Syntax-guided Neural Module Distillation to Probe Compositionality in Sentence Embeddings

TL;DR

This paper introduces a syntax-guided neural module distillation method to investigate the causal role of syntax in sentence embedding models, revealing that many models' compositionality is largely linear and syntax plays a significant role.

Contribution

It proposes a novel neural module distillation approach based on syntax parse trees to probe the causal impact of syntax on sentence embeddings, and analyzes the geometry and linearity of semantic composition.

Findings

Distillability correlates with model performance.

Distillability does not vary much with model size.

Syntax-guided composition is largely linear.

Abstract

Past work probing compositionality in sentence embedding models faces issues determining the causal impact of implicit syntax representations. Given a sentence, we construct a neural module net based on its syntax parse and train it end-to-end to approximate the sentence's embedding generated by a transformer model. The distillability of a transformer to a Syntactic NeurAl Module Net (SynNaMoN) then captures whether syntax is a strong causal model of its compositional ability. Furthermore, we address questions about the geometry of semantic composition by specifying individual SynNaMoN modules' internal architecture & linearity. We find differences in the distillability of various sentence embedding models that broadly correlate with their performance, but observe that distillability doesn't considerably vary by model size. We also present preliminary evidence that much syntax-guided composition in sentence embedding models is linear, and that non-linearities may serve primarily to handle non-compositional phrases.