Are Representations Built from the Ground Up? An Empirical Examination of Local Composition in Language Models

TL;DR

This paper investigates whether modern language models can accurately represent the compositional and non-compositional meanings of phrases, revealing limitations in their ability to distinguish and generate such representations.

Contribution

It introduces a method to predict phrase representations from their parts and analyzes the models' ability to handle compositionality, highlighting gaps in current language models.

Findings

Representation of parent phrases can be predicted from children with some accuracy.

Predictive accuracy does not align well with human judgments of compositionality.

Language models often struggle to distinguish between compositional and non-compositional phrases.

Abstract

Compositionality, the phenomenon where the meaning of a phrase can be derived from its constituent parts, is a hallmark of human language. At the same time, many phrases are non-compositional, carrying a meaning beyond that of each part in isolation. Representing both of these types of phrases is critical for language understanding, but it is an open question whether modern language models (LMs) learn to do so; in this work we examine this question. We first formulate a problem of predicting the LM-internal representations of longer phrases given those of their constituents. We find that the representation of a parent phrase can be predicted with some accuracy given an affine transformation of its children. While we would expect the predictive accuracy to correlate with human judgments of semantic compositionality, we find this is largely not the case, indicating that LMs may not accurately distinguish between compositional and non-compositional phrases. We perform a variety of analyses, shedding light on when different varieties of LMs do and do not generate compositional representations, and discuss implications for future modeling work.