GATech at AbjadMed: Bidirectional Encoders vs. Causal Decoders in 82‑Class Arabic Medical Classification

Lead

Which modeling paradigm wins for fine‑grained Arabic medical tagging — bidirectional encoders or causal decoders? A new arXiv preprint (arXiv:2603.10008) from a GATech submission to the AbjadMed task tackles that question head‑on by describing a system for classifying Arabic medical text into 82 distinct categories. The paper lays out an encoder‑centric approach built around AraBERTv2 and reports systematic comparisons with decoder‑style alternatives as part of the shared‑task benchmarking effort.

Method

The core system is a fine‑tuned AraBERTv2 encoder enhanced with a hybrid pooling strategy that combines attention and mean representations, plus multi‑sample dropout for stronger regularization — standard moves in modern classification work, applied here to Arabic clinical language. The authors detail training recipes and ablations and position their setup against causal, autoregressive decoders to probe practical trade‑offs: complexity, calibration, and label discrimination over many classes. The writeup is available as a cross‑posted arXiv submission and serves mainly as a system description for the AbjadMed evaluation.

Significance and caveats

Arabic clinical NLP is under‑resourced compared with English, so a focused 82‑class benchmark is useful for the community. But clinical deployment is another step: datasets, privacy controls, and regulatory validation all matter before systems touch patient care. The paper contributes reproducible engineering and comparative insight; full peer review and clinical validation remain necessary. The preprint can be read at arXiv:2603.10008 for researchers who want the implementation and experimental details.