# Learning Latent Trees with Stochastic Perturbations and Differentiable Dynamic Programming

## TL;DR

This paper introduces a novel method for learning latent projective dependency trees using stochastic perturbations and differentiable dynamic programming, improving downstream NLP tasks without direct tree supervision.

## Contribution

It presents a fully differentiable, stochastic approach to latent tree learning that outperforms previous methods and emphasizes the importance of stochasticity and projectivity constraints.

## Key findings

- Effective on sentiment analysis and natural language inference
- Stochastic sampling improves structure induction
- Ablation studies highlight the role of stochasticity and projectivity

## Abstract

We treat projective dependency trees as latent variables in our probabilistic model and induce them in such a way as to be beneficial for a downstream task, without relying on any direct tree supervision. Our approach relies on Gumbel perturbations and differentiable dynamic programming. Unlike previous approaches to latent tree learning, we stochastically sample global structures and our parser is fully differentiable. We illustrate its effectiveness on sentiment analysis and natural language inference tasks. We also study its properties on a synthetic structure induction task. Ablation studies emphasize the importance of both stochasticity and constraining latent structures to be projective trees.