Automatic Noun Compound Interpretation using Deep Neural Networks and Word Embeddings

The present paper reports on the results of automatic noun compound interpretation for English using a deep neural network classifier and a selection of publicly available word embeddings to represent the individual compound constituents. The task at hand consists of identifying the semantic relation that holds between the constituents of a compound (e.g. WHOLE+PART_OR_MEMBER_OF in the case of ‘robot arm’, LOCATION in the case of ‘hillside home’). The experiments reported in the present paper use the noun compound dataset described in Tratz (2011), a revised version of the dataset used by Tratz and Hovy (2010) for training their Maximum Entropy classifier. Our experiments yield results that are comparable to those reported in Tratz and Hovy (2010) in a crossvalidation setting, but outperform their system on unseen compounds by a large margin.