| This work explores the linguistic factors that influence the lexical choice of prepositions in machine translation. To determine the relative importance of bilingual and monolingual influences on the lexical selection of prepositions we conduct an empirical investigation, using several bilingual corpora. A simple statistical translation model incorporating these factors is proposed. The model is trained on bilingual data automatically extracted from the French-English Hansard Corpus.; In addition, we present a novel machine translation evaluation methodology, based on cloze testing (Taylor, 1953). This methodology, called the Appropriate Cloze MT Evaluation (ACME) method, makes fine-grained distinctions in the evaluation of MT output, and has six principal advantages: (1) It is objective; (2) It is replicable; (3) It produces unambiguously quantifiable results; (4) It provides the effect of basing the evaluation of the MT systems on multiple human translations of the test text, while only requiring a single full professional translation; (5) It enforces a focus on the intended object of evaluation; (6) It can provide insight into the factors that influence the success of the translation. The ACME methodology is used to evaluate the performance of two commercial (COTS) MT engines, providing a baseline for comparison with our prepositional translation component.; The simple statistical preposition translation component containins monolingual and bilingual prepositional selection models, at least one of which produces the correct preposition at a rate of 82.47%. The models are controlled by an arbiter, which judges which model can provide the correct preposition in a given situation. In cases where at least one of the models provides a correct response, the arbiter, based on Resnik's (1993, 1996) selectional preference strength, identifies the correct model 76.88% of the time, resulting in an overall rate of 63.40% correct for the prepositional translation component, ranking between the two COTS systems; The prepositional selection models are trainable, and could be integrated into existing NIT architectures, or added to machine-assisted post-editing tools. Both cross-lingual and monolingual information retrieval could benefit by integrating elements of the models to distinguish prepositions that should be excluded as stopwords from those that should be included among the search terms. |
