Relaxing Unanswerable Geographic Questions Using A Spatially Explicit Knowledge Graph Embedding Model.
Lecture Notes in Geoinformation and Cartography(2020)
摘要
Recent years have witnessed a rapid increase in Question Answering (QA) research and products in both academic and industry. However, geographic question answering remained nearly untouched although geographic questions account for a substantial part of daily communication. Compared to general QA systems, geographic QA has its own uniqueness, one of which can be seen during the process of handling unanswerable questions. Since users typically focus on the geographic constraints when they ask questions, if the question is unanswerable based on the knowledge base used by a QA system, users should be provided with a relaxed query which takes distance decay into account during the query relaxation and rewriting process. In this work, we present a spatially explicit translational knowledge graph embedding model called TransGeo which utilizes an edge-weighted PageRank and sampling strategy to encode the distance decay into the embedding model training process. This embedding model is further applied to relax and rewrite unanswerable geographic questions. We carry out two evaluation tasks: link prediction as well as query relaxation/rewriting for an approximate answer prediction task. A geographic knowledge graph training/testing dataset, DB18, as well as an unanswerable geographic query dataset, GeoUQ, are constructed. Compared to four other baseline models, our TransGeo model shows substantial advantages in both tasks.
更多查看译文
关键词
Geographic question answering,Query relaxation,Knowledge graph embedding,Spatially explicit model
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络