Abstract
Recent research on the multi-head attention mechanism, especially that in pre-trained modelssuch as BERT, has shown us heuristics and clues in analyzing various aspects of the this http URL most of the research focus on probing tasks or hidden states, previous works have found someprimitive patterns of attention head behavior by heuristic analytical methods, but a more system-atic analysis specific on the attention patterns still remains primitive. In this work, we clearlycluster the attention heatmaps into significantly different patterns through unsupervised cluster-ing on top of a set of proposed features, which corroborates with previous observations. Wefurther study their corresponding functions through analytical study. In addition, our proposedfeatures can be used to explain and calibrate different attention heads in Transformer models.
Abstract (translated)
URL
https://arxiv.org/abs/2011.00943