Uma arquitetura de pré-processamento para análise de sentimento em mídias sociais em português brasileiro

Abstract

The Web 2.0 and the evolution of Information Technologies have brought novel interaction and relationship channels. In this context, the Online Social Networks (OSN) are an example as platforms which allow interactions and sharing of information between people. In this scenario, it is possible to observe the adoption of OSN as a channel for posting opinions regarding products and experience. This scene presents an excellent opportunity for companies that aim to improve products, services and marketing strategies, given OSNs are powerful sources of massive unstructured data generated by consumers (UGC), with opinions and reviews concerning offers, in platforms such as Facebook, Twitter and Instagram. Brazil is a highlight in this scenario, where this phenomenon can be observed, as the Brazilian population is one of the most active in social media platforms in the world. This makes it a country full of opportunities to market exploitation. In this context, computational techniques of Opinion Mining and Sentiment Analysis (SA) are applied aiming to infer the polarity (positive, negative, neutral) regarding a sentiment associated to texts, and can also be applied in data from OSN to evaluate the feedback from a target audience. Although the existing diversity of SA strategies reported in the literature, there are still challenges faced in the application of SA in text data from OSN, given the characteristics of the language adopted in such platforms. The state of art is focused on SA towards the English language, and the existing proposals for Brazilian Portuguese do not have a standardized methodology for preprocessing steps. In this context, this research investigates an approach with no translation, and proposes a novel preprocessing architecture for SA towards Brazilian Portuguese, aiming to provide enriched features to SA algorithms. The proposal was compared with well-established baselines from the literature, and the obtained results indicate that this architecture can overcome the state of art recall in at least 3% , for 6 out of 7 datasets evaluated.