An Interpretability Illusion for BERT
This highlights a methodological pitfall for interpretability research in NLP, emphasizing the need for multi-dataset testing.
The paper identifies an 'interpretability illusion' in BERT where neuron activations appear to encode simple concepts but actually represent complex patterns, attributing this to geometric properties of embeddings and limited text corpora.
We describe an "interpretability illusion" that arises when analyzing the BERT model. Activations of individual neurons in the network may spuriously appear to encode a single, simple concept, when in fact they are encoding something far more complex. The same effect holds for linear combinations of activations. We trace the source of this illusion to geometric properties of BERT's embedding space as well as the fact that common text corpora represent only narrow slices of possible English sentences. We provide a taxonomy of model-learned concepts and discuss methodological implications for interpretability research, especially the importance of testing hypotheses on multiple data sets.