Smoothing parameter estimation framework for IBM word alignment models

IBM models are important word alignment models in Machine Translation. Based on the Maximum Likelihood Estimation principle to estimate their parameters, the models could easily overfit training data when data are sparse. Even though smoothing is a very popular solution in Language Model, there is still a lack of studies on smoothing for word alignment. In this paper, we propose a framework which generalizes the notable work Moore (2004) of applying additive smoothing to word alignment models. The framework allows developers to customize the smoothing amount for each pair of words. The added amount will be scaled appropriately by a common factor which reflects how much the framework trusts the adding strategy according to the performance on data. We also carefully examine various performance criteria and propose a smoothened version of the error count, which generally gives the best result.