Attention for Implicit Discourse Relation Recognition

Andre Cianflone, Leila Kosseim

Abstract: Implicit discourse relation recognition remains a challenging task as state-of-the-art approaches reach F1 scores ranging from 9.95% to 37.67% on the 2016 CoNLL shared task. In our work, we explore the use of a neural network which exploits the strong correlation between pairs of words across two discourse arguments that implicitly signal a discourse relation. We present a novel approach to Implicit Discourse Relation Recognition that uses an encoder-decoder model with attention. Our approach is based on the assumption that a discourse argument is "generated" from a previous argument and conditioned on a latent discourse relation, which we detect. Experiments show that our model achieves an F1 score of 38.25% on fine-grained classification, outperforming previous approaches and performing comparatively with state-of-the-art on coarse-grained classification, while computing alignment parameters without the need for additional pooling and fully connected layers.