the inclusion of new messages in the training set. We evaluated the proposed solution in different languages, in cases where the sentiment distribution changed in different way over time and in cases where the initial training seed is rather small.