Whatcha lookin' at? DeepLIFTing BERT's Attention in Question Answering

TL;DR

This paper uses DeepLIFT to analyze BERT's attention mechanisms in question answering, revealing patterns that may explain its superior performance and similarities to human reasoning.

Contribution

It introduces a method to interpret BERT's attention in question answering by applying DeepLIFT and analyzing attention shifts and patterns.

Findings

Attention shifts correlate with question types

Patterns resemble aspects of human reasoning

Provides insights into BERT's decision process

Abstract

There has been great success recently in tackling challenging NLP tasks by neural networks which have been pre-trained and fine-tuned on large amounts of task data. In this paper, we investigate one such model, BERT for question-answering, with the aim to analyze why it is able to achieve significantly better results than other models. We run DeepLIFT on the model predictions and test the outcomes to monitor shift in the attention values for input. We also cluster the results to analyze any possible patterns similar to human reasoning depending on the kind of input paragraph and question the model is trying to answer.