Improving Syntactic Parsing with Consistency Learning

Abstract

In this paper, we propose using consistency learning to improve constituency and dependency parsing performances on a multi-task setting. It utilizes a consistent constraint between the predictions. While multi-task learning implicitly learns shared representations for multiple sub-tasks, our method introduces an explicit consistency objective, which encourages shared representations that result in consistent predictions. Our intuition is that correct predictions are more likely consistent ones. To introduce consistent constraints, we propose a general method for introducing consistency objectives, as well as other prior knowledge, into existing neural models. This method only requires a boolean function that tells whether or not the multiple predictions are consistent, which does not need to be differentiable. We demonstrate the efficacy of our method by showing that it out-performs a state-of-the-art joint dependency and constituency parser on CTB.

Tasks

Reproductions