Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3376–3381, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3376 Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations Xin Lv1,2,3, Yuxian Gu1, Xu Han1,2, Lei Hou1,2,3∗, Juanzi Li1,2,3, Zhiyuan Liu1,2 1Department of Computer Science and Technology, Tsinghua University 2KIRC, Institute for Artificial Intelligence, Tsinghua University 3Beijing National Research Center for Information Science and Technology {lv-x18,gu-yx17,hanxu17}@mails.tsinghua.edu.cn {houlei,lijuanzi,liuzy}@tsinghua.edu.cn Abstract Multi-hop knowledge graph (KG) reasoning is an effective and explainable method for pre- dicting the target entity via reasoning paths in query answering (QA) task. Most pre- vious methods assume that every relation in KGs has enough training triples, regardless of those few-shot relations which cannot pro- vide sufficient triples for training robust rea- soning models. In fact, the performance of existing multi-hop reasoning methods drops significantly on few-shot relations. In this paper, we propose a meta-based multi-hop reasoning method (Meta-KGR), which adopts meta-learning to learn effective meta parame- ters from high-frequency relations that could quickly adapt to few-shot relations. We eval- uate Meta-KGR on two public datasets sam- pled from Freebase and NELL, and the exper- imental results show that Meta-KGR outper- forms the current state-of-the-art methods in few-shot scenarios. Our code and datasets can be obtained from https://github.com/ THU-KEG/MetaKGR. 1 Introduction Recently, large-scale knowledge graphs (KGs) have been demonstrated to be beneficial for many NLP tasks like query answering (QA). A triple query for QA is generally formalized as (es, rq, ?), where es is the source entity and rq is the query relation. For example, given a language query “What is the nationality of Mark Twain?”, we can transform it into (Mark Twain, nationality, ?) and then search the target entity America from KGs as the answer. However, as many KGs are constructed automatically and face serious incom- pleteness problems (Bordes et al., 2013), it is often hard to directly get target entities for queries. To alleviate this issue, some knowledge graph embedding methods (Bordes et al., 2013; ∗ Corresponding Author 0 2000 4000 6000 8000 10000 Relation frequency 0 200 400 600 800 Number Figure 1: The histogram of relation frequency in the real-world knowledge graph Wikidata. Dettmers et al., 2018) have been proposed to em- bed entities and relations into semantic spaces to capture inner connections, and then use the learned embeddings for final predictions. Al- though these embedding-based methods have shown strong abilities in predicting target entities for queries, they only give answers and lack inter- pretability for their decisions (Lin et al., 2018). In order to make models more intuitive, Das et al. (2018) and Lin et al. (2018) propose multi-hop reasoning methods, which leverage the symbolic compositionality of relations in KGs to achieve explainable reasoning results. For example, when queried with (Mark Twain, nationality, ?), multi-hop reasoning models can give not only the target entity America but also multi-hop ex- plainable paths (Mark Twain, bornIn, Florida) …

Content truncated to 500 words