CAKE: A Scalable Commonsense-Aware Framework For Multi-View Knowledge Graph Completion

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2867 - 2877 May 22-27, 2022 c⃝2022 Association for Computational Linguistics CAKE: A Scalable Commonsense-Aware Framework For Multi-View Knowledge Graph Completion Guanglin Niu1,2, Bo Li2,3, Yongfei Zhang1,4∗, Shiliang Pu5 1 Beijing Key Laboratory of Digital Media, Beihang University, Beijing, China 2 Institute of Artificial Intelligence, Beihang University, Beijing, China 3 Hangzhou Innovation Institute, Beihang University, Hangzhou, China 4 State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China 5 Hikvision Research Institute, Hangzhou, China {beihangngl, boli, yfzhang}@buaa.edu.cn, pushiliang.hri@hikvision.com Abstract Knowledge graphs store a large number of fac- tual triples while they are still incomplete, in- evitably. The previous knowledge graph com- pletion (KGC) models predict missing links be- tween entities merely relying on fact-view data, ignoring the valuable commonsense knowl- edge. The previous knowledge graph embed- ding (KGE) techniques suffer from invalid neg- ative sampling and the uncertainty of fact-view link prediction, limiting KGC’s performance. To address the above challenges, we propose a novel and scalable Commonsense-Aware Knowledge Embedding (CAKE) framework to automatically extract commonsense from factual triples with entity concepts. The gen- erated commonsense augments effective self- supervision to facilitate both high-quality nega- tive sampling (NS) and joint commonsense and fact-view link prediction. Experimental results1 on the KGC task demonstrate that assembling our framework could enhance the performance of the original KGE models, and the proposed commonsense-aware NS module is superior to other NS techniques. Besides, our proposed framework could be easily adaptive to various KGE models and explain the predicted results. 1 Introduction In recent years, knowledge graphs (KGs) such as Freebase (Bollacker et al., 2008), DBpe- dia (Lehmann et al., 2015) and NELL (Mitchell et al., 2018) have been widely used in many knowledge-intensive applications, including ques- tion answering (Sun et al., 2020; Saxena et al., 2020), dialogue systems (Yang et al., 2020; Zhou et al., 2018) and recommender systems (Wang et al., 2021, 2019a). However, the KGs constructed man- ually or automatically are inevitably incomplete, requiring KGC to infer new facts. ∗Corresponding author. 1The source code and datasets of this paper are available at https://github.com/ngl567/CAKE. (Los Angeles, LocatedIn, California) (Los Angeles, LocatedIn, California) Positive triple: High-quality negative triple: False-negative triples: (Los Angeles, LocatedIn, Georgia) (Los Angeles, LocatedIn, Georgia) (San Francisco, LocatedIn, California) (San Francisco, LocatedIn, California) Low-quality negative triples: (Los Angeles, LocatedIn, Apple Pie) (Los Angeles, LocatedIn, Apple Pie) Challenge 1: Invalid Negative Sampling Negative sampling (Apple Inc., LocatedIn, California) (Apple Inc., LocatedIn, California) (English, LocatedIn, California) (English, LocatedIn, California) Challenge 2: Uncertainty of Fact-View Link Prediction (David, Nationality, ?) (David, Nationality, ?) Query: (David, Nationality, ?) Query: (David, Nationality, U.S.A.) (David, Nationality, U.S.A.) Correct triple: (David, Nationality, U.S.A.) Correct triple: David U.S.A. California Rank: California 1 California 1 U.S.A. 2 3 Commonsense Canada ... Unsatisfy Satisfy Satisfy ||eDavid+rNationality-eCalifornia|| < ||eDavid+rNationality-eU.S.A.|| Country belongs to Figure 1: Two examples exhibit the challenges that needed to be addressed. Challenge 1: Given a positive triple, some generated negative triples …

Content truncated to 500 words