I-NOWJ: adaptive context selection and knowledge-guided learning for substance use named entity recognition.

Abstract

Substance use information in clinical narratives is clinically important but is typically documented in unstructured and informal text, making automatic extraction challenging, particularly in low-resource and non-English settings. The BioCreative IX ToxHabits shared task addresses this problem in Spanish clinical text by identifying substance use triggers and their arguments. In this paper, we introduce I-NOWJ, which addresses these challenges through two complementary strategies: a low-resource strategy based on domain knowledge expansion, controlled data augmentation with quality-aware self-training, and an adaptive context selection mechanism that determines the appropriate contextual scope for each instance. Experimental results on the ToxHabits benchmark show that I-NOWJ achieves an F1 score of 94.93% on trigger detection and 91.30% on argument extraction. These results demonstrate the effectiveness of knowledge-guided modelling and adaptive context selection for robust clinical named entity recognition in low-resource settings. Our implementation and resources are publicly available at: https://github.com/candleMind/toxhabit.