Accurate Text-Enhanced Knowledge Graph Representation Learning

Proceedings of NAACL-HLT 2018, pages 745–755 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Accurate Text-Enhanced Knowledge Graph Representation Learning Bo An1,2, Bo Chen1,2, Xianpei Han1, Le Sun1 1State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China {anbo, chenbo, xianpei, sunle}@iscas.ac.cn Abstract Previous representation learning techniques for knowledge graph representation usually represent the same entity or relation in differ- ent triples with the same representation, with- out considering the ambiguity of relations and entities. To appropriately handle the semantic variety of entities/relations in distinct triples, we propose an accurate text-enhanced knowl- edge graph representation learning method, which can represent a relation/entity with dif- ferent representations in different triples by ex- ploiting additional textual information. Specif- ically, our method enhances representations by exploiting the entity descriptions and triple- specific relation mention. And a mutual atten- tion mechanism between relation mention and entity description is proposed to learn more accurate textual representations for further improving knowledge graph representation. Experimental results show that our method achieves the state-of-the-art performance on both link prediction and triple classification tasks, and significantly outperforms previous text-enhanced knowledge representation mod- els. 1 Introduction Knowledge graphs such as Freebase (Bollacker et al., 2008), YAGO (Suchanek et al., 2007) and WordNet (Miller, 1995) are among the most widely used resources in NLP applications. Typi- cally, a knowledge graph consists of a set of triples {(h, r, t)}, where h, r, t stand for head entity, rela- tion and tail entity respectively. Learning distributional representation of knowl- edge graph has attracted many research attentions in recent years. By projecting all elements in a knowledge graph into a dense vector space, the se- mantic distance between all elements can be eas- ily calculated, and thus enables many applications Figure 1: A demonstration of our accurate text- enhanced model. The meanings of relation parentOf in different triples are distinguished by their entity descriptions, and the relation parentOf emphasizes words which describe its social relationship in entity descriptions. such as link prediction and triple classification (Socher et al., 2013). Recently, translation-based models, including TransE (Bordes et al., 2013), TransH (Wang et al., 2014), TransD (Ji et al., 2015) and TransR (Lin et al., 2015b), have achieved promising results in distributional representation learning of knowl- edge graph. ComplEx (Trouillon et al., 2016) has achieved the state-of-the-art performance on multiple tasks, such as triple classification and link prediction. Unfortunately, all of these meth- ods only utilize the structure information of knowledge graph, which inevitably suffer from the sparseness and incompleteness of knowledge graph. Even worse, structure information usually cannot distinguish the different meanings of rela- tions and entities in different triples. To address the above problem, additional in- formation is introduced to enrich the knowledge representations, including entity types and logic rules. However, most researches of this line are limited by manually constructed logic rules, which 745 are knowledge graph sensitive and require the ex- pert knowledge. …

Content truncated to 500 words