All in Strings: a Powerful String-based Automatic MT Evaluation Metric with Multiple Granularities

摘要

String-based metrics of automatic ma-chine translation (MT) evaluation are widely applied in MT research. Mean-while, some linguistic motivated me-trics have been suggested to improve the string-based metrics in sentence-level evaluation. In this work, we at-tempt to change their original calcula-tion units (granularities) of string-based metrics to generate new features. We then propose a powerful string-based automatic MT evaluation metric, com-bining all the features with various granularities based on SVM rank and regression models. The experimental results show that i) the new features with various granularities can contri-bute to the automatic evaluation of translation quality; ii) our proposed string-based metrics with multiple gra-nularities based on SVM regression model can achieve higher correlations with human assessments than the state-of- art automatic metrics.