ABSTRACT
Sentiment Analysis (SA) is a field of research within Natural Language
Processing that has been growing in the last decades due
to social media and smartphones popularization. Many SA applications
make use of a corpus: a collection of data in textual form
used to train and/or test SA resources. This work describes the
construction of a corpus intended for document-level SA. The corpus
contains reviews of supermarkets throughout Brazil, extracted
from Google Places. The data were collected taking into account
the Brazilian geographic distribution and linguistic variations, and
were carefully reviewed. The corpus was then evaluated using a
k-fold cross-validation method applied in both machine learning
and deep learning techniques in which precision, accuracy, recall
and f1-score were collected and compared among the techniques.
It was also tested by a lexical approach using a domain specific
lexicon.