Extending Sense Collocations in Interpreting Noun Compounds

Extending Sense Collocations in Interpreting Noun Compounds Su Nam Kim,♠♥ Meladel Mistica♠ and Timothy Baldwin♠♥ ♠ CSSE and ♥ NICTA VRL University of Melbourne VIC 3010, Australia {snkim,mmistica,tim}@csse.unimelb.edu.au Abstract This paper investigates the task of noun compound interpretation, building on the sense collocation approach proposed by Moldovan et al. (2004). Our primary task is to evaluate the impact of similar words on the sense collocation method, and decrease the sensitivity of the classifiers by expanding the range of sense collocations via different semantic relations. Our method combines hypernyms, hyponyms and sister words of the component nouns, based on WORDNET. The data used in our experiments was taken from the nominal pair interpretation task of SEMEVAL-2007 (4th International Work- shop on Semantic Evaluation 2007). In our evaluation, we test 7-way and 2-way class data, and show that the inclusion of hyper- nyms improves the performance of the sense collocation method, while the inclusion of hyponym and sister word information leads to a deterioration in performance. 1 Introduction This paper investigates the automatic interpretation of noun compounds (i.e. NCs), such as paper sub- mission and computer science department. NC in- terpretation is a well-known problem that aims to predict the semantic relation (i.e. SR) between a head noun and its modifying nominal(s) (i.e. mod- ifiers). SRs, simply put, encapsulate how the head noun and the other nominals in a noun compound are related. As English noun phrases are right- headed, the head noun occurs after all modifying nouns. For example, brick house is interpreted as a house that is modified by the word brick, which ex- hibits a PRODUCT-PRODUCER relationship between the two nouns in the compound. In contrast, the modifier and head in house brick exhibits a PART- WHOLE relationship, which is interpreted as a brick from a house, rather than the former interpretation of a house made of bricks. The set of SRs that we are concerned with in this paper is defined in Sec- tion 5.1. Research on NCs can be categorised into four main groups: defining SRs, disambiguating the syn- tax of NCs, disambiguating the semantics of NCs, and interpreting NCs via SRs. Each task is detailed in Section 2.1. Interpreting NCs has received much attention of late, and the problem has been addressed in areas of machine translation (MT), information extraction (IE), and applications such as question- answering (QA). NCs pose a considerable challenge to computational linguistics due to the following is- sues (Lapata, 2002): (1) NCs are extremely produc- tive; (2) the semantic relationship between the head noun and its modifier(s) is implicit; and (3) the in- terpretation of an NC can vary due to contextual and pragmatic factors. Due to these challenges, current NC interpretation methods are too error-prone to be employed directly in NLP applications without hu- man intervention or preprocessing. In this paper, we investigate the task of NC in- terpretation based on sense collocation. It has been shown that NCs with semantically similar compo- nents share the same SR (Kim and Baldwin, …

Content truncated to 500 words