| Resumo: |
A huge number of short informal messages are posted every day in social network sites, discussion forums and customer surveys. Emotions seem to be frequently important in these texts. The challenge of identifying and understanding an emotion present in this type of communication is important in distinguishing the sentiment in the text and also in identifying anomalous and inappropriate behaviors, eventually offering some kind of risk. This work proposes the implementation of a sentiment analysis solution based on machine learning. Using supervised learning techniques, it is desired to discern whether a message has a positive, neutral, or negative sentiment. The messages to be analyzed are IT service satisfaction surveys. Two models were used in the analysis, the first model where only the ”Comment”, a nonstructured text field was considered and the second model, where besides the ”Comment”field, two objective questions were considered. The results obtained indicate that the techniques of machine learning, are not behind the results produced by human-produced baselines. The accuracy obtained was up to 86.8% accuracy for a three class model: ”praise”, ”neutral”and ”complaint”. Accuracy was significantly higher, reaching up to 94.5 % in an alternative model of only two classes: ”praise”and ”non-praise”. |
|---|
