Zero-Shot Relation Extraction via Reading Comprehension

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 333–342, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Zero-Shot Relation Extraction via Reading Comprehension Omer Levy1 Minjoon Seo1 Eunsol Choi1 Luke Zettlemoyer1,2 1Paul G. Allen School of Computer Science and Engineering University of Washington, Seattle WA {omerlevy,minjoon,eunsol,lsz}@cs.washington.edu 2Allen Institute for Artificial Intelligence, Seattle, WA Abstract We show that relation extraction can be re- duced to answering simple reading com- prehension questions, by associating one or more natural-language questions with each relation slot. This reduction has sev- eral advantages: we can (1) learn relation- extraction models by extending recent neural reading-comprehension techniques, (2) build very large training sets for those models by combining relation-specific crowd-sourced questions with distant su- pervision, and even (3) do zero-shot learn- ing by extracting new relation types that are only specified at test-time, for which we have no labeled training examples. Ex- periments on a Wikipedia slot-filling task demonstrate that the approach can gen- eralize to new questions for known rela- tion types with high accuracy, and that zero-shot generalization to unseen relation types is possible, at lower accuracy levels, setting the bar for future work on this task. 1 Introduction Relation extraction systems populate knowledge bases with facts from an unstructured text corpus. When the type of facts (relations) are predefined, one can use crowdsourcing (Liu et al., 2016) or distant supervision (Hoffmann et al., 2011) to col- lect examples and train an extraction model for each relation type. However, these approaches are incapable of extracting relations that were not specified in advance and observed during training. In this paper, we propose an alternative approach for relation extraction, which can potentially ex- tract facts of new types that were neither specified nor observed a priori. Relation Question Template educated at(x, y) Where did x graduate from? In which university did x study? What is x’s alma mater? occupation(x, y) What did x do for a living? What is x’s job? What is the profession of x? spouse(x, y) Who is x’s spouse? Who did x marry? Who is x married to? Figure 1: Common knowledge-base relations de- fined by natural-language question templates. We show that it is possible to reduce relation ex- traction to the problem of answering simple read- ing comprehension questions. We map each re- lation type R(x, y) to at least one parametrized natural-language question qx whose answer is y. For example, the relation educated at(x, y) can be mapped to “Where did x study?” and “Which university did x graduate from?”. Given a par- ticular entity x (“Turing”) and a text that men- tions x (“Turing obtained his PhD from Prince- ton”), a non-null answer to any of these questions (“Princeton”) asserts the fact and also fills the slot y. Figure 1 illustrates a few more examples. This reduction enables new ways of framing the learning problem. In particular, it allows us to perform zero-shot learning: define new rela- tions “on the …

Content truncated to 500 words