Abstract
Transformers have obtained significant success modeling natural language as a sequence of text tokens. However, in many real world scenarios, textual data inherently exhibits structures beyond a linear sequence such as trees and graphs; an important one being multi-hop question answering, where evidence required to answer questions are scattered across multiple related documents. This paper presents Transformer-XH, which uses eXtra Hop attention to enable the intrinsic modeling of structured texts in a fully data-driven way. Its new attention mechanism naturally “hops” across the connected text sequences in addition to attending over tokens within each sequence. Thus, Transformer-XH better answers multi-hop questions by propagating information between multiple documents, constructing global contextualized representations, and jointly reasoning over multiple pieces of evidence. This leads to a simpler multi-hop QA system which outperforms previous state-of-the-art on the HotpotQA FullWiki setting by large margins.