Aplicação de técnicas de análise de sentimento para classificação de postagens em ambiente virtual de aprendizagem

Abstract

In this age of technologies aimed at helping on human labor, it is possible to relegate the activity of discovering concordancy with a subject via sentiment analysis. As the professors on a Virtual Learning Environment have to work with a lot of text, we propose the conception of a tool which can help them on this matter. This work has the objective of demonstrating, via the implementation of such tool, how the utilization of Sentiment Analysis may be useful in Virtual Learning Environments, as a way for a professor/tutor to discover the general tendency of a group about some subject that they are interested in without the need of reading all the texts – those may be of great length and in great quantity. The methodology applied here has four distinct steps. First, we collect the login data of professors/students that have active accounts on SOLAR. Then, with the data of no less than ten students, or one student with ate least ten forums, or one professor with active course subjects, the forum texts are collected and written to a database by the first module of this application. The second module is then executed, which does the Sentiment Analysis per se and classifies the texts as positive, negative or neutral. Then we get a third person, unrelated to the experiment, and give them a percentage of the classified texts so we can make a cross-reference to test the application accuracy. The results show that the application has the overall capability of getting 79% of the classifications right, and individually some classifications can go as high as 93% - which characterizes the application as being trustworthy enough to do the job it’s supposed to. As a conclusion, we can attest that this classification method – which is considered very simple – can show extremely satisfactory results, depending only on the context it is applied to; that, as of the time of writing, there are no brazilian portuguese dictionaries that are good enough for this type of application, but even a relatively simple one can still show satisfactory results; and that the developed application is enough to cover all the objectives it is built to do.