Attention Flows: Analyzing and Comparing Attention Mechanisms in Language Models

Advances in language modeling have led to the development of deep attention-based models that are performant across a wide variety natural processing (NLP) problems. These typified by pre-training process on large unlabeled text corpora and subsequently fine-tuned for specific tasks. Although considerable work has been devoted understanding attention mechanisms pre-trained models, it is less understood how model's change when trained target NLP task. In this paper, we propose visual analytics approach fine-tuning models. Our visualization, Attention Flows, designed support users querying, tracing, comparing within layers, amongst heads Transformer-based To help gain insight classification decision made, our design centered depicting classification-based at deepest layer from prior layers flows throughout words input. Flows supports analysis single model, as well comparison between via their similarities differences. We use study various sentence tasks highlight evolves address nuances solving these