Bidirectional Macro-level Discourse Parser Based on Oracle Selection

Longwang He, Feng Jiang, Xiaoyi Bao, Yaxin Fan, Weihao Liu, Peifeng Li, Xiaomin Chu

Abstract: Most existing studies construct a discourse structure tree following two popular methods: top-down or bottom-up strategy. However, they often suffered from cascading errors because they can not switch the strategy of building a structure tree to avoid mistakes caused by uncertain decision-making. Moreover, due to the different basis of top-down and bottom-up methods in building discourse trees, thoroughly combining the advantages of the two methods is challenging. To alleviate these issues, we propose a Bidirectional macro-level dIscourse Parser based on OracLe selEction (BIPOLE), which combines the top-down and bottom-up strategies by selecting the suitable decision-making strategy. BIPOLE consists of a basic parsing module composed of top-down and bottom-up sub-parsers and a decision-maker for selecting a prediction strategy by considering each sub-parser state. Moreover, we propose a label-based data-enhanced oracle training strategy to generate the training data of the decision-maker. Experimental results on MCDTB and RST-DT show that our model can effectively alleviate cascading errors and outperforms the SOTA baselines significantly.