Abstract
Prepositions are among the most frequent words in English and play complexroles in the syntax and semantics of sentences. Not surprisingly, they posewell-known difficulties in automatic processing of sentences (prepositionalattachment ambiguities and idiosyncratic uses in phrases). Existing methods onpreposition representation treat prepositions no different from content words(e.g., word2vec and GloVe). In addition, recent studies aiming at solvingprepositional attachment and preposition selection problems depend heavily onexternal linguistic resources and use dataset-specific word representations. Inthis paper we use word-triple counts (one of the triples being a preposition)to capture a preposition's interaction with its attachment and complement. Wethen derive preposition embeddings via tensor decomposition on a largeunlabeled corpus. We reveal a new geometry involving Hadamard products andempirically demonstrate its utility in paraphrasing phrasal verbs. Furthermore,our preposition embeddings are used as simple features in two challengingdownstream tasks: preposition selection and prepositional attachmentdisambiguation. We achieve results comparable to or better than thestate-of-the-art on multiple standardized datasets.