How do Transformer Embeddings Represent Compositions? A Functional Analysis

TL;DR

This paper investigates how transformer-based language models represent compositional structures, revealing that most models are highly compositional, with BERT showing weaker compositionality, through a comprehensive analysis of various models and methods.

Contribution

It provides a comparative analysis of compositionality in different transformer models, highlighting the effectiveness of simple models like addition and regression in capturing compositional representations.

Findings

Ridge regression best models compositionality among tested methods.

Most embedding models are highly compositional, except BERT.

Vector addition performs nearly as well as more complex models.

Abstract

Compositionality is a key aspect of human intelligence, essential for reasoning and generalization. While transformer-based models have become the de facto standard for many language modeling tasks, little is known about how they represent compound words, and whether these representations are compositional. In this study, we test compositionality in Mistral, OpenAI Large, and Google embedding models, and compare them with BERT. First, we evaluate compositionality in the representations by examining six diverse models of compositionality (addition, multiplication, dilation, regression, etc.). We find that ridge regression, albeit linear, best accounts for compositionality. Surprisingly, we find that the classic vector addition model performs almost as well as any other model. Next, we verify that most embedding models are highly compositional, while BERT shows much poorer compositionality. We verify and visualize our findings with a synthetic dataset consisting of fully transparent adjective-noun compositions. Overall, we present a thorough investigation of compositionality.