Frustratingly Easy Meta-Embedding -- Computing Meta-Embeddings by Averaging Source Word Embeddings

TL;DR

This paper demonstrates that simple averaging of different pre-trained word embeddings can produce effective meta-embeddings, challenging the assumption that complex transformations are necessary for high-quality results.

Contribution

It shows that straightforward averaging of source embeddings can outperform or match complex meta-embedding methods, providing a surprisingly simple yet effective approach.

Findings

Averaging two distinct embeddings yields competitive meta-embeddings.

Simple mean-based meta-embeddings outperform some complex methods.

The approach is effective despite the incommensurability of source vector spaces.

Abstract

Creating accurate meta-embeddings from pre-trained source embeddings has received attention lately. Methods based on global and locally-linear transformation and concatenation have shown to produce accurate meta-embeddings. In this paper, we show that the arithmetic mean of two distinct word embedding sets yields a performant meta-embedding that is comparable or better than more complex meta-embedding learning methods. The result seems counter-intuitive given that vector spaces in different source embeddings are not comparable and cannot be simply averaged. We give insight into why averaging can still produce accurate meta-embedding despite the incomparability of the source vector spaces.