Contrastive heterogeneous graphs learning for multi-hop machine reading comprehension

Machine reading comprehension (MRC) enables a machine to find from documents the answer to a given question. The task is challenging when there is a so-called reasoning process among several documents before eventually arriving at the answer. We investigate multi-hop MRC in the following formulation: given a question in the form of the triplet with a missing entity, along with a collection of supporting documents, to choose an answer to the question from a set of candidates. In order to handle the problem, we propose to leverage selected sentences, as well as candidates and entities in the supporting documents to construct a heterogeneous graph, on which graph learning and reasoning using graph neural networks are performed, followed by an answer prediction layer. To better differentiate candidates with the answer, we come up with a novel contrastive learning module over the heterogeneous graph such that the produced representations of candidates and the answer are more distinguishable. The overall model is learned under a multi-task learning framework by taking both of the losses of heterogeneous graph learning and contrastive learning into consideration. The new model, namely, CHGL is superior to the competing methods on benchmark datasets in experiments.