Minimally Supervised Morphological Segmentation with Applications to Machine Translation

Minimally Supervised Morphological Segmentation with Applications to Machine Translation Jason Riesa and David Yarowsky Department of Computer Science Johns Hopkins University Baltimore, MD 21218 {riesa, yarowsky}@jhu.edu Abstract Inflected languages in a low-resource set- ting present a data sparsity problem for statistical machine translation. In this paper, we present a minimally super- vised algorithm for morpheme segmen- tation on Arabic dialects which reduces unknown words at translation time by over 50%, total vocabulary size by over 40%, and yields a significant increase in BLEU score over a previous state-of-the- art phrase-based statistical MT system. 1 Introduction A significant problem for statistical machine transla- tion in a low-resource setting is data sparsity caused by highly inflected languages. Languages such as Arabic and Turkish, for example, are rich with com- plex morphology. For statistical machine transla- tion, this means much more parallel data is required in order to accurately learn translations for the in- creased number of unique words and phrases that result. In this paper, we focus on the problem of pre- fix and suffix morpheme segmentation with the ul- timate goal of increasing the accuracy of statistical machine translation systems by reducing the amount of unknown words at translation time as well as re- ducing total vocabulary size. Table 1 shows that it is nontrivial to make a correct analysis by inspection of the word surface form alone. A na¨ıve rule-based scheme employing a stemmed word list may yield Possible Inflected Uninflected Correct Affix Analysis Analysis Gloss [b] b+ syArp bsyArp by car [b] b+ TAryAt bTAryAt batteries [Al] Al+ mATwrAt AlmATwrAt the motors [Al] Al+ ly Ally that/which [y] Ax +y Axy my brother [y] mksyk +y mksyky Mexican Table 1. Examples of correct and incorrect Iraqi Arabic morpheme segmentation analyses for three different affixes: [b], [Al], and [y]. The cor- rect analysis in each case is shown in bold. adequate results, but only if the word list is com- plete and covers all unique word types found in a corpus. Not only are these resources difficult to ob- tain for low-resource languages, such methods are not robust to gaps in the lexicon and fail to indicate a context-based preference when multiple analyses are present. With the aid of a small lexicon annotated for morphological segmentation and part-of-speech, we employ a supervised trie-based segmentation model trained on relatively little data. This model for morpheme segmentation is then evaluated on both Iraqi and Levantine Arabic – two dialectal variants of Modern Standard Arabic (MSA). Few linguis- tic resources and little parallel or monolingual data are currently available for these two dialects. We 185 Proceedings of the 7th Conference of the Association for Machine Translation in the Americas, pages 185-192, Cambridge, August 2006. ©2006 The Association for Machine Translation in the Americas Corpus Statistic English Iraqi Levantine Utterances 36, 895 36, 895 43, 471 Running words 438, 911 305, 889 247, 741 Words per utterance 11.9 8.3 5.7 Unique words 8, 776 29, 238 26, 147 Table 2. Summary statistics for the …

Content truncated to 500 words