Commonsense Knowledge Base Completion

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1445–1455, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Commonsense Knowledge Base Completion Xiang Li∗‡ Aynaz Taheri† Lifu Tu‡ Kevin Gimpel‡ ∗University of Chicago, Chicago, IL, 60637, USA †University of Illinois at Chicago, Chicago, IL, 60607, USA ‡Toyota Technological Institute at Chicago, Chicago, IL, 60637, USA lix1@uchicago.edu, ataher2@uic.edu, {lifu,kgimpel}@ttic.edu Abstract We enrich a curated resource of common- sense knowledge by formulating the prob- lem as one of knowledge base comple- tion (KBC). Most work in KBC focuses on knowledge bases like Freebase that re- late entities drawn from a fixed set. How- ever, the tuples in ConceptNet (Speer and Havasi, 2012) define relations between an unbounded set of phrases. We develop neural network models for scoring tuples on arbitrary phrases and evaluate them by their ability to distinguish true held-out tuples from false ones. We find strong performance from a bilinear model using a simple additive architecture to model phrases. We manually evaluate our trained model’s ability to assign quality scores to novel tuples, finding that it can propose tu- ples at the same quality level as medium- confidence tuples from ConceptNet. 1 Introduction Many ambiguities in natural language process- ing (NLP) can be resolved by using knowledge of various forms. Our focus is on the type of knowledge that is often referred to as “common- sense” or “background” knowledge. This knowl- edge is rarely expressed explicitly in textual cor- pora (Gordon and Van Durme, 2013). Some re- searchers have developed techniques for inferring this knowledge from patterns in raw text (Gor- don, 2014; Angeli and Manning, 2014), while oth- ers have developed curated resources of common- sense knowledge via manual annotation (Lenat and Guha, 1989; Speer and Havasi, 2012) or games with a purpose (von Ahn et al., 2006). Curated resources typically have high preci- sion but suffer from a lack of coverage. For cer- relation right term conf. MOTIVATEDBYGOAL relax 3.3 USEDFOR relaxation 2.6 MOTIVATEDBYGOAL your muscle be sore 2.3 HASPREREQUISITE go to spa 2.0 CAUSES get pruny skin 1.6 HASPREREQUISITE change into swim suit 1.6 Table 1: ConceptNet tuples with left term “soak in hotspring”; final column is confidence score. tain resources, researchers have developed meth- ods to automatically increase coverage by infer- ring missing entries. These methods are com- monly categorized under the heading of knowl- edge base completion (KBC). KBC is widely- studied for knowledge bases like Freebase (Bol- lacker et al., 2008) which contain large sets of enti- ties and relations among them (Mintz et al., 2009; Nickel et al., 2011; Riedel et al., 2013; West et al., 2014), including recent work using neural net- works (Socher et al., 2013; Yang et al., 2014). We improve the coverage of commonsense re- sources by formulating the problem as one of knowledge base completion. We focus on a par- ticular curated commonsense resource called Con- ceptNet (Speer and Havasi, 2012). ConceptNet contains tuples consisting of a left term, a rela- tion, …

Content truncated to 500 words