S2QL: Retrieval Augmented Zero-Shot Question Answering over Knowledge Graph

Zan Daoguang, Sirui Wang, Hongzhi Zhang, Yuanmeng Yan, Wei Wu, Bei Guan, Yongji Wang

Abstract: Knowledge Graph Question Answering (KGQA) is a challenging task that aims to obtain the entities from the given Knowledge Graph (KG) to answer the user’s natural language questions. Most existing studies are focused on the traditional KGQA task, where the test distribution is the same as the training distribution over questions. In contrast, few efforts have been made to explore the zero-shot KGQA task. Logically, the existing models for the traditional KGQA task naturally show poor performance on the zero-shot setting. It is a non-trivial task to migrate the off-the-shelf zero-shot solutions in other common tasks to KGQA since an intrinsical gap exists between other common tasks and the KGQA task under the zero-shot settings. Furthermore, we observed that Similar Questions tend to have Similar Logic forms. Motivated by this, we propose a simple yet effective framework S $$^2$$ QL. In detail, we first elaborately devise three similarity measurement units to category the user’s questions. Then based on the Similarity Relation Graph (SRG) constructed by the above similarity measurement units, we devise a retrieval augmented strategy to further answer arduous zero-shot questions with its retrieved similar questions. Extensive experiments on the GrailQA and WebQSP benchmarks demonstrate that our approach is more effective than a number of competitive KGQA baselines on the zero-shot setting.