Representation biases in sentence transformers

TL;DR

This paper investigates the biases of sentence transformers, revealing a strong preference for noun participant overlap over syntactic or predicate similarities in sentence representations.

Contribution

It demonstrates that state-of-the-art sentence transformers are predominantly biased towards noun participant sets, providing insights into their underlying representation properties.

Findings

Sentence transformers show a strong bias towards noun participant overlap.

Syntactic functions of participants are largely irrelevant to representations.

Lexical overlap in noun sets influences cosine similarity more than predicate similarity.

Abstract

Variants of the BERT architecture specialised for producing full-sentence representations often achieve better performance on downstream tasks than sentence embeddings extracted from vanilla BERT. However, there is still little understanding of what properties of inputs determine the properties of such representations. In this study, we construct several sets of sentences with pre-defined lexical and syntactic structures and show that SOTA sentence transformers have a strong nominal-participant-set bias: cosine similarities between pairs of sentences are more strongly determined by the overlap in the set of their noun participants than by having the same predicates, lengthy nominal modifiers, or adjuncts. At the same time, the precise syntactic-thematic functions of the participants are largely irrelevant.