Exploiting Inductive Bias in Transformers for Unsupervised Disentanglement of Syntax and Semantics with VAEs

TL;DR

This paper introduces QKVAE, a Transformer-based generative model that learns disentangled syntax and semantics without explicit syntactic or semantic supervision, relying solely on the inductive bias of attention mechanisms.

Contribution

QKVAE is the first model to achieve unsupervised disentanglement of syntax and semantics using only Transformer attention biases, without requiring parse trees or paraphrase pairs.

Findings

QKVAE demonstrates clear disentanglement of syntax and semantics.

The model achieves competitive syntax transfer performance.

Supervised models need over 50K samples to outperform QKVAE.

Abstract

We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available.