Multihop Question Answering by Using Sequential Path Expansion With Backtracking

Abstract

Multi-hop question answering from knowledge graphs has gained a growing attention in the past few years due to its vast applications in AI. Existing approaches in this regard mostly rely on exhaustively traversing all paths in the graph until reaching a candidate answer entity, leading to high time complexity and reasoning errors when training the model. Recently, a few works have tried to reduce the search space by proposing sequential-decision techniques so that only the path that is likely to lead to the answer is traced. However, the sequential decision on the relations is likely to accumulate errors due to the potential incorrect matching with the question. Alternatively, this work proposes a method that leverages sequence matching and a backtracking algorithm to identify the correct path to the answer while minimizing the accumulated error. The process starts from the question entity and grows the path iteratively by reasoning over the outgoing relations. The aim is to find the relation with the highest similarity to the question and to transit through it to the next entity. Meanwhile, a termination/backtracking check is performed at each iteration to validate the path and act accordingly either by proceeding to the next entity, stopping the process, or backtracking to correct a wrongly expanded path. The proposed method does not require setting a maximum number of hops in advance and uses a compare-aggregate model with attention mechanism for effective sequence matching. Experimental results show that our method significantly outperforms existing methods on three benchmarking datasets.