MELB-KB: Nominal Classification as Noun Compound Interpretation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 231–236, Prague, June 2007. c⃝2007 Association for Computational Linguistics MELB-KB: Nominal Classification as Noun Compound Interpretation Su Nam Kim and Timothy Baldwin Computer Science and Software Engineering University of Melbourne, Australia {snkim,tim}@csse.unimelb.edu.au Abstract In this paper, we outline our approach to interpreting semantic relations in nominal pairs in SemEval-2007 task #4: Classifica- tion of Semantic Relations between Nomi- nals. We build on two baseline approaches to interpreting noun compounds: sense col- location, and constituent similarity. These are consolidated into an overall system in combination with co-training, to expand the training data. Our two systems attained an average F-score over the test data of 58.7% and 57.8%, respectively. 1 Introduction This paper describes two systems entered in SemEval-2007 task #4: Classification of Semantic Relations between Nominals. A key contribution of this research is that we examine the compatibility of noun compound (NC) interpretation methods over the extended task of nominal classification, to gain empirical insight into the relative complexity of the two tasks. The goal of the nominal classification task is to identify the compatibility of a given semantic re- lation with each of a set of test nominal pairs, e.g. between climate and forest in the fragment the climate in the forest with respect to the CONTENT- CONTAINER relation. Semantic relations (or SRs) in nominals represent the underlying interpretation of the nominal, in the form of the directed relation between the two nominals. The proposed task is a generalisation of the more conventional task of interpreting noun compounds (NCs), in which we take a NC such as cookie jar and interpret it according to a pre-defined inventory of semantic relations (Levi, 1979; Vanderwende, 1994; Barker and Szpakowicz, 1998). Examples of seman- tic relations are MAKE,1, as exemplified in apple pie where the pie is made from apple(s), and POSSES- SOR, as exemplified in family car where the car is possessed by a family. In the SemEval-2007 task, SR interpretation takes the form of a binary decision for given nominal pair in context and a given SR,a in judging whether that nominal pair conforms to the SR. Seven relations were used in the task: CAUSE-EFFECT, INSTRUMENT-AGENCY, PRODUCT-PRODUCER, ORIGIN-ENTITY, THEME- TOOL, PART-WHOLE and CONTENT-CONTAINER. Our approach to the task was to: (1) naively treat all nominal pairs as NCs (e.g. the climate in the for- est is treated as an instance of climate forest); and (2) translate the individual binary classification tasks into a single multiclass classification task, in the in- terests of benchmarking existing SR interpretation methods over a common dataset. That is, we take all positive training instances for each SR and pool them together into a single training dataset. For each test instance, we make a prediction according to one of the seven relations in the task, which we then map onto a binary classification for final evaluation purposes. This mapping is achieved by determining which binary SR classification the test instance was sourced from, and returning a …

Content truncated to 500 words