Discourse element and relation identification in argumentative essay as self-knowledge exploring based question answering with boosting

Qing Zhang, Jiaxing Sun, Jie Liu, Yanxin Xu, Li He, Jianyong Duan

Abstract: Recently, large language models (LLMs) such as GPT\(-\)3.5 have demonstrated significant performance on many information extraction (IE) tasks. However, they still fall behind smaller supervised state-of-the-art (SOTA) models in addressing discourse-level issues. This is primarily due to the scarcity of rich discourse-level datasets for supervised fine-tuning (SFT) and the absence of explicit discourse-level knowledge, making these tasks particularly challenging. In this paper, we focus on two challenging tasks in this line of research: discourse element identification and relation identification. The two tasks are traditionally addressed separately and have never been explored jointly. We argue that solving these two related tasks together is beneficial, especially given the challenging low-resource setting. To explore this problem, we first propose a novel dataset for joint discourse-level extraction task in argumentative essays. Then, we propose a novel joint discourse-level learning framework that innovatively introduces a self-knowledge-exploring question-answering (QA) method combined with a boosting approach for joint extraction in a multi-turn fashion. This not only enhances the semantic interaction between the two tasks but also addresses the inherent issue of error propagation that arises when considering the tasks independently. The proposed framework can be viewed as a novel retrieval-augmented approach, which uses the training data itself as the retrieval source, incorporates training error guidance into data augmentation and combines it with a multi-turn QA approach to improve the performance of joint tasks. Experimental results show that our method outperforms existing strong baseline methods including GPT\(-\)3.5 on the English argumentative essay dataset.