ReNoun: Fact Extraction for Nominal Attributes

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 325–335, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics ReNoun: Fact Extraction for Nominal Attributes Mohamed Yahya∗ Max Planck Institute for Informatics myahya@mpi-inf.mpg.de Steven Euijong Whang, Rahul Gupta, Alon Halevy Google Research {swhang,grahul,halevy}@google.com Abstract Search engines are increasingly relying on large knowledge bases of facts to provide direct answers to users’ queries. How- ever, the construction of these knowledge bases is largely manual and does not scale to the long and heavy tail of facts. Open information extraction tries to address this challenge, but typically assumes that facts are expressed with verb phrases, and there- fore has had difficulty extracting facts for noun-based relations. We describe ReNoun, an open information extraction system that complements pre- vious efforts by focusing on nominal at- tributes and on the long tail. ReNoun’s ap- proach is based on leveraging a large on- tology of noun attributes mined from a text corpus and from user queries. ReNoun creates a seed set of training data by us- ing specialized patterns and requiring that the facts mention an attribute in the ontol- ogy. ReNoun then generalizes from this seed set to produce a much larger set of ex- tractions that are then scored. We describe experiments that show that we extract facts with high precision and for attributes that cannot be extracted with verb-based tech- niques. 1 Introduction One of the major themes driving the current evo- lution of search engines is to make the search experience more efficient and mobile friendly for users by providing them concrete answers to queries. These answers, that apply to queries about entities that the search engine knows about (e.g., famous individuals, organizations or loca- tions) complement the links that the search en- ∗Work done during an internship at Google Research. gine typically returns (Sawant and Chakrabati, 2013; Singhal, 2012; Yahya et al., 2012). To support such answers, the search engine main- tains a knowledge base that describes various at- tributes of an entity (e.g., (Nicolas Sarkozy, wife, Carla Bruni)). Upon receiving a query, the search engine tries to recognize whether the answer is in its knowledge base. For the most part, the aforementioned knowl- edge bases are constructed using manual tech- niques and carefully supervised information ex- traction algorithms. As a result, they obtain high coverage on head attributes, but low coverage on tail ones, such as those shown in Table 1. For ex- ample, they may have the answer for the query “Sarkozy’s wife”, but not for “Hollande’s ex- girlfriend” or “Google’s philanthropic arm”. In addition to broadening the scope of query answer- ing, extending the coverage of the knowledge base to long tail attributes can also facilitate providing Web answers to the user. Specifically, the search engine can use lower-confidence facts to corrob- orate an answer that appears in text in one of the top Web results and highlight them to the user. This paper describes ReNoun, an open- information extraction system …

Content truncated to 500 words