Improving alignment quality in statistical machine translation using context-dependent maximum entropy models
Ismael García-Varea, Franz Josef Och, Hermann Ney, Francisco Casacuberta · 2002
Typically, statistical alignment models are based on single-word dependencies. These models do not include contextual information, which can lead to inadequate alignments. In this paper, we present an approach to include contextual dependencies in the statistical alignment model by using a refined lexicon model. Unlike previous work, we directly integrate this in the EM algorithm of statistical alignment models. Experimental results are given for the French-English Canadian Parliament Hansards task and the Verbmobil task. The evaluation is performed by comparing the obtained alignments with a manually annotated reference alignment.