Deep Syntax in Statistical Machine Translation

This work investigates the current performance capabilities of LFG f-structure based transfer machine translation. Our empirical evaluation compares transfer-based machine translation performance to state of the art machine translation. Our investigation reveals that although the LFG-based approach under-performs compared to state of the art method in general, when the evaluation is restricted to translations where the target language f-structure falls within coverage of the generation grammar, the LFG-based system can in fact achieve higher coverage of unseen data in addition to improvements in translation quality.