Assassinating Lady Gaga with a banana?   A practical model of fine-grained selectional preferences and its application to coreference resolution


Selectional preferences -- and semantics in general -- have long been theorized to be essential for coreference resolution. In his seminal work on "Resolving Pronoun References" (1978), Hobbs proposed a semantic approach that, among other things, requires reasoning about the "demands the predicate makes on its arguments".

Despite advances in modeling and automatic acquisition of selectional preferences (Dagan & Atai, 1990; Resnik, 1992, Agirre & Martinez, 2001; Pantel et al., 2007; Erk, 2007; Ritter et al., 2010), they are conspicuously absent in today's coreference resolvers (Martschat & Strube, 2015; Wiseman et al., 2016; Clark & Manning, 2016b). In fact, use of semantics in automatic coreference resolution is mostly limited to shallow features capturing number, gender, animacy, and entity type agreement of potential coreferents (Manning et al., 2014). The apparent non-utility of semantics, and selectional preferences in particular, has been lamented repeatedly (Kehler et al., 2004; Durrett & Klein, 2013; Strube, 2015).

We present a new attempt at integrating selectional preferences into coreference resolution. Mainly driven by practical considerations, we model predicate arguments using distributed representations derived from fine-grained entity types (Ling & Weld, 2012), named entities, and common noun phrases. Predicate slots are defined via universal dependencies (De Marneffe et al., 2014; Schuster & Manning, 2016).

We report on ongoing experiments which show that our selectional preference model, trained on large automatically annotated corpora and integrated into a state-of-the-art system, yields absolute improvements on English CoNLL'12 of up to 1.0 F1 with gold mentions and 0.8 F1 with system mentions.