Topic-aware multi-hop machine reading comprehension using weighted graphs

The problem of Machine Reading Comprehension (MRC) aims to answer a question based on a natural language context. Multi-hop MRC is a challenging task as it requires a deep comprehension and reasoning of disjoint pieces of information to find the answer. Recently, graph-based methods have become very popular in multi-hop MRC as they well model the problem and ease the reasoning task. However, most existing studies ignore some valuable information of the context. As a result, they focus on partial information instead of covering the full information of the context. To fill this gap, a new approach is presented in this study for the graph-based multi-hop MRC that takes more important information of the context into consideration, including the topic of sentences, the topic of relationships, and the importance and strength of relationships to generate and reason an enrich weighted graph. Several experiments have been conducted to demonstrate the usefulness of the proposed approach. Experiments on the HotpotQA benchmark show that our proposed approach has achieved the new state-of-the-art results.