Improving Translation Fluency with Search-Based Decoding and a Monolingual Statistical Machine Translation Model for Automatic Post-Editing

The BLEU scores and translation fluency for the current state-of-the-art SMT systems based on IBM models are still too low for publication purposes. The major issue is that stochastically generated sentences hypotheses, produced through a stack decoding process, may not strictly follow the natural target language grammar, since the decoding process is directed by a highly simplified translation model and n-gram language model, and a large number of noisy phrase pairs may introduce significant search errors. This paper proposes a statistical post-editing (SPE) model, based on a special monolingual SMT paradigm, to “translate”disfluent sentences into fluent sentences. However, instead of conducting a stack decoding process, the sentence hypotheses are searched from fluent target sentences in a large target language corpus or on the Web to ensure fluency. Phrase-based local editing, if necessary, is then applied to correct weakest phrase alignments between the disfluent and searched hypotheses using fluent target language phrases; such phrases are segmented from a large target language corpus with a global optimization criterion to maximize the likelihood of the training sentences, instead of using noisy phrases combined from bilingually wordaligned pairs. With such search-based decoding, the absolute BLEU scores are much higher than automatic post editing systems that conduct a classical SMT decoding process. We are also able to fully correct a significant number of disfluent sentences into completely fluent versions. The BLEU scores are significantly improved. The evaluation shows that on average 46% of translation errors can be fully recovered, and the BLEU score can be improved by about 26%.