CaRe: Open Knowledge Graph Embeddings

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 378–388, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 378 CaRe: Open Knowledge Graph Embeddings Swapnil Gupta Indian Institute of Science Bangalore, India swapnilgupta.229@gmail.com Sreyash Kenkre Walmart Labs Bangalore, India sreyash@gmail.com Partha Talukdar Indian Institute of Science Bangalore, India ppt@iisc.ac.in Abstract Open Information Extraction (OpenIE) meth- ods are effective at extracting (noun phrase, re- lation phrase, noun phrase) triples from text, e.g., (Barack Obama, took birth in, Honolulu). Organization of such triples in the form of a graph with noun phrases (NPs) as nodes and relation phrases (RPs) as edges results in the construction of Open Knowledge Graphs (OpenKGs). In order to use such OpenKGs in downstream tasks, it is often desirable to learn embeddings of the NPs and RPs present in the graph. Even though several Knowl- edge Graph (KG) embedding methods have been recently proposed, all of those methods have targeted Ontological KGs, as opposed to OpenKGs. Straightforward application of existing Ontological KG embedding methods to OpenKGs is challenging, as unlike Onto- logical KGs, OpenKGs are not canonicalized, i.e., a real-world entity may be represented us- ing multiple nodes in the OpenKG, with each node corresponding to a different NP refer- ring to the entity. For example, nodes with labels Barack Obama, Obama, and President Obama may refer to the same real-world en- tity Barack Obama. Even though canonical- ization of OpenKGs has received some at- tention lately, output of such methods has not been used to improve OpenKG embed- dings. We fill this gap in the paper and propose Canonicalization-infused Representa- tions (CaRe) for OpenKGs. Through exten- sive experiments, we observe that CaRe en- ables existing models to adapt to the chal- lenges in OpenKGs and achieve substantial improvements for the link prediction task. 1 Introduction Open Information Extraction (OpenIE) methods such as ReVerb (Fader et al., 2011), OLLIE (Mausam et al., 2012), BONIE (Saha et al., 2017) and CALMIE (Saha and Mausam, 2018) can au- tomatically extract (noun phrase, relation phrase, Sentences in Text Corpus: • Barack Obama, 44th president of USA, took birth in Honolulu. • Michelle Obama wife of Barack was born in Chicago. OpenIE OpenKG Barack Obama USA Michelle Obama Chicago 44th president of Honolulu Took birth in Barack Wife of Was born in First lady of ? Figure 1: Challenges in OpenKG. KG embedding methods are effective in downstream tasks such as Link Prediction. Existing KG embedding methods are in- effective on OpenKGs as they are not canonicalized. CaRe can effectively utilize canonicalization informa- tion to learn better embeddings in OpenKGs. Green dotted line represents the missing NP canonicalization information. Please see Section 1 for more details. noun phrase) triples from text, e.g., (Barack Obama, took birth in, Honolulu), (Michelle Obama, wife of, Barack), etc. An Open Knowl- edge Graph (OpenKG) can be constructed out of such triples by representing noun phrases (NPs) as nodes and …

Content truncated to 500 words