Interaction between encoder fine-tuning and distance-based augmentation

Determine whether fine-tuning the RoBERTa encoder on the full augmented training set produces further gains and how encoder adaptation interacts with distance-based synthetic data augmentation for discourse-pragmatic function classification.

Background

The experiments use frozen mean-pooled RoBERTa representations so that performance differences can be attributed to the augmentation strategy rather than changes in the feature space. The paper notes that this design prevents the representations from adapting to the distributional properties of discourse-pragmatic data.

The unresolved issue is whether jointly adapting the encoder through fine-tuning and controlling the geometric placement of synthetic examples would improve classification performance, or whether the effects of the two interventions interact in a meaningful way.