Old is Gold: Linguistic Driven Approach for Entity and Relation Linking of Short Text

Proceedings of NAACL-HLT 2019, pages 2336–2346 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2336 Old is Gold: Linguistic Driven Approach for Entity and Relation Linking of Short Text Ahmad Sakor∗1, Isaiah Onando Mulang’2, Kuldeep Singh2, Saeedeh Shekarpour3, Maria-Esther Vidal4, Jens Lehmann2, and Sören Auer4 1L3S Research Center, Hannover, Germany 2Fraunhofer IAIS, Sankt Augustin, Germany 3University of Dayton, Dayton, USA 4TIB, Hannover, Germany {sakor}@l3s.de,{maria.vidal,auer}@tib.eu {isaiah.mulang.onando,kuldeep.singh}@iais.fraunhofer.de {sshekarpour1}@udayton.edu {jens.lehmann}@iais.fraunhofer.de Abstract Short texts challenge NLP tasks such as named entity recognition, disambiguation, linking and relation inference because they do not pro- vide sufficient context or are partially mal- formed (e.g. wrt. capitalization, long tail en- tities, implicit relations). In this work, we present the Falcon approach which effectively maps entities and relations within a short text to its mentions of a background knowledge graph. Falcon overcomes the challenges of short text using a light-weight linguistic ap- proach relying on a background knowledge graph. Falcon performs joint entity and rela- tion linking of a short text by leveraging sev- eral fundamental principles of English mor- phology (e.g. compounding, headword iden- tification) and utilizes an extended knowledge graph created by merging entities and rela- tions from various knowledge sources. It uses the context of entities for finding relations and does not require training data. Our empirical study using several standard benchmarks and datasets show that Falcon significantly outper- forms state-of-the-art entity and relation link- ing for short text query inventories. 1 Introduction Entity Linking (EL) task annotates surface forms in the text with the corresponding reference men- tions in knowledge bases such as Wikipedia. It involves the two sub-tasks, i.e. Named Entity Recognition and Disambiguation (NER and NED) tasks. The state of the art contains considerable research body for EL from text to its Wikipedia mention (Cucerzan, 2007; Ferragina and Scaiella, 2010; Hoffart et al., 2011; Balog, 2018; Shen ∗First three authors have equal contribution. et al., 2015; Ferragina and Scaiella, 2010; Hof- fart et al., 2014). With the emergence of Knowl- edge Graphs (KGs) which represent data in a higher structured and semantic format such as DBpedia (Auer et al., 2007), Freebase (Bollacker et al., 2008) and Wikidata (Vrandecic, 2012) that utilize Wikipedia as familiar knowledge source, retrieval-based applications such as question an- swering (QA) systems or keyword-based seman- tic search systems are empowered to provide more cognitive capabilities. Entity linking is a crucial component for a variety of applications built on knowledge graphs. For instance, an ideal NED tool on DBpedia recognizes the entities embedded in the question ‘Who wrote the book The Pillars of The Earth?’ and links them to the correspond- ing DBpedia entity (e.g. ‘Pillars of The Earth’ to dbr:The_Pillars_of_the_Earth)1. An- other important NLP task is relation linking; it is about linking surface forms in text repre- senting a relation to equivalent relations (pred- icates) of a KG. In our example question, an ideal relation linking (RL) tool links ‘wrote’ to dbo:author2. There are existing approaches which address EL and RL tasks either jointly or …

Content truncated to 500 words