Jane 2: Open Source Phrase-based and Hierarchical Statistical Machine Translation

We present Jane 2, an open source toolkit supporting both the phrase-based and the hierarchical phrase-based paradigm for statistical machine translation. It is implemented in C++ and provides efficient decoding algorithms and data structures. This work focuses on the description of its phrase-based functionality. In addition to the standard pipeline, including phrase extraction and parameter optimization, Jane 2 contains several state-of-the-art extensions and tools. Forced alignment phrase training can considerably reduce rule table size while learning the translation scores in a more principled manner. Word class language models can be used to integrate longer context with a reduced vocabulary size. Rule table interpolation is applicable for different tasks, e.g. domain adaptation. The decoder distinguishes between lexical and coverage pruning and applies reordering constraints for efficiency.