Remove Noise and Keep Truth: A Noisy Channel Model for Semantic Role Labeling

Abstract

Semantic role labeling usually models structures using sequences, trees, or graphs. Past works focused on researching novel modeling methods and neural structures and integrating more features. In this paper, we re-examined the noise in neural semantic role labeling models, a problem that has been long-ignored. By proposing a noisy channel model structure, we effectively eliminate the noise in the labeling flow and thus improve performance. Without relying on additional features, our proposed novel model significantly outperforms a strong baseline on multiple popular semantic role labeling benchmarks, which demonstrates the effectiveness and robustness of our proposed model.

Tasks

Reproductions