The article examines the cross-domain generalizability of different model architectures through a blind zero-shot evaluation. By training on various datasets (Paheli, M-CL, M-IT) with a unified rhetorical label space, the study finds that while the baseline model showcases a commendable ability to transfer knowledge across domains, models specifically designed for discourse or prototypical learning exhibit distinct performance dynamics. Notably, single prototype models outshine their counterparts by representing core data characteristics, thus ensuring better cross-domain applicability, despite the peculiar challenges presented by certain datasets like M-IT.
The baseline model demonstrates a robust ability to transfer knowledge across different domains, maintaining performance above random guessing despite variations in training and testing datasets.
Disc. Contr. marginally reduces cross-domain performance, suggesting it captures too many domain-specific features while struggling to generalize across diverse datasets.
Single and multi-prototypical models show enhanced cross-domain transfer capabilities, where single prototypes excel by encapsulating core features and being less sensitive to domain variations.
The coupling of discourse-aware contrastive models with prototypical learning boosts cross-domain performance, although M-IT presents challenges likely tied to its in-domain characteristics.
Collection
[
|
...
]