Are Morpho-Syntactic Features More Predictive for the Resolution of Noun Phrase Coordination Ambiguity than Lexico-Semantic Similarity Scores?

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 89–96 Manchester, August 2008 Are Morpho-Syntactic Features More Predictive for the Resolution of Noun Phrase Coordination Ambiguity than Lexico-Semantic Similarity Scores? Ekaterina Buyko and Udo Hahn Jena University Language & Information Engineering (JULIE) Lab F¨urstengraben 30, 07743 Jena, Germany ekaterina.buyko|udo.hahn@uni-jena.de Abstract Coordinations in noun phrases often pose the problem that elliptified parts have to be reconstructed for proper semantic inter- pretation. Unfortunately, the detection of coordinated heads and identification of el- liptified elements notoriously lead to am- biguous reconstruction alternatives. While linguistic intuition suggests that semantic criteria might play an important, if not su- perior, role in disambiguating resolution alternatives, our experiments on the re- annotated WSJ part of the Penn Treebank indicate that solely morpho-syntactic crite- ria are more predictive than solely lexico- semantic ones. We also found that the combination of both criteria does not yield any substantial improvement. 1 Introduction Looking at noun phrases such as ‘cat and dog owner’ ‘novels and travel books’ their proper coordination reading (and asymmetric distribution of coordinated heads) as ‘cat owner’ AND ‘dog owner’ ‘novels’ AND ‘travel books’ seems to be licensed by the striking semantic sim- ilarity between ‘cat’ and ‘dog’, and ‘novels’ and ‘books’, respectively. If this were a general rule, then automatic procedures for the resolution of co- ordination ambiguities had to rely on the a priori provision of potentially large amounts of semantic c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. background knowledge to make this similarity ex- plicit. Furthermore, any changes in languages or domains where such resources were missing (or, were incomplete) would severely hamper coordi- nation analysis. Indeed, previous research has gathered lot of evidence that conjoined elements tend to be se- mantically similar. The important role of seman- tic similarity criteria for properly sorting out con- juncts was first tested by Resnik (1999). He in- troduced an information-content-based similarity measure that uses WORDNET (Fellbaum, 1998) as a lexico-semantic resource and came up with the claim that semantic similarity is helpful to achieve higher coverage in coordination resolution for co- ordinated noun phrases of the form ‘noun1 and noun2 noun3’ than similarity measures based on morphological information only. In a similar vein, Hogan (2007b) inspected WORDNET similarity and relatedness measures and investigated their role in conjunct identifi- cation. Her data reveals that several measures of semantic word similarity can indeed detect conjunct similarity. For the majority of these similarity measures, the differences between the mean similarity of coordinated elements and non- coordinated ones were statistically significant. However, it also became evident that these were only slight differences, and not all coordinated heads were semantically related as evidenced, e.g., by ‘work’/‘harmony’ in ‘hard work and harmony’. The significance tests did also not reveal particu- larly useful measures for conjunct identification. Rus et al. (2002) in an earlier study presented an alternative heuristics-based approach to conjunct identification for coordinations of the form ‘noun1 and noun2 …

Content truncated to 500 words