Learning-based Multi-Sieve Co-reference Resolution with Knowledge

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1234–1244, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Learning-based Multi-Sieve Co-reference Resolution with Knowledge∗ Lev Ratinov Google Inc.† ratinov@google.com Dan Roth University of Illinois at Urbana-Champaign danr@illinois.edu Abstract We explore the interplay of knowledge and structure in co-reference resolution. To inject knowledge, we use a state-of-the-art system which cross-links (or “grounds”) expressions in free text to Wikipedia. We explore ways of using the resulting grounding to boost the performance of a state-of-the-art co-reference resolution system. To maximize the utility of the injected knowledge, we deploy a learning- based multi-sieve approach and develop novel entity-based features. Our end system outper- forms the state-of-the-art baseline by 2 B3 F1 points on non-transcript portion of the ACE 2004 dataset. 1 Introduction Co-reference resolution is the task of grouping men- tions to entities. For example, consider the text snip- pet in Fig. 11. The correct output groups the men- tions {m1, m2, m5} to one entity while leaving m3 ∗ We thank Nicholas Rizzolo and Kai Wei Chang for their invaluable help with modifying the baseline co-reference sys- tem. We thank the anonymous EMNLP reviewers for con- structive comments. This research was supported by the Army Research Laboratory (ARL) under agreement W911NF-09-2- 0053 and by the Defense Advanced Research Projects Agency (DARPA) Machine Reading Program under Air Force Research Laboratory (AFRL) prime contract no. FA8750-09-C-0181. Any opinions, findings, conclusions or recommendations are those of the authors and do not necessarily reflect the view of the ARL, DARPA, AFRL, or the US government. † The majority of this work was done while the first author was at the University of Illinois. 1Throughout this paper, curly brackets {} denote the extent and square brackets [] denote the head. “After the {[vessel]}m1 suffered a catastrophic torpedo detonation, {[Kursk]}m2 sank in the waters of {[Barents Sea]}m3 with all hands lost. Though rescue attempts were offered by a nearby {Norwegian [ship]}m4, Russia declined initial rescue offers, and all 118 sailors and officers aboard {[Kursk]}m5 perished.” Figure 1: Example illustrating the challenges in co-reference resolution. and m4 as singletons. Resolving co-reference is fun- damental for understanding natural language. For example in Fig. 1, to infer that Kusrk has suffered a torpedo detonation, we have to understand that {[vessel]}m1 refers to {[Kursk]}m2. This inference is typically trivial for humans, but proves extremely challenging for state-of-the-art co- reference resolution systems. We believe that it is world knowledge that gives people the ability to un- derstand text with such ease. A human reader can in- fer that since Kursk sank, it must be a vessel and ves- sels which suffer catastrophic torpedo detonations can sink. Moreover, some readers might just know that Kursk is a Russian submarine named after the city of Kursk, where the largest tank battle in his- tory took place in 1943. In this work we are using Wikipedia as a source of encyclopedic knowledge. The key …

Content truncated to 500 words