Learning to Find Translations and Transliterations on the Web based on Conditional Random Fields

In recent years, state-of-the-art cross-linguistic systems have been based on parallel corpora. Nevertheless, it is difficult at times to find translations of a certain technical term or named entity even with a very large parallel corpora. In this paper, we present a new method for learning to find translations on the Web for a given term. In our approach, we use a small set of terms and translations to obtain mixed-code snippets returned by a search engine. We then automatically annotate the data with translation tags, automatically generate features to augment the tagged data, and automatically train a conditional random fields model for identifying translations. At runtime, we obtain mixed-code webpages containing the given term and run the model to extract translations as output. Preliminary experiments and evaluation results show our method cleanly combines various features, resulting in a system that outperforms previous works.