An Empirical Study on Compositionality in Compound Nouns

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 210–218, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP An Empirical Study on Compositionality in Compound Nouns Siva Reddy University of York, UK siva@cs.york.ac.uk Diana McCarthy Lexical Computing Ltd, UK diana@dianamccarthy.co.uk Suresh Manandhar University of York, UK suresh@cs.york.ac.uk Abstract A multiword is compositional if its mean- ing can be expressed in terms of the mean- ing of its constituents. In this paper, we collect and analyse the compositionality judgments for a range of compound nouns using Mechanical Turk. Unlike exist- ing compositionality datasets, our dataset has judgments on the contribution of con- stituent words as well as judgments for the phrase as a whole. We use this dataset to study the relation between the judg- ments at constituent level to that for the whole phrase. We then evaluate two differ- ent types of distributional models for com- positionality detection – constituent based models and composition function based models. Both the models show competi- tive performance though the composition function based models perform slightly better. In both types, additive models per- form better than their multiplicative coun- terparts. 1 Introduction Compositionality is a language phenomenon where the meaning of an expression can be ex- pressed in terms of the meaning of its constituents. Multiword expressions (Sag et al., 2002, MWEs) are known to display a continuum of composi- tionality (McCarthy et al., 2003) where some of them are compositional e.g. “swimming pool”, some are non-compositional e.g. “cloud nine”, and some in between e.g. “zebra crossing”. The past decade has seen interest in develop- ing computational methods for compositionality in MWEs (Lin, 1999; Schone and Jurafsky, 2001; Baldwin et al., 2003; Bannard et al., 2003; Mc- Carthy et al., 2003; Venkatapathy and Joshi, 2005; Katz and Giesbrecht, 2006; Sporleder and Li, 2009). Recent developments in vector-based se- mantic composition functions (Mitchell and Lap- ata, 2008; Widdows, 2008) have also been applied to compositionality detection (Giesbrecht, 2009). While the existing methods of compositional- ity detection use constituent word level seman- tics to compose the semantics of the phrase, the evaluation datasets are not particularly suitable to study the contribution of each constituent word to the semantics of the phrase. Existing datasets (McCarthy et al., 2003; Venkatapathy and Joshi, 2005; Katz and Giesbrecht, 2006; Biemann and Giesbrecht, 2011) only have the compositionality judgment of the whole expression without con- stituent word level judgment, or they have judg- ments on the constituents without judgments on the whole (Bannard et al., 2003). Our dataset al- lows us to examine the relationship between the two rather than assume the nature of it. In this paper we collect judgments of the con- tribution of constituent nouns within noun-noun compounds (section 2) alongside judgments of compositionality of the compound. We study the relation between the contribution of the parts with the compositionality of the whole (section 3). We propose various constituent based models (section 4.3) which are intuitive and related to existing models of compositionality …

Content truncated to 500 words