Research Goal
This research explored multilabel emotion classification using the GoEmotions dataset. It investigated continual learning strategies to help models learn sequential tasks while reducing catastrophic forgetting.
My Contribution
I joined a six-person research team as a research assistant. I contributed to writing the research proposal and paper, and conducted experiments with continual learning methods, including EWC, replay-based approaches, and an AGCN-based model architecture.
Challenges
Because the research was already underway when I joined, I needed to catch up on the team’s research questions, methods, and prior work. I read relevant papers to build that understanding. Model training was also computationally demanding.
Results and Reflection
Among the reported approaches, AESL achieved the strongest results in the evaluation table, with a macro-F1 score of 79.46. This experience taught me how to join an ongoing research effort, build context through literature review, and contribute to experiments despite the computational demands of model training.