| Sumario: | Human evaluations of translation are extensive but expensive. Human evaluations can take weeks to finish and involve human labor that cannot be reused. This paper proposes a method of automatic translation evaluation that is quick, inexpensive, and correlates highly with human evaluation. It uses natural language processing techniques to extract linguistic features from translated sentences and to model human analytic ratings. We discuss the methodology, development, and reliability of this system. Then we describe main approaches, such as the segmentation of target text, the recognition of passive voice and negative tense, the assignment of weighting scheme for phrases based on their scarcity, and the calculation of the phrase- and sentence-level semantic similarity between candidates' translation and several reference ones. Statistical results show that compared to human judges, this method is manageable, fast, and can effectively simulate human translation evaluation.
|
|---|