Unsupervised word clustering algorithms — which form word clusters based on a measure of distributional similarity — have proven to be useful in providing beneficial features for various natural language processing tasks involving supervised learning. This work explores the utility of such word clusters as factors in statistical machine translation.
Although some of the language pairs in this work clearly benefit from the factor augmentation, there is no consistent improvement in translation accuracy across the board. For all language pairs, the word clusters clearly improve translation for some proportion of the sentences in the test set, but has a weak or even detrimental effect on the rest.
It is shown that if one could determine whether or not to use a factor when translating a given sentence, rather substantial improvements in precision could be achieved for all of the language pairs evaluated. While such an “oracle” method is not identified, evaluations indicate that unsupervised word cluster are most beneficial in sentences without unknown words.