Avaliação de modelos de processamento de linguagem natural para análise de sentimentos em avaliações de restaurantes

Abstract

The food sector has a significant impact on the Brazilian economy, and with the digitalization of how society seeks information about establishments, review platforms like Google Maps have become crucial in consumer perception and decision-making. However, simply viewing overall star ratings masks important nuances in customer sentiment and hinders the extraction of detailed metrics from large volumes of text. To address this gap, this study evaluates the performance of different Natural Language Processing (NLP) models in the task of sentiment analysis in Google Maps restaurant reviews. To this end, the methodology consisted of automated data collection (Web Scraping) from establishment pages on Google Maps, followed by processing the textual review database using the BERT, XLM-ROBERTa, and LeIA models. The results demonstrated that the BERT model showed the best suitability for direct comparisons with user ratings. XLM-ROBERTa, in turn, demonstrated high reliability, although its categorical classification limited precision in numerical comparisons. The LeIA model, despite using a more classical lexical approach, produced consistent results comparable to the others. Furthermore, the analysis revealed a behavioral disparity: the sentiment polarity extracted exclusively from the texts tended to be lower than the establishment's overall rating. It is concluded that the application of Natural Language Processing models provides complementary metrics that can be useful for the management of gastronomic businesses, overcoming the analytical limitations of ratings based solely on stars.