The heads hypothesis: A unifying statistical approach towards understanding multi-headed attention in BERT

TL;DR

This paper introduces a unified statistical framework for analyzing multi-headed attention in BERT, enabling robust, comparable, and significance-aware classification of attention head roles across different contexts.

Contribution

It proposes a generalized score and hypothesis testing method for classifying attention head roles, addressing inconsistencies and statistical significance issues in prior approaches.

Findings

Identifies co-location of multiple roles within the same attention head

Analyzes distribution of attention heads across BERT layers

Examines impact of fine-tuning on attention head roles

Abstract

Multi-headed attention heads are a mainstay in transformer-based models. Different methods have been proposed to classify the role of each attention head based on the relations between tokens which have high pair-wise attention. These roles include syntactic (tokens with some syntactic relation), local (nearby tokens), block (tokens in the same sentence) and delimiter (the special [CLS], [SEP] tokens). There are two main challenges with existing methods for classification: (a) there are no standard scores across studies or across functional roles, and (b) these scores are often average quantities measured across sentences without capturing statistical significance. In this work, we formalize a simple yet effective score that generalizes to all the roles of attention heads and employs hypothesis testing on this score for robust inference. This provides us the right lens to systematically analyze attention heads and confidently comment on many commonly posed questions on analyzing the BERT model. In particular, we comment on the co-location of multiple functional roles in the same attention head, the distribution of attention heads across layers, and effect of fine-tuning for specific NLP tasks on these functional roles.