Lejing Yu

Rel. These results suggest that RexDrug’s reasoning-driven generative paradigm is more robust than surface-pattern matching when modeling highly complex pharmacological semantics, reducing reliance on human-annotated entity information and enabling more accurate relation extraction. We further analyze model performance on negative samples (NO_COMB) and higher-order n-ary combinations in Supplementary Section S4, along with a representative case study. The results show that RexDrug improves discrimination of true interactions and the modeling of complex pharmacological regimens. Also in Supplementary Section S4, under a relation classification setting with gold entity spans (consistent with PURE* and RCFIND*), RexDrug improves POS-Exact F1 by 8.4% over the best baseline. We also extend the task to joint entity–relation extraction. Despite the increased task complexity, RexDrug achieves 95.1% F1 on the NER subtask (Supplementary Section S5), demonstrating strong robustness under more demanding structured extraction settings. 5.3 Performance on the External DDI13 Dataset To assess the applicability of RexDrug, we evaluate its performance on the binary DDI13 dataset (Herrero-Zazo et al., 2013) under both relation extraction and relation classification settings. Accordingly, we slightly adapt the metric-based reward to match the DDI13 evaluation protocol: instead of the Exact/Partial matching used for n-ary DCE, we compute the Combination Metric Reward using the micro-averaged F1 score over DDI relation types. This adjustment keeps the RL objective aligned with the dataset-specific metric while leaving the other reward components unchanged. In addition to GPT-based models, we include the following baselines: Relation Extraction: TP-DDI (Zaikis and Vlahavas, 2021) uses a BioBERT-based pipeline for sequential entity and relation prediction; MRC