R2D2: Recursive Transformer based on Differentiable Tree for Interpretable Hierarchical Language Modeling

TL;DR

This paper introduces R2D2, a recursive Transformer model that explicitly encodes hierarchical language structures using differentiable trees, improving language modeling and unsupervised parsing efficiency.

Contribution

It proposes a novel recursive Transformer architecture with differentiable trees and an efficient pruning algorithm for scalable hierarchical language modeling.

Findings

Effective in language modeling tasks

Improves unsupervised parsing accuracy

Scales linearly with sequence length

Abstract

Human language understanding operates at multiple levels of granularity (e.g., words, phrases, and sentences) with increasing levels of abstraction that can be hierarchically combined. However, existing deep models with stacked layers do not explicitly model any sort of hierarchical process. This paper proposes a recursive Transformer model based on differentiable CKY style binary trees to emulate the composition process. We extend the bidirectional language model pre-training objective to this architecture, attempting to predict each word given its left and right abstraction nodes. To scale up our approach, we also introduce an efficient pruned tree induction algorithm to enable encoding in just a linear number of composition steps. Experimental results on language modeling and unsupervised parsing show the effectiveness of our approach.