An Unsupervised Ranking Model for Noun-Noun Compositionality

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 132–141, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics An Unsupervised Ranking Model for Noun-Noun Compositionality Karl Moritz Hermann, Phil Blunsom, and Stephen Pulman Department of Computer Science University of Oxford Wolfson Building, Parks Road Oxford OX1 3QD, UK {karl.moritz.hermann,phil.blunsom,stephen.pulman}@cs.ox.ac.uk Abstract We propose an unsupervised system that learns continuous degrees of lexicality for noun-noun compounds, beating a strong base- line on several tasks. We demonstrate that the distributional representations of compounds and their parts can be used to learn a fine- grained representation of semantic contribu- tion. Finally, we argue such a representation captures compositionality better than the cur- rent status-quo which treats compositionality as a binary classification problem. 1 Introduction A Multiword Expressions (MWE) can be defined as a sequence of words whose meaning cannot nec- essarily be derived from the meaning of the words making up that sequence, for example: Rat Race — self-defeating or pointless pursuit1 MWEs are considered a “key problem for the de- velopment of large-scale, linguistically sound nat- ural language processing technology” (Sag et al., 2002). The challenge posed by MWEs is three- fold, consisting of MWE identification, classifica- tion and interpretation. Following the identification of a MWE, it needs to be established whether the expression should be treated as lexical (idiomatic) or as compositional. The final step, learning the se- mantics of the MWE, strongly depends on this deci- sion. 1Definition taken from Wikipedia, and clearly not recover- able if one only knows the meaning of the words ‘rat’ and ‘race’. The problem posed by MWEs is considered hard, but at the same time it is highly relevant and inter- esting. MWEs occur frequently in language and in- terpreting them correctly would directly improve re- sults in a number of tasks in NLP such as translation and parsing (Korkontzelos and Manandhar, 2010). By extension this makes deciding the lexicality of MWEs an important challenge for various fields in- cluding machine translation, question answering and information retrieval. In this paper we discuss com- positionality with respect to noun-noun compounds. Most Computational Linguistics literature treats compositionality as a binary problem, classifying compounds as either lexical or compositional. We show that this approach is too simplistic and argue for the real-valued treatment of compositionality. We propose two unsupervised models that learn compositionality rankings for compounds, placing them on a scale between lexical and compositional extremes. We develop a fine-grained representa- tion of compositionality using a novel generative ap- proach that models context as generated by com- pound constituents. This representation differenti- ates between the semantic contribution of both com- pound constituents as well as the compound itself. Comparing it with existing work in the field, we demonstrate the competitiveness of our approach. We evaluate on an existing corpus of noun com- pounds with ranked compositionality data, as well as on a large corpus with a binary annotation for lex- ical and compositional compounds. We analyse the impact of data sparsity and propose an interpolation …

Content truncated to 500 words