Morphology Modeling for Statistical Machine Translation

Word-alignment is the initial step in most state-of-the art approaches to statistical machine translation. Morphologically rich languages pose problems for current statistical machine translation systems, including word-alignment, the most common problem being data sparsity.;Current word-alignment models do not take into account morphology beyond merely treating morphemes as words. We present a new word alignment model that distinguishes between words and morphemes. Our model does not collapse words and morphemes into one single category, therefore we can legitimately talk about words and their morphemes in line with the linguistic conception of these terms.;We adopt a two-level representation of alignment: the first level involves word alignment, the second level involves morpheme alignment in the scope of a given word alignment. Two-level alignment models (TAM) can align rarely occurring words through their frequently occurring morphemes. Our model induces word and morpheme alignments jointly using the expectation maximization algorithm.;We present the HMM extension of TAM, which is an instance of a multi-scale HMM. The two-level HMM we present addresses reordering between morpheme positions and word positions simultaneously.