همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Divide and Conquer: A Cascaded Architecture for Relation Extraction on Long-Tailed Datasets

نویسندگان :

Parham Rahimi¹ Behrouz Minaei-Bidgoli²

1- دانشگاه علم و صنعت ایران 2- دانشگاه علم و صنعت ایران

کلمات کلیدی :

Relation Extraction،Distant Supervision،Long-Tailed Datasets،Prompt Engineering Framework،Small Language Models

چکیده :

Relation Extraction (RE) from unstructured text is a critical task for populating Knowledge Graphs (KGs). The Distant Supervision (DS) paradigm, while effective for data annotation, suffers from two critical challenges: the wrong label problem (noise) and severe class imbalance in benchmark datasets. While recent generative models have shown promise, they struggle with these real-world data imperfections. This paper proposes a novel framework to address these challenges. First, we introduce a prompt-engineering methodology to convert traditional DS datasets (e.g., NYT) into a format suitable for fine-tuning generative models. Second, we demonstrate that a T5 model fine-tuned with this framework achieves state-of-the-art results on the NYT24 joint RE task, surpassing prior baselines. Finally, to combat extreme data imbalance, we propose a novel cascaded T5 architecture that improves classification F1-score by over 6% compared to a single-stage model. Our work provides a robust, generative solution for RE in noisy and imbalanced settings.