Positional Artefacts Propagate Through Masked Language Model Embeddings

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.

arXiv commentAccepted as long paper at ACL 2021

References in corpus (8)

• Efficient Estimation of Word Representations in Vector Space
• Deep Residual Learning for Image Recognition
• Multilingual Denoising Pre-training for Neural Machine Translation
• DeBERTa: Decoding-enhanced BERT with Disentangled Attention
• Whitening Sentence Representations for Better Semantics and Faster Retrieval
• Visualizing and Measuring the Geometry of BERT
• Does BERT Make Any Sense? Interpretable Word Sense Disambiguation with Contextualized Embeddings
• Rethinking Positional Encoding in Language Pre-training