(A) hierarchical SRL graph network for multi-hop question answering

Abstract

Multi-hop question answering requires the aggregation of information from several documents to find the answer to a question. Most prominent works approach this aggregation through entity graphs. However, they tend to overlook intra-sentence reasoning. In this work, we propose a graph structure whose main innovation is the use of semantic role labeling (SRL) arguments to explicitly model all the multi-hop reasoning steps, including the intra-sentence reasoning. Additionally, we propose a novel hierarchical graph2seq mechanism to fuse multi-hop and entity boundary information from the graph into the token embeddings of the context to enhance the answer span prediction task. We achieve competitive performance compared to the current state of the art and prove through extensive qualitative and quantitative experiments the effectiveness of SRL to model multi-hop reasoning, as well as the capabilities of our hierarchical graph2seq mechanism, which outperforms all previous approaches, to fuse graph information into the token embeddings.