Statistical Phrase-Based Translation

The authors propose a new phrase-based translation model and decoding algorithm that enables evaluation and comparison of several previously proposed phrase-based translation models within a single framework. Using this framework they run a large number of experiments aimed at better understanding and explaining why phrase-based models outperform word-based models.

The empirical results, which hold for all examined language pairs, suggest the highest performance is obtained through relatively simple means: heuristic learning of phrase translations from word-based alignments and lexical weighting of phrase translations. Surprisingly, learning phrases longer than three words or from high-accuracy word-level alignment models does not strongly affect performance, and restricting learning to only syntactically motivated phrases degrades the systems, insights that shaped subsequent statistical machine translation.