Taming Structured Perceptrons on Wild Feature Vectors

Structured perceptrons are attractive due to their simplicity and speed, and have been used successfully for tuning the weights of binary features in a machine translation system. In attempting to apply them to tuning the weights of real-valued features with highly skewed distributions, we found that they did not work well. This paper describes a modification to the update step and compares the performance of the resulting algorithm to standard minimum error-rate training (MERT). In addition, preliminary results for combining MERT or structured-perceptron tuning of the log-linear feature weights with coordinate ascent of other translation system parameters are presented.