Learning to Map Natural Language Statements into Knowledge Base Representations for Knowledge Base Construction

Learning to Map Natural Language Statements into Knowledge Base Representations for Knowledge Base Construction Chin-Ho Lin, Hen-Hsen Huang, and Hsin-Hsi Chen Department of Computer Science and Information Engineering National Taiwan University, Taipei, Taiwan {chlin, hhhuang}@nlg.csie.ntu.edu.tw, hhchen@ntu.edu.tw Abstract Directly adding the knowledge triples obtained from open information extraction systems into a knowledge base is often impractical due to a vocabulary gap between natural language (NL) expressions and knowledge base (KB) representation. This paper aims at learning to map relational phrases in triples from natural-language-like statement to knowledge base predicate format. We train a word representation model on a vector space and link each NL relational pattern to the semantically equivalent KB predicate. Our mapping result shows not only high quality, but also promising coverage on relational phrases compared to previous research. Keywords: knowledge base, knolwedge base construction, relation mapping 1. Introduction Knowledge bases (KBs) such as Freebase (Bollacker et al., 2008) and DBpedia (Auer et al., 2007) are fundamental resources for many intelligent applications. Currently, the construction and updating of KBs, directly or indirectly, rely on human labor. Keeping KBs up-to-date by humans is cost intensive and impractical. To reduce the cost and to minimize the updating latency, automatically updating a KB with knowledge extracted from natural language (NL) content is a feasible strategy. In KBs, a fact is represented by a triple (subject, predicate, object), where subject and object are two entities in the KB, and predicate describes their relation. With an information extraction system (Carlson et al., 2010; Fader et al., 2011), we can also extract facts from NL text in the format (np1, pattern, np2), where np1 and np2 are two entities and pattern is the relational phrase between them. However, the extracted triples from NL text do not always follow the paradigm of KB, and that becomes a challenging issue for KB construction. For example, the NL triple (Garnett, was born in, Mauldin) extracted by ReVerb (Fader et al., 2011) shows a fact identical to the KB triple (Kevin Garnett, birthPlace, Mauldin (South Carolina)). Although these two triples state the same fact, there is a vocabulary gap between them. In the former triple, “was born in” is an NL- like expression and “birthPlace” in the latter triple is a formatted predicate used in KB. These two relational phrases are different in surface forms. They cannot be mapped by string matching directly. In addition, a KB predicate may be described in multiple NL statements. A number of ReVerb patterns such as “is the hometown of”, “was raised in”, and “grew up in” are related to the predicate “hometown” in DBpedia. That makes the mapping between KB and NL even more challenging. Recently, more and more works show their interests in the issue of KB construction. Knowledge graph embedding models (Bordes et al., 2013; Yang et al., 2015; Xie et al., 2016a; Xie et al., 2016b) focus on learning the vector representation on the KB side only. Previous works (Nakashole et al., 2012; Riedel et al., 2013; Dutta et …

Content truncated to 500 words