Commonsense Knowledge Base Completion and Generation

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 141–150 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 141 Commonsense Knowledge Base Completion and Generation Itsumi Saito Kyosuke Nishida Hisako Asano NTT Media Intelligence LaboratoriesJunji Tomita {saito.itsumi, nishida.kyosuke}@lab.ntt.co.jp, {asano.hisako, tomita.junji}@lab.ntt.co.jp Abstract This study focuses on acquisition of common- sense knowledge. A previous study proposed a commonsense knowledge base completion (CKB completion) method that predicts a con- fidence score of triplet-style knowledge for improving the coverage of CKBs. To im- prove the accuracy of CKB completion and ex- pand the size of CKBs, we formulate a new commonsense knowledge base generation task (CKB generation) and propose a joint learning method that incorporates both CKB comple- tion and CKB generation. Experimental re- sults show that the joint learning method im- proved completion accuracy and the genera- tion model created reasonable knowledge. Our generation model could also be used to aug- ment data and improve the accuracy of com- pletion. 1 Introduction Knowledge bases (KBs) are a kind of information network, and they have been applied to many nat- ural language processing tasks such as question answering (Yang and Mitchell, 2017; Long et al., 2017) and dialog tasks (Young et al., 2018). In this paper, we focus on commonsense knowledge bases (CKBs). Commonsense knowledge is also referred to as background knowledge and is used in natural language application tasks that require reasoning based on implicit knowledge. For exam- ple, machine comprehension tasks that need com- monsense reasoning have been proposed very re- cently (Lin et al., 2017; Ostermann et al., 2018). In particular, Wang et al. (2018) used commonsense knowledge provided by ConceptNet (Speer et al., 2017) to efficiently resolve ambiguities and infer implicit information. Information in CKB is represented in RDF- style triples ⟨t1, r, t2⟩, where t1 and t2 are ar- bitrary words or phrases, and r ∈ R is a relation between t1 and t2. For example, ⟨go to restaurant, subevent, order food⟩ means “order food” happens as a subevent of “go to restaurant”. Although researchers have devel- oped techniques for acquiring CKB from raw text with patterns (Angeli and Manning, 2013), it has been pointed out that some sorts of knowledge are rarely expressed explicitly in textual corpora (Gordon and Van Durme, 2013). Therefore, re- searchers have developed curated CKB resources by manual annotation (Speer et al., 2017). While manually created knowledge has high precision, these resources suffer from lack of coverage. Knowledge base completion methods are used to improve the coverage of existing general- purpose KBs, such as Freebase (Bollacker et al., 2008; Bordes et al., 2013; Lin et al., 2015). For example, given a node pair ⟨Athens, Greece⟩, a completion method predicts the missing rela- tion “IsLocatedIn”. Such KBs consist of well- connected entities; thus, the completion meth- ods are mainly used to find missing links of the existing nodes. On the other hand, CKBs are very sparse because their nodes contain arbitrary phrases and it is difficult …

Content truncated to 500 words