Jointly Embedding Relations and Mentions for Knowledge Population

Proceedings of Recent Advances in Natural Language Processing, pages 186–191, Hissar, Bulgaria, Sep 7–9 2015. Jointly Embedding Relations and Mentions for Knowledge Population Miao Fan†,‡, Kai Cao‡, Yifan He‡ and Ralph Grishman‡ † CSLT, Division of Technical Innovation and Development, Tsinghua National Laboratory for Information Science and Technology, Tsinghua University, Beijing, 100084, China. ‡Proteus Group, New York University, NY, 10003, U.S.A. fanmiao.cslt.thu@gmail.com, carson529@gmail.com; me@yifanhe.org, grishman@cs.nyu.edu Abstract This paper contributes a joint embedding model for predicting relations between a pair of entities in the scenario of rela- tion inference. It differs from most stand- alone approaches which separately oper- ate on either knowledge bases or free texts. The proposed model simultaneously learn- s low-dimensional vector representation- s for both triplets in knowledge reposito- ries and the mentions of relations in free texts, so that we can leverage the evidence both resources to make more accurate pre- dictions. We use NELL to evaluate the performance of our approach, compared with cutting-edge methods. Results of ex- tensive experiments show that our model achieves significant improvement on rela- tion extraction. 1 Introduction Relation extraction (Bach and Badaskar, 2007; Gr- ishman, 1997; Sarawagi, 2008), which aims at dis- covering the relationships between a pair of en- tities, is a significant research direction for dis- covering more beliefs for knowledge bases. Most stand-alone approaches, however, either use lo- cal graph patterns in knowledge repositories, or extract features from text mentions, to individu- ally help predict relations between two entities. The heterogeneity brings about a gap between structured repositories and unstructured free texts, which spoils the dream of sharing the evidence from both knowledge and natural language. For studies in decades, scientists either com- pete the performance of their methods on the public text datasets such as ACE1 (GuoDong et al., 2005) and MUC2 (Zelenko et al., 2003), or look for effective approaches (Gardner et al., 2013; Lao et al., 2011) on improving the accu- racy of link prediction within knowledge bases such as NELL3 (Carlson et al., 2010) and Free- base4 (Bollacker et al., 2007). Thanks to the research of distantly supervised relation extrac- tion (Fan et al., 2014a; Mintz et al., 2009) which facilitates the manual annotation via auto- matically aligning with the relation mentions in free texts, NELL can not only extract triplets, i.e. ⟨head entity, relation, tail entity⟩, but al- so collect the texts between two entities as the evidence of relation mention. We take an exam- ple from NELL which originally records a be- lief: ⟨concept : city : caroline, concept : citylocatedinstate, concept : stateorprovince : maryland, County and State of⟩, where “County and State of” is the mention between the head entity concept : city : caroline, and the tail entity concept : stateorprovince : maryland, to indicate the relation concept : citylocatedinstate. Fortunately, the embedding techniques (Fan et al., 2014b; Mikolov et al., 2013) enlighten us to break through the limitation of heterogeneous re- sources, and to establish a connection between a relation and its corresponding mention via learn- …

Content truncated to 500 words