Abstract
Commonsense explanation generation aims to empower the machine's sense-makingcapability by generating plausible explanations to statements againstcommonsense. While this task is easy to human, the machine still struggles togenerate reasonable and informative explanations. In this work, we propose amethod that first extracts the underlying concepts which are served as\textit{bridges} in the reasoning chain and then integrates these concepts togenerate the final explanation. To facilitate the reasoning process, we utilizeexternal commonsense knowledge to build the connection between a statement andthe bridge concepts by extracting and pruning multi-hop paths to build asubgraph. We design a bridge concept extraction model that first scores thetriples, routes the paths in the subgraph, and further selects bridge conceptswith weak supervision at both the triple level and the concept level. Weconduct experiments on the commonsense explanation generation task and ourmodel outperforms the state-of-the-art baselines in both automatic and humanevaluation.