Latent Question Reformulation and Information Accumulation for Multi-Hop Machine Reading

Multi-hop text-based question-answering is a current challenge in machine comprehension. This task requires to sequentially integrate facts from multiple passages to answer complex natural language questions. In this paper, we propose a novel architecture, called the Latent Question Reformulation Network (LQR-net), a multi-hop and parallel attentive network designed for question-answering tasks that require reasoning capabilities. LQR-net is composed of an association of \textbf{reading modules} and \textbf{reformulation modules}. The purpose of the reading module is to produce a question-aware representation of the document. From this document representation, the reformulation module extracts essential elements to calculate an updated representation of the question. This updated question is then passed to the following hop. We evaluate our architecture on the \hotpotqa question-answering dataset designed to assess multi-hop reasoning capabilities. Our model achieves competitive results on the public leaderboard and outperforms the best current \textit{published} models in terms of Exact Match (EM) and $F_1$ score. Finally, we show that an analysis of the sequential reformulations can provide interpretable reasoning paths.