The Problem
Sarcasm in Indonesian social media posts can be mistaken for literal praise or criticism. Detecting it is challenging because the meaning often depends on context, while larger language models can be costly to run.
My Contribution
As first author, I conducted most of the experiments, wrote most of the paper, and presented the research. I worked with an Indonesian Twitter sarcasm dataset and investigated whether a classical classifier and a transformer could improve each other’s predictions.
Process and Challenges
Many approaches I tried failed to improve the results. The paper documents one important finding: a more complex ensemble of classical models did not outperform an optimized logistic regression baseline. I then combined that baseline with transformer predictions using late fusion.
Results
The best combination, optimized logistic regression with XLM-R-base, achieved an F1 score of 0.7900 on the test set, compared with 0.7692 for the previously reported benchmark model. It used about half as many model parameters. The observed F1 improvement did not reach conventional statistical significance. The paper was accepted and published at iSemantic 2026, where I received Best Paper and Best Presenter awards.
Reflection
This research taught me to examine failed experiments and strengthen a simple baseline before adding complexity. It also taught me to report both promising results and the limits of the evidence behind them.