Multi-Level Implicit Discourse Relation Classification via Sequence-Level Reinforcement Learning
Abstract
AbstractImplicit Discourse Relation Classification (IDRC) identifies discourse relations between text spans when connectives are absent. The task is typically evaluated on the Penn Discourse Treebank (PDTB), where discourse senses are organized in a hierarchical structure. Recent unified models predict discourse labels level by level under teacher forcing, which leads to exposure bias and cascading errors across the hierarchy. Moreover, token-level training objectives fail to assess the correctness of the entire label sequence. We reformulate multi-level IDRC as hierarchical label sequence generation and optimize the full label sequence using sequence-level reinforcement learning. By sampling and ranking complete label sequences, the proposed framework directly optimizes global sequence correctness rather than isolated token-level decisions. Experiments on PDTB 2.0 and 3.0 demonstrate consistent improvements over strong baselines. Further analysis shows that fine-to-coarse decoding outperforms the conventional coarse-to-fine strategy, highlighting the importance of decoding order in hierarchical discourse modeling.