A Linguistically Motivated MT Evaluation System Based on SVM Regression

This paper describes the automatic MT evaluation system by Machine Intelligence and Translation Lab. of Harbin Institute of Technology, for NIST MetricsMATR 2008 evaluation. The system is based on SVM regression and employed many linguistic features. Machine-learning based automatic MT evaluation has recently gained more attention in NLP research. And feature mining is essential for ML. In this work, a novel feature, letterbased BLEU, is introduced, and it turns out to be of good capability to correlate with human assessments and contributes to the system’s final performance.