Extended word similarity based clustering on unsupervised PoS induction to improve English-Indonesian statistical machine translation

In this paper, we present the unsupervised Part-of-Speech (PoS) induction algorithm to improve translations quality on statistical machine translation. The proposed algorithm is an extension of the algorithm Word-Similarity-Based (WSB) clustering. In the clustering, the similarity between words is measured by its grammatical relation with other words. The grammatical relation is represented as the n-gram relation. We extend the WSB clustering by take into account for the previous words in measuring the grammatical relation. The clustering results are then used in the English-Indonesia statistical machine translation. The experiments were conducted using MOSES as the machine translation decoder, and were evaluated by its BLEU score. Using 14.000 English-Indonesian sentence pairs, the clustering improved the BLEU score of 2.07%.