Probing with Noise: Unpicking the Warp and Weft of Embeddings

TL;DR

This paper introduces a noise-based probing method to analyze how vector norms encode linguistic information in embeddings, revealing that different models use norms to store various types of information.

Contribution

It extends the probing framework with noise-based analysis to interpret the role of vector norms in encoding linguistic features in embeddings.

Findings

GloVe encodes syntactic and sentence length information in the vector norm.

BERT uses the vector norm to encode contextual incongruity.

The method confirms the existence of separate information containers in embeddings.

Abstract

Improving our understanding of how information is encoded in vector space can yield valuable interpretability insights. Alongside vector dimensions, we argue that it is possible for the vector norm to also carry linguistic information. We develop a method to test this: an extension of the probing framework which allows for relative intrinsic interpretations of probing results. It relies on introducing noise that ablates information encoded in embeddings, grounded in random baselines and confidence intervals. We apply the method to well-established probing tasks and find evidence that confirms the existence of separate information containers in English GloVe and BERT embeddings. Our correlation analysis aligns with the experimental findings that different encoders use the norm to encode different kinds of information: GloVe stores syntactic and sentence length information in the vector norm, while BERT uses it to encode contextual incongruity.