Joint morphological-lexical language modeling for machine translation

We present a joint morphological-lexical language model (JMLLM) for use in statistical machine translation (SMT) of language pairs where one or both of the languages are morphologically rich. The proposed JMLLM takes advantage of the rich morphology to reduce the Out-Of-Vocabulary (OOV) rate, while keeping the predictive power of the whole words. It also allows incorporation of additional available semantic, syntactic and linguistic information about the morphemes and words into the language model. Preliminary experiments with an English to Dialectal-Arabic SMT system demonstrate improved translation performance over trigram based baseline language model.