System Combination for Machine Translation Based on Text-to-Text Generation

In this paper, we introduce a novel translation system combination framework using a text-to-text generation technique. We are moti-vated by the observation that for many transla-tion sentences, some of their constituents may be poorly translated, while others may be well translated. Moreover, it is often the case that other systems can provide good alternatives to problematic constituents. In our approach, the system constructs paraphrase lattices repre-senting all possible hypotheses for the same source sentence. We then use a text-to-text generator operating on those lattices to gener-ate those hypotheses. We filter ungrammatical combinations using a feature-based lexical-ized tree adjoining grammars (FB-LTAG) and then use a TER-based metric to compute a consensus score function to select the best translation among grammatical hypotheses. The system combination gains 1.38 BLEU points over the best individual system.