Taming Structured Perceptrons on Wild Feature Vectors

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 384–391, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Taming Structured Perceptrons on Wild Feature Vectors Ralf D. Brown Carnegie Mellon University Language Technologies Institute 5000 Forbes Avenue, Pittsburgh PA 15213 USA ralf+@cs.cmu.edu Abstract 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 up- date step and compares the performance of the resulting algorithm to standard min- imum error-rate training (MERT). In ad- dition, preliminary results for combining MERT or structured-perceptron tuning of the log-linear feature weights with coordi- nate ascent of other translation system pa- rameters are presented. 1 Introduction Structured perceptrons are a relatively recent (Collins, 2002) update of the classic perceptron algorithm which permit the prediction of vec- tors of values. Initially developed for part of speech taggers, they have been applied to tuning the weights of the features in the log-linear mod- els used by statistical machine translation (Arun and Koehn, 2007), and found to have performance similar to the Margin-Infused Relaxed Algorithm (MIRA) by Crammer and Singer (2003; 2006) and Minimum-Error Rate Training (MERT) by Och (2003). Parameter tuning is an important aspect of current data-driven machine translation systems, as an improper selection of feature weights can dramatically reduce scores on evaluation metrics such as BLEU (Papineni et al., 2002) or METEOR (Banerjee and Lavie, 2005). When we recently added new features to the CMU-EBMT translation system (Brown, 1996; Brown, 2008)1, in addition to splitting a number of composite features into their components, our pre- vious method of parameter tuning via coordinate ascent2 became impractical. With now more than 50 features partaking in the scoring model, MERT no longer seemed a good choice, as the common wisdom is that it is not able to reliably optimize more than about 20 features (Chiang et al., 2008). We had been using coordinate ascent because of a need to tune a substantial number of parameters which are not directly part of the log-linear model which can be tuned by MERT or similar methods. Our system generates a translation lattice by run- time lookup in the training corpus rather than us- ing a precomputed phrase table, so important pa- rameters include • the size of the sample of retrieved training instances for a given input phrase which are aligned, • the weight of source features for ranking training instances during sampling, and • the minimum alignment score to accept a translation instance Decoder parameters which are important to tune, but which are generally not mentioned in the liter- ature include • how many alternative translations of a phrase to consider during decoding, • the size of the reordering window, and • the rank of the …

Content truncated to 500 words