Disambiguation of newly derived nominalizations in context: A Distributional Semantics approach

One of the central problems in the semantics of derived words is polysemy (see, for example, the recent contributions by Lieber 2016 and Plag et al. 2018 ). In this paper, we tackle the problem of disambiguating newly derived words in context by applying Distributional Semantics ( Firth 1957 ) to deverbal -ment nominalizations (e.g. bedragglement, emplacement). We collected a dataset containing contexts of low frequency deverbal -ment nominalizations (55 types, 406 tokens, see Appendix B) extracted from large corpora such as the Corpus of Contemporary American English. We chose low frequency derivatives because high frequency formations are often lexicalized and thus tend to not exhibit the kind of polysemous readings we are interested in. Furthermore, disambiguating low-frequency words presents an especially difficult task because there is little to no prior knowledge about these words from which their semantic properties can be extrapolated. The data was manually annotated according to eventive vs. non-eventive interpretations, allowing also an ambiguous label in those cases where the context did not disambiguate. Our question then was to what extent, and under which conditions, context-derived representations such as those of Distributional Semantics can be successfully employed in the disambiguation of low-frequency derivatives. Our results show that, first, our models are able to distinguish between eventive and non-eventive readings with some success. Second, very small context windows are sufficient to find the intended interpretation in the majority of cases. Third, ambiguous instances tend to be classified as events. Fourth, the performance of the classifier differed for different subcategories of nouns, with non-eventive derivatives being harder to classify correctly. We present indirect evidence that this is due to the semantic similarity of abstract non-eventive nouns to eventive nouns. Overall, this paper demonstrates that distributional semantic models can be fruitfully employed for the disambiguation of low frequency words in spite of the scarcity of available contextual information. 1