Lexical Features for Statistical Machine Translation

In modern phrasal and hierarchical statistical machine translation systems, two major features model translation: rule translation probabilities and lexical smoothing scores. The rule translation probabilities are computed as maximum likelihood estimates (MLEs) of an entire source (or target) phrase translating to a target (or source) phrase. The lexical smoothing scores are also a likelihood estimate of a source (target) phrase translating to a target (source) phrase, but they are computed using independent word-to-word translation probabilities. Intuitively, it would seem that the lexical smoothing score is a less powerful estimate of translation likelihood due to this independence assumption, but I present the somewhat surprising result that lexical smoothing is far more important to the quality of a state-of-the-art hierarchical SMT system than rule translation probabilities. I posit that this is due to a fundamental data sparsity problem: The average word-to-word translation is seen many more times than the average phrase-to-phrase translation, so the word-to-word translation probabilities (or lexical probabilities) are far better estimated. Motivated by this result, I present a number of novel methods for modifying the lexical probabilities to improve the quality of our MT output. First, I examine two methods of lexical probability biasing, where for each test document, a set of secondary lexical probabilities are extracted and interpolated with the primary lexical probability distribution. Biasing each document with the probabilities extracted from its own first-pass decoding output provides a small but consistent gain of about 0.4 BLEU. Second, I contextualize the lexical probabilities by factoring in additional information such as the previous or next word. The key to the success of this context-dependent lexical smoothing is a backoff model, where our "trust" of a context-dependent probability estimation is directly proportional to how many times it was seen in the training. In this way, I avoid the estimation problem seen in translation rules, where the amount of context is high but the probability estimation is inaccurate. When using the surrounding words as context, this feature provides a gain of about 0.6 BLEU on Arabic and Chinese. Finally, I describe several types of discriminatively trained lexical features, along with a new optimization procedure called Expected-BLEU optimization. This new optimization procedure is able to robustly estimate weights for thousands of decoding features, which can in effect discriminatively optimize a set of lexical probabilities to maximize BLEU. I also describe two other discriminative feature types, one of which is the part-of-speech analogue to lexical probabilities, and the other of which estimates training corpus weights based on lexical translations. The discriminative features produce a gain of 0.8 BLEU on Arabic and 0.4 BLEU on Chinese.