The Life and Death of Discourse Entities: Identifying Singleton Mentions

Proceedings of NAACL-HLT 2013, pages 627–633, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics The Life and Death of Discourse Entities: Identifying Singleton Mentions Marta Recasens Linguistics Department Stanford University Stanford, CA 94305 recasens@google.com Marie-Catherine de Marneffe Linguistics Department The Ohio State University Columbus, OH 43210 mcdm@ling.osu.edu Christopher Potts Linguistics Department Stanford University Stanford, CA 94305 cgpotts@stanford.edu Abstract A discourse typically involves numerous en- tities, but few are mentioned more than once. Distinguishing discourse entities that die out after just one mention (singletons) from those that lead longer lives (coreferent) would ben- efit NLP applications such as coreference res- olution, protagonist identification, topic mod- eling, and discourse coherence. We build a lo- gistic regression model for predicting the sin- gleton/coreferent distinction, drawing on lin- guistic insights about how discourse entity lifespans are affected by syntactic and seman- tic features. The model is effective in its own right (78% accuracy), and incorporating it into a state-of-the-art coreference resolution sys- tem yields a significant improvement. 1 Introduction Not all discourse entities are created equal. Some lead long lives and appear in a variety of discourse contexts (coreferent), whereas others never escape their birthplaces, dying out after just one mention (singletons). The ability to make this distinction based on properties of the NPs used to identify these referents (mentions) would benefit not only corefer- ence resolution, but also topic analysis, textual en- tailment, and discourse coherence. The existing literature provides numerous gen- eralizations relevant to answering the question of whether a given discourse entity will be singleton or coreferent. These involve the internal syntax and morphology of the target NP (Prince, 1981a; Prince, 1981b; Wang et al., 2006), the grammatical function and discourse role of that NP (Chafe, 1976; Hobbs, 1979; Walker et al., 1997; Beaver, 2004), and the in- teraction of all of those features with semantic oper- ators like negation, modals, and attitude predicates (Karttunen, 1973; Karttunen, 1976; Kamp, 1981; Heim, 1982; Heim, 1992; Roberts, 1990; Groe- nendijk and Stokhof, 1991; Bittner, 2001). The first step in our analysis is to bring these insights together into a single logistic regression model — the lifespan model — and assess their predictive power on real data. We show that the features generally behave as the existing literature leads us to expect, and that the model itself is highly effective at predicting whether a given mention is singleton or coreferent. We then provide an initial assessment of the engineering value of making the singleton/coreferent distinction by incorporating our lifespan model into the Stanford coreference resolu- tion system (Lee et al., 2011). This addition results in a significant improvement on the CoNLL-2012 Shared Task data, across the MUC, B3, CEAF, and CoNLL scoring algorithms. 2 Data All the data used throughout the paper come from the CoNLL-2012 Shared Task (Pradhan et al., 2012), which included the 1.6M English words from OntoNotes v5.0 (Hovy et al., 2006) that have been annotated with different layers of annotation (coref- erence, parse trees, etc.). We used the …

Content truncated to 500 words