Reinforced Subject-Aware Graph Neural Network for Related Work Generation

Luyao Yu, Qi Zhang, Chongyang Shi, An Lao, Liang Xiao

Abstract: The objective of automatic related work generation is to gather the primary contributions of relevant prior work in a research field and provide a comprehensive analysis, which assists authors in drafting a related work section efficiently, saving them time and effort. The unique characteristic of the related work generation makes the task challenging. However, most existing abstractive related work generation methods are implemented at a coarse granularity, which leads to the complex relationships and interactions among multiple papers are not effectively modeled. In this study, we propose an abstractive Reinforced Subject-aware Graph Neural Network for Related work Generation (RSG) to explore the relationships between the target and the related reference papers based on the writing style of the related work section. Since these relationships are often not explicit, we first leverage the capability of the large language model (LLM) to extract keyphrases among the given papers. Building upon this, we introduce a keyphrase-guided selective encoding mechanism to augment the representations of the given papers. Considering the keyphrases as the subjects discussed within the papers, we propose a subject-aware graph to model the relationships between the papers and the subjects by constructing a hierarchical structure. In the decoding phase, we extend the transformer decoder by keyphrases augmented attention mechanism to integrate various information into the generation process. Extensive experiments on two benchmark datasets demonstrate the effectiveness of the proposed model.