All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality
This addresses a methodological problem for researchers analyzing representational quality in transformer language models, though it is incremental as it builds on existing similarity measures.
The paper identifies that standard similarity measures like cosine similarity are dominated by a few 'rogue dimensions' in contextual language models, which do not reflect actual model behavior, and shows that simple postprocessing can correct this issue.
Similarity measures are a vital tool for understanding how language models represent and process language. Standard representational similarity measures such as cosine similarity and Euclidean distance have been successfully used in static word embedding models to understand how words cluster in semantic space. Recently, these measures have been applied to embeddings from contextualized models such as BERT and GPT-2. In this work, we call into question the informativity of such measures for contextualized language models. We find that a small number of rogue dimensions, often just 1-3, dominate these measures. Moreover, we find a striking mismatch between the dimensions that dominate similarity measures and those which are important to the behavior of the model. We show that simple postprocessing techniques such as standardization are able to correct for rogue dimensions and reveal underlying representational quality. We argue that accounting for rogue dimensions is essential for any similarity-based analysis of contextual language models.