Factored Machine Translation
Speaker: Hieu Hoang
Institution: University of Edinburgh
Abstract:
Current phrase-based statistical machine translation (SMT) systems typically translate by using phrase dictionaries learned from a parallel corpus. However, the decoder uses only the surface word forms to translate.
We have developed a system, Moses, which can make use of other word factors, such as lemma, POS tags or morphology in a consistent and flexible way, allowing linguistic knowledge of the source and target language to improve translation quality.
Also, an important ability of Moses is the decoding of ambiguous input which are modeled with confusion networks. We describe how this can be used to translate output from automatic speech recognition systems or source sentences with ambiguous tags.