Date of Publication :17th January 2024

Abstract:Several consumer protection entities have their own online platforms to collect consumers' experience and to mediate the resolution of conflicts between them and the suppliers. In Brazil, one of the most relevant tools for this purpose is the Consumidor.gov.br platform: it enables the communication of consumers with product/service suppliers in order to solve conflicts/disputes between the parts; at the end of the mediation process, the consumer writes his/her final opinion about the service provided and about the fulfillment of the initial expectations. This work proposes to build a sentiment analysis model through Transfer Learning, performing the fine-tuning process of the BERT model via a training set which aggregates well-known public data-sets of customer reviews in Portuguese, which is subsequently evaluated in the task of sentiment analysis of Consumidor.gov.br complaints, through a test set exclusively labeled for this purpose. The model is deployed in Senacon data analysis environment and is able to perform sentiment analysis of the finalized disputes on the platform, on a daily basis, with expected average F1-score of 0.72.

Reference :