Learning to Explain Entity Relationships in Knowledge Graphs

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 564–574, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Learning to Explain Entity Relationships in Knowledge Graphs Nikos Voskarides∗ University of Amsterdam n.voskarides@uva.nl Edgar Meij Yahoo Labs, London emeij@yahoo-inc.com Manos Tsagkias 904Labs, Amsterdam manos@904labs.com Maarten de Rijke University of Amsterdam derijke@uva.nl Wouter Weerkamp 904Labs, Amsterdam wouter@904labs.com Abstract We study the problem of explaining re- lationships between pairs of knowledge graph entities with human-readable de- scriptions. Our method extracts and en- riches sentences that refer to an entity pair from a corpus and ranks the sentences ac- cording to how well they describe the re- lationship between the entities. We model this task as a learning to rank problem for sentences and employ a rich set of fea- tures. When evaluated on a large set of manually annotated sentences, we find that our method significantly improves over state-of-the-art baseline models. 1 Introduction Knowledge graphs are a powerful tool for support- ing a large spectrum of search applications includ- ing ranking, recommendation, exploratory search, and web search (Dong et al., 2014). A knowl- edge graph aggregates information around enti- ties across multiple content sources and links these entities together, while at the same time provid- ing entity-specific properties (such as age or em- ployer) and types (such as actor or movie). Although there is a growing interest in au- tomatically constructing knowledge graphs, e.g., from unstructured web data (Weston et al., 2013; Craven et al., 2000; Fan et al., 2012), the prob- lem of providing evidence on why two entities are related in a knowledge graph remains largely unaddressed. Extracting and presenting evidence for linking two entities, however, is an impor- tant aspect of knowledge graphs, as it can enforce trust between the user and a search engine, which in turn can improve long-term user engagement, e.g., in the context of related entity recommenda- tion (Blanco et al., 2013). Although knowledge ∗This work was carried out while this author was visiting Yahoo Labs. graphs exist that provide this functionality to a certain degree (e.g., when hovering over Google’s suggested entities, see Figure 1), to the best of our knowledge there is no previously published re- search on methods for entity relationship explana- tion. Figure 1: Part of Google’s search result page for the query “barack obama”. When hovering over the related entity “Michelle Obama”, an explana- tion of the relationship between her and “Barack Obama” is shown. In this paper we propose a method for explain- ing the relationship between two entities, which we evaluate on a newly constructed annotated dataset that we make publicly available. In par- ticular, we consider the task of explaining rela- tionships between pairs of Wikipedia entities. We aim to infer a human-readable description for an entity pair given a relationship between the two entities. Since Wikipedia does not explicitly de- fine relationships between entities we use a knowl- edge graph to …

Content truncated to 500 words