Enriquecimento de dicionários para aprimoramento da classificação automática de sentimento em postagens relacionadas ao uso de sistemas.

Abstract

Textual evaluation consists of using user narratives in order to assess or gain some insight into the system through its postings. This evaluation is very important, since the increasing popularization and dissemination of the developed applications and the increase of the quality standard demanded by the users are characteristics more and more present at the moment. It then becomes necessary for software developers to enhance their applications on a regular basis, since competition is always present and the user’s need may change suddenly. For this reason, it is indispensable to study ways to perform textual evaluation automatically, since manually classifying the large amount of information made available by users in a viable time is an impossible task. There is also the difficulty of detecting the user’s positive, neutral or negative feelings about the system through written Portuguese posts. Based on this problem, this work presents an investigation to improve the classification of a lexical-based classifier, the SentiStrength, for automatic classification of sentiments in postings related to the use of systems. To achieve this goal, the TF-IDF metric was used to select words that are within the domain of the related posts, which will enrich the dictionary used by the SentiStrength to generate the polarity of the posts. We also investigated the efficiency of dictionaries enriched with words in their root (stemming) and with lemmatized words. The research was conducted with 2108 sentences extracted from the reviews section of the Play Store on urban mobility applications, such as Waze, Google Maps and GPS Brasil. One of the results obtained was an average increase of 7.3 % in the accuracy of the classifier when using the enriched dictionaries.