Move Structure Recognition in Scientific Papers with Saliency Attribution

Move analysis is a primary research topic in computational linguistics that relates to pragmatics. It plays a crucial role in analyzing the intent and coherence of the text. This paper introduces a innovative exploration of move analysis to scientific papers and presents a novel task - move structure recognition in scientific papers. Existing datasets are inadequate to support this task. Thus, we manually annotated a dataset called Scientific Abstract Moves Dataset (SAMD). The implicit mixture and counterfactual reasoning in the move structure’s content has led to poor performance in move recognition. This research examines the issue in depth and presents a new concept of move saliency attribution, which can illuminate the contribution of words to specific move structures. On this foundation, we design a new move recognition training mechanism, which fully consider the context information of the move to achieve promising performance on SAMD and NLPContributionGraph shared task dataset (NCG). This is the first attempt at interpretability of move recognition, giving us the possibility to understand how the model makes decisions and identify potential biases or errors in the model.