Towards Robust Animacy Classification Using Morphosyntactic Distributional Features

This paper presents results from experiments in automatic classification of animacy for Norwegian nouns using decision-tree classifiers. The method makes use of relative frequency measures for linguistically motivated morphosyntactic features extracted from an automatically annotated corpus of Norwegian. The classifiers are evaluated using leave-one-out training and testing and the initial results are promising (approaching 90% accuracy) for high frequency nouns, however deteriorate gradually as lower frequency nouns are classified. Experiments attempting to empirically locate a frequency threshold for the classification method indicate that a subset of the chosen morphosyntactic features exhibit a notable resilience to data sparseness. Results will be presented which show that the classification accuracy obtained for high frequency nouns (with absolute frequencies >1000) can be maintained for nouns with considerably lower frequencies (~50) by backing off to a smaller set of features at classification.