The APVA-TURBO Approach To Question Answering in Knowledge Base

Proceedings of the 27th International Conference on Computational Linguistics, pages 1998–2009 Santa Fe, New Mexico, USA, August 20-26, 2018. 1998 The APVA-TURBO Approach To Question Answering in Knowledge Base Yue Wang†, Richong Zhang†∗, Cheng Xu† and Yongyi Mao‡ †BDBC and SKLSDE, School of Computer Science and Engineering, Beihang University, China ‡School of Electrical Engineering and Computer Science, University of Ottawa, Canada †{wangyue16, zhangrc, xucheng}@act.buaa.edu.cn ‡yymao@site.uottawa.ca Abstract In this paper, we study the problem of question answering over knowledge base. We identify that the primary bottleneck in this problem is the difficulty in accurately predicting the relations connecting the subject entity to the object entities. We advocate a new model architecture, APVA, which includes a verification mechanism responsible for checking the correctness of predicted relations. The APVA framework naturally supports a well-principled iterative training procedure, which we call turbo training. We demonstrate via experiments that the APVA-TUBRO approach drastically improves the question answering performance. Title and Abstract in Chinese 面向知识库问答的 APVA-TURBO 方法 本文主要围绕目前基于知识库的问答方法中的存在的问题展开研究。经过对现有各类问 答方法的调查与分析，我们发现目前知识库问答的瓶颈主要体现在关系预测上，即如何 准确地预测出问题中的实体与答案实体之间的关联，是需要面临的最大挑战。因此本 文在现有问答框架的基础上，加入了负责对预测关系的可靠性进行评价检验的验证机 制，用以增强关系预测效果，并提出了一种新的知识库问答框架“APVA”。另外，文中 为 APVA 设计了一种具有良好理论基础的迭代训练过程，我们称之为“涡轮式(turbo)训 练”。通过实验证明，APVA-TUBRO 方法可以在问答数据集上取得优异效果，大大提升 了目前问答方法的准确性。 1 Introduction Knowledge bases (KBs), such as YAGO, DBPedia, and Freebase contain an enormous volume of facts, or knowledge, about the real world. Each of these facts is represented as a triple (s, r, o), where s is a subject entity, r is a relation, and o is an object entity. This representation allows the KB to be interpreted as a graph in which entities are vertices, triples are edges, and relations are edge labels. Such a structured representation has enabled the development of query languages, such as RQL(Karvounarakis et al., 2002), Versa(Ogbuji, 2005) and SPARQL(Prud et al., 2006), for retrieving knowledge stored in the KB. Despite the usefulness of these languages, it is highly desirable that the query-processing engine of a KB is capable of processing queries presented in human language directly. This stimulates intense research in question answering over KB. Briefly, question answering over KB, or QA-KB, aims at taking as input a question presented in a natural language and outputs the answer in terms of a fact triple or a collection of fact triples in the KB. In a general setting, the answer is a set of paths, where each path connects the subject entity of the question to the object entity via a sequence of relations. We refer to this relation sequence as a path label. There are two main lines of solutions to QA-KB. The first centres around semantic parsing, which translates the natural-language question into a logical form and uses it to query the KB. This line of ∗Corresponding Author: Richong Zhang This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 1999 solutions includes, for example, (Berant et al., 2013; Dong et al., 2017; Yih et al., 2015; Xu et al., 2016; Reddy et al., 2017). The second line of solutions centres around learning with neural network models, which follows what we call an APA architecture or its variants. Specifically, the APA architecture includes three …

Content truncated to 500 words