Standardised Evaluation of English Noun Compound Interpretation

We present a tagged corpus for English noun compound interpretation and describe the method used to generate them. In order to collect noun compounds, we extracted binary noun compounds (i.e. noun-noun pairs) by looking for sequences of two nouns in the POS tag data of the Wall Street Journal. We then manually filtered out all noun compounds which were incorrectly tagged or included proper nouns. This left us with a data set of 2,169 noun compounds, which we annotated using a set of 20 semantic relations defined by Barker and Szpakowicz (1998) allowing the annotators to assign multiple semantic relations if necessary. The initial agreement was 52.31%. The final data set contains 1,081 test noun compounds and 1,088 training noun compounds.