Sentiment analysis, which involves the computational study of people’s opinions and emotions, has been applied in the student context to understand satisfaction and difficulties and mitigate school dropout. This article is part of a project aimed at mitigating dropout in programming courses in Brazilian higher education, utilizing sentiment analysis in conjunction with psycho-pedagogical methods and Natural Language Processing (NLP) techniques. The present research compares the effectiveness of three automated sentiment extraction tools - two based on Large Language Models (LLMs) and a lexical analyzer. This study focuses on distinguishing ambiguous sentiment with a phrase-to-phrase refinement approach. Results show that the approach improves ambiguous classification from 6% to 84% in the best case.