An Entity-centric Approach for Overcoming Knowledge Graph Sparsity

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 530–535, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. An Entity-centric Approach for Overcoming Knowledge Graph Sparsity Manjunath Hegde Indian Institute of Science manjunath@ssl.serc.iisc.in Partha Talukdar Indian Institute of Science ppt@serc.iisc.in Abstract Automatic construction of knowledge graphs (KGs) from unstructured text has received considerable attention in recent research, resulting in the con- struction of several KGs with millions of entities (nodes) and facts (edges) among them. Unfortunately, such KGs tend to be severely sparse in terms of number of facts known for a given en- tity, i.e., have low knowledge density. For example, the NELL KG consists of only 1.34 facts per entity. Unfor- tunately, such low knowledge density makes it challenging to use such KGs in real-world applications. In contrast to best-effort extraction paradigms fol- lowed in the construction of such KGs, in this paper we argue in favor of ENTIty Centric Expansion (ENTICE), an entity-centric KG population frame- work, to alleviate the low knowledge density problem in existing KGs. By using ENTICE, we are able to increase NELL’s knowledge density by a factor of 7.7 at 75.5% accuracy. Additionally, we are also able to extend the ontology discovering new relations and entities. 1 Introduction Over the last few years, automatic construc- tion of knowledge graphs (KGs) from web- scale text data has received considerable at- tention, resulting in the construction of sev- eral large KGs such as NELL (Mitchell et al., 2015), Google’s Knowledge Vault (Dong et al., 2014). These KGs consist of millions of en- tities and facts involving them. While mea- suring size of the KGs in terms of number of entities and facts is helpful, they don’t read- ily capture the volume of knowledge needed in Known Target Entity New Target Entity Known Relation KR-KE KR-NE New Relation NR-KE NR-NE Table 1: Any new fact involving a source en- tity from a Knowledge Graph (i.e., facts of the form entity1-relation-entity2 where entity1 is already in the KG) can be classified into one of the four extraction classes shown above. Most KG population techniques tend to focus on ex- tracting facts of the KR-KE class. ENTICE, the entity-centric approach proposed in this paper, is able to extract facts of all four classes. real-world applications. When such a KG is used in an application, one is often interested in known facts for a given entity, and not nec- essarily the overall size of the KG. In particu- lar, knowing the average number of facts per entity is quite informative. We shall refer to this as the knowledge density of the KG. Low knowledge density (or high sparsity) in automatically constructed KGs has been rec- ognized in recent research (West et al., 2014). For example, NELL KG has a knowledge den- sity of 1.34. Such low knowledge density puts significant limitations on the utility of these KGs. Construction of such KGs tend to follow a batch paradigm: the knowledge extraction system makes a …

Content truncated to 500 words