An Elastic-Phrase Model for Statistical Machine Translation

We present some on-going research on phrase-based Statistical Machine Translation using flexible phrases that may contain gaps of variable lengths. This allows us to naturally handle various linguistic phenomena such as negations or separable particles. We integrate this within the standard Maximum Entropy model using some dedicated feature functions, and describe a beam-search stack decoder that handles these noncontiguous, elastic phrases. Preliminary experimental results show that the translation performance compares favourably with phrase-based MT using fixed gap size. We expect that future results may allow us to leverage the added flexibility of elastic chunks to further increase translation performance.