Related Work Generation with Variational Sequential Planning

The RWG(Related Work Generation) aims to thoroughly explore research topics by referencing a provided list of cited papers. The organization of related work content generally falls into integrative and descriptive styles. Most research in this field tends to create integrated content, concentrating on combining ideas and findings from reference papers while offering fewer specifics about individual studies. In this paper, we address the descriptive style of related work generation, which offers more detailed information about each referenced study, including methods, results, and interpretation, presented in a logical order. We propose a Related Work Generation model with Variational Sequential Planning (RWG-VSP) to achieve this. RWG-VSP learns a sequence of discrete variables that organize high-level information coherently and meaningfully. During the planning phase, we introduce a keyphrase-guided attention mechanism to highlight important parts of each cited paper in descriptive related work. During decoding, we enhance the decoder by incorporating a keyphrase-augmented attention mechanism to bring hierarchical context to the generation model. The proposed method’s efficacy is supported by thorough experiments on two collected datasets.