Revisiting Selectional Preferences for Coreference Resolution

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1332–1339 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Revisiting Selectional Preferences for Coreference Resolution Benjamin Heinzerling∗ AIPHES Heidelberg Institute for Theoretical Studies benjamin.heinzerling@h-its.org Nafise Sadat Moosavi∗ Heidelberg Institute for Theoretical Studies nafise.moosavi@h-its.org Michael Strube Heidelberg Institute for Theoretical Studies michael.strube@h-its.org Abstract Selectional preferences have long been claimed to be essential for coreference res- olution. However, they are mainly mod- eled only implicitly by current corefer- ence resolvers. We propose a dependency- based embedding model of selectional preferences which allows fine-grained compatibility judgments with high cover- age. We show that the incorporation of our model improves coreference resolu- tion performance on the CoNLL dataset, matching the state-of-the-art results of a more complex system. However, it comes with a cost that makes it debatable how worthwhile such improvements are. 1 Introduction Selectional preferences have long been claimed to be useful for coreference resolution. In his sem- inal work on “Resolving Pronominal References” Hobbs (1978) proposed a semantic approach that requires reasoning about the “demands the pred- icate makes on its arguments.” For example, se- lectional preferences allow resolving the pronoun it in the text “The Titanic hit an iceberg. It sank quickly.” Here, the predicate sink ‘prefers’ certain subject arguments over others: It is plausible that a ship sinks, but implausible that an iceberg does. Work on the automatic acquisition of selectional preferences has shown considerable progress (Da- gan and Itai, 1990; Resnik, 1993; Agirre and Mar- tinez, 2001; Pantel et al., 2007; Erk, 2007; Ritter et al., 2010; Van de Cruys, 2014). However, to- day’s coreference resolvers (Martschat and Strube, 2015; Wiseman et al., 2016; Clark and Manning, 2016a, i.a.) capture selectional preferences only ∗ These authors contributed equally to this work. implicitly at best, e.g., via a given mention’s de- pendency governor and other contextual features. Since negative results do not often get reported, there is no clear evidence in the literature re- garding the non-utility of particular knowledge sources. Consequently, an absence of the explicit modeling of selectional preferences in the recent literature is an indicator that incorporating this knowledge source has not been very successful for coreference resolution. More than ten years ago, Kehler et al. (2004) declared the “non-utility of predicate-argument structures for pronoun resolution” and observed that minor improvements on a small dataset were due to fortuity rather than selectional preferences having captured meaningful world knowledge re- lations. The claim by Kehler et al. (2004) is based on selectional preferences extracted from a, by cur- rent standards, small number of 2.8m predicate- argument pairs. Furthermore, they employ a sim- ple (linear) maximum entropy classifier, which re- quires manual definition of feature combinations and is unlikely to fully capture the complex inter- action between selectional preferences and other coreference features. Therefore, it is worth revis- iting how a better selectional preference model af- fects the performance of a more complex corefer- ence resolver. In this work, we propose a fine-grained, high- …

Content truncated to 500 words