Positional Artefacts Propagate Through Masked Language Model Embeddings

TL;DR

This paper investigates positional artefacts in masked language models, revealing outlier neurons related to positional embeddings that affect vector anisotropy and can be mitigated to improve semantic tasks without harming performance.

Contribution

The study identifies positional artefacts in BERT and RoBERTa, introduces a neuron-level analysis method, and shows that removing outliers enhances embedding quality without degrading task performance.

Findings

Outliers are linked to positional embeddings.

Removing outliers improves word sense discrimination.

Clipping outliers enhances sentence embedding quality.

Abstract

In this work, we demonstrate that the contextualized word vectors derived from pretrained masked language model-based encoders share a common, perhaps undesirable pattern across layers. Namely, we find cases of persistent outlier neurons within BERT and RoBERTa's hidden state vectors that consistently bear the smallest or largest values in said vectors. In an attempt to investigate the source of this information, we introduce a neuron-level analysis method, which reveals that the outliers are closely related to information captured by positional embeddings. We also pre-train the RoBERTa-base models from scratch and find that the outliers disappear without using positional embeddings. These outliers, we find, are the major cause of anisotropy of encoders' raw vector spaces, and clipping them leads to increased similarity across vectors. We demonstrate this in practice by showing that clipped vectors can more accurately distinguish word senses, as well as lead to better sentence embeddings when mean pooling. In three supervised tasks, we find that clipping does not affect the performance.