Latent Tree Learning with Differentiable Parsers: Shift-Reduce Parsing and Chart Parsing

TL;DR

This paper introduces a new latent tree learning model using differentiable shift-reduce parsing, demonstrating competitive performance and analyzing the nature of the induced parse trees compared to chart-based models.

Contribution

It presents a novel shift-reduce based latent tree learning model and provides an analysis of the induced parse trees, advancing understanding of learned syntactic structures.

Findings

The shift-reduce model achieves competitive downstream task performance.

Induced trees are non-trivial and comparable to those from chart-based models.

Analysis reveals insights into the structure of learned parse trees.

Abstract

Latent tree learning models represent sentences by composing their words according to an induced parse tree, all based on a downstream task. These models often outperform baselines which use (externally provided) syntax trees to drive the composition order. This work contributes (a) a new latent tree learning model based on shift-reduce parsing, with competitive downstream performance and non-trivial induced trees, and (b) an analysis of the trees learned by our shift-reduce model and by a chart-based model.