A Systematic Comparison of SVM and Maximum Entropy Classifiers for Translation Error Detection

In recent years, the translation error detection or confidence estimation for SMT has been becoming a hot question, especially in the localization industry. This paper mainly focuses on a systematic comparison on two different classifiers Maximum Entropy (MaxEnt) and SVM over different features to illustrate their error detection capabilities. Three typical word posterior probabilities (WPP) and three linguistic features are introduced and fairly compared over two classifiers on Chinese to-English NIST datasets. Experimental results show that the combination of WPP with linguistic features can significantly reduce the CER, and the SVM classifier outperforms the MaxEnt classifier in terms of the CER and F measure.