1. INTRODUCTION
The ability to retrieve and reason about world knowledge is the core of question answering (QA) [1], [2], [3]. Prevailing textual QA tasks focus on answering a question given a large collection of documents, each question-answer pair involves long passages (e.g. SQuAD[4], HotpotQA[5], trivalQA[6]) without order. Recent work propose NLDB [7], that similar to database queries and each query requires reasoning over hundreds or thousands of facts. In NLDB task, models are required to perform detailed and specific reasoning over fine-grained knowledge facts to answer complex queries such as join, max, set et al. To take a closer step of knowledge reasoning, ReasonchainQA dataset [8] introduce explicit reasoning chain for each query, and the explicit reasoning chains depths vary from 1 to 7 with 12 inference types and much more correlative relations. As is shown in Fig. 1, model needs retrieve evidence set {4, 1, 7} from a database and construct reasoning chains 4 → 1 → 7 (length =3). The reasoning evidence chains can be used to explain why the answer "Polish" is correct.