Unsupervised Learning of Prototypical Fillers for Implicit Semantic Role Labeling

Proceedings of NAACL-HLT 2016, pages 1473–1479, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Unsupervised Learning of Prototypical Fillers for Implicit Semantic Role Labeling Niko Schenk and Christian Chiarcos Applied Computational Linguistics Lab Goethe University Frankfurt am Main, Germany {n.schenk,chiarcos}@em.uni-frankfurt.de Abstract Gold annotations for supervised implicit se- mantic role labeling are extremely sparse and costly. As a lightweight alternative, this paper describes an approach based on unsupervised parsing which can do without iSRL-specific training data: We induce prototypical roles from large amounts of explicit SRL annota- tions paired with their distributed word repre- sentations. An evaluation shows competitive performance with supervised methods on the SemEval 2010 data, and our method can eas- ily be applied to predicates (or languages) for which no training annotations are available. 1 Introduction Semantic role labeling (SRL) (Gildea and Jurafsky, 2002) has become a well-established and highly im- portant NLP component which directly benefits var- ious downstream applications, such as text summa- rization (Trandab˘at¸, 2011), recognizing textual en- tailment (Sammons et al., 2012) or QA systems (Shen and Lapata, 2007; Moreda et al., 2011). Its goal is to detect verbal or nominal predicates, to- gether with their associated arguments and semantic roles, either by PropBank/Nombank (Palmer et al., 2005; Meyers et al., 2004) or FrameNet (Baker et al., 1998) analysis. In its traditional form, however, SRL is restricted to the local syntactic context of the predicate as in the following example from Ruppen- hofer et al. (2010): [GOAL/NIIn the centre of this room] there was an upright beam, [THEMEwhich] had been placed [TIMEat some period] as a support for the old worm-eaten baulk of timber which spanned the roof. In a FrameNet-style analysis of the sentence, the predicate place evokes the PLACING frame, with two frame elements (roles) overtly expressed (THEME and TIME) but with one role – GOAL – be- yond the embedded relative clause and thus beyond the scope of the SRL parser. Such implicit roles, or null instantiations (NIs) (Fillmore, 1986; Ruppen- hofer, 2005) are much harder to detect automatically, as they require to broaden the analysis to the sur- rounding discourse, commonly also to preceding (or following) sentences. State-of-the-art approaches to implicit SRL (iSRL) are supervised and need a groundwork of hand-annotated training data – which is costly, ex- tremely sparse, limited to only a handful of predi- cates, and requires careful feature engineering (Ger- ber and Chai, 2012; Silberer and Frank, 2012; Li et al., 2015). A first attempt has been made to combine the scarce resources available (Feizabadi and Pad´o, 2015), but given the great diversity of predicate- specific roles and enormous complexity of the task, the main issues remain (Chen et al., 2010). A promising exploratory effort recently made by Gorinski et al. (2013) aims to overcome the anno- tation bottleneck by using distributional methods to infer evidence for elements filling null instantiated roles. The authors do not rely on gold annotations but instead learn distributional properties of fillers induced from a large corpus. Our …

Content truncated to 500 words