Start and End Interactions in Bidirectional Attention Flow for Reading Comprehension

The reading comprehension machine learning task involves reading in a question and returning an answer from an associated context paragraph. This task has proven to be difficult, as the performance of state-of-the-art models still do not compare with human performance. The difficulty of the tasks comes from understanding two separate pieces of information as well as the relationship between them before picking a response from the context. Recently, machine learning models have used attention to extend other natural language processing tools to reading comprehension, and among these methods is the Bidirectional Attention Flow (BiDAF) model, which was first introduced by Seo et al. [1]. In this paper, we present experimental results from variations of the default BiDAF model.