Inclusive yet Selective: Supervised Distributional Hypernymy Detection

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1025–1036, Dublin, Ireland, August 23-29 2014. Inclusive yet Selective: Supervised Distributional Hypernymy Detection Stephen Roller∗, Katrin Erk†, Gemma Boleda† ∗ Department of Computer Science † Department of Linguistics The University of Texas at Austin roller@cs.utexas.edu, katrin.erk@mail.utexas.edu, gemma.boleda@upf.edu Abstract We test the Distributional Inclusion Hypothesis, which states that hypernyms tend to occur in a superset of contexts in which their hyponyms are found. We find that this hypothesis only holds when it is applied to relevant dimensions. We propose a robust supervised approach that achieves accuracies of .84 and .85 on two existing datasets and that can be interpreted as selecting the dimensions that are relevant for distributional inclusion. 1 Introduction One of the main criticisms of distributional models has been that they fail to distinguish between semantic relations: Typical nearest neighbors of dog are words like cat, animal, puppy, tail, or owner, all obviously related to dog, but through very different types of semantic relations. On these grounds, Murphy (2002) argues that distributional models cannot be a valid model of conceptual representation. Distinguishing semantic relations are also crucial for drawing inferences from distributional data, as different semantic relations lead to different inference rules (Lenci, 2008). This is of practical import for tasks such as Recognizing Textual Entailment or RTE (Geffet and Dagan, 2004). For these reasons, research has in recent years started to attempt the detection of specific semantic relationships, and current results suggest that distributional models can, in fact, distinguish between semantic relations, given the right similarity measures (Weeds et al., 2004; Kotlerman et al., 2010; Lenci and Benotto, 2012; Herbelot and Ganesalingam, 2013; Santus, 2013). Because of its relevance for RTE and other tasks, much of this work has focused on hypernymy. Hypernymy is the semantic relation between a superordinate term in a taxonomy (e.g. animal) and a subordinate term (e.g. dog). Distributional approaches to date for detecting hypernymy, and the related but broader relation of lexical entailment, have been unsupervised (except for Baroni et al. (2012)) and have mostly been based on the Distributional Inclusion Hypothesis (Zhitomirsky-Geffet and Dagan, 2005; Zhitomirsky-Geffet and Dagan, 2009), which states that more specific terms appear in a subset of the distributional contexts in which more general terms appear. So, animal can occur in all the contexts in which dog can occur, plus some contexts in which dog cannot – for instance, rights can be a typical cooccurrence for animal (e.g. “animal rights”), but not so much for dog (e.g. #“dog rights”). This paper takes a closer look at the Distributional Inclusion Hypothesis for hypernymy detection. We show that the current best unsupervised approach is brittle in that their performance depends on the space they are applied to. This raises the question of whether the Distributional Inclusion Hypothesis is correct, and if so, under what circumstances it holds. We use a simple supervised approach to relation detection that has good performance (accuracy .84 on BLESS, .85 on the …

Content truncated to 500 words