Análise de Sentimentos: Impacto da Tradução Neural na Avaliação de Desempenho

Abstract

With the increasing ease of internet access in recent years, there has been a significant increase in user-generated content, including product reviews and social media analyses, among other types of content. This data-rich environment has proven to be beneficial for sentiment analysis. However, many languages face a lack of resources and datasets for effective sentiment analysis. To address these data gaps, for example in the Portuguese language, this study proposed the use of neural machine translation as a strategy to obtain sentiment analysis data from more widely explored languages, such as English. The study presented two experimental scenarios for sentiment analysis after automatic neural translation of texts, one with tweets and the other with product reviews. The results obtained from the original tweets in the Portuguese language showed an F1-Score of 59.08%, precision of 60.94%, and recall of 59.31% for three sentiment classes: Positive, Negative, and Neutral. After neural translation, the results showed an F1-Score of 53.29%, recall of 54.73%, and precision of 52.62% for the same classes. In the scenario of product reviews, where only two sentiment classes, Negative and Positive, were considered, the results in the original language yielded an F1-Score of 83.00%, with precision and recall also at 83.00%. After translation, the results showed an F1-Score of 80.05%, with precision and recall also at 80.05%. These results indicate errors in the neural translation process that hindered sentence classification and incorrect classification, particularly in sentences that did not contain words directly related to sentiment, as sentiment was represented by the entire sentence. Thus, the study suggests that neural translation can be a valuable tool for gathering sentiment analysis resources in languages with limited resources, as a means to overcome the lack of data for research in specific languages.