Improving Learning and Inference in a Large Knowledge-Base using Latent Syntactic Cues

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 833–838, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Improving Learning and Inference in a Large Knowledge-base using Latent Syntactic Cues Matt Gardner, Partha Pratim Talukdar, Bryan Kisiel, and Tom Mitchell Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213, USA {mg1,ppt,bkisiel,tom.mitchell}@cs.cmu.edu Abstract Automatically constructed Knowledge Bases (KBs) are often incomplete and there is a gen- uine need to improve their coverage. Path Ranking Algorithm (PRA) is a recently pro- posed method which aims to improve KB cov- erage by performing inference directly over the KB graph. For the first time, we demon- strate that addition of edges labeled with la- tent features mined from a large dependency parsed corpus of 500 million Web documents can significantly outperform previous PRA- based approaches on the KB inference task. We present extensive experimental results val- idating this finding. The resources presented in this paper are publicly available. 1 Introduction Over the last few years, several large scale Knowl- edge Bases (KBs) such as Freebase (Bollacker et al., 2008), NELL (Carlson et al., 2010), and YAGO (Suchanek et al., 2007) have been developed. Each such KB consists of millions of facts (e.g., (Tiger Woods, playsSport, Golf)) spanning over multiple relations. Unfortunately, these KBs are often incom- plete and there is a need to increase their coverage of facts to make them useful in practical applications. A strategy to increase coverage might be to per- form inference directly over the KB represented as a graph. For example, if the KB contained the follow- ing facts, (Tiger Woods, participatesIn, PGA Tour)) and (Golf, sportOfTournament, PGA Tour), then by putting these two facts together, we could potentially infer that (Tiger Woods, playsSport, Golf). The Figure 1: Example demonstrating how lexicalized syn- tactic edges can improve connectivity in the KB enabling PRA (Lao and Cohen, 2010) to discover relationships be- tween Alex Rodriguez and World Series. Edges with la- tent labels can improve inference performance by reduc- ing data sparsity. See Section 1.1 for details. recently proposed Path Ranking Algorithm (PRA) (Lao and Cohen, 2010) performs such inference by automatically learning semantic inference rules over the KB (Lao et al., 2011). PRA uses features based off of sequences of edge types, e.g., ⟨playsSport, sportOfTournament⟩, to predict missing facts in the KB. PRA was extended by (Lao et al., 2012) to per- form inference over a KB augmented with depen- dency parsed sentences. While this opens up the possibility of learning syntactic-semantic inference rules, the set of syntactic edge labels used are just the unlexicalized dependency role labels (e.g., nobj, dobj, etc., without the corresponding words), thereby limiting overall expressitivity of the learned inference rules. To overcome this limitation, in this paper we augment the KB graph by adding edges with more expressive lexicalized syntactic labels (where the labels are words instead of dependen- 833 cies). These additional edges, e.g., (Alex Rodriguez, “plays for”, NY Yankees), are mined by extracting 600 million …

Content truncated to 500 words