Morphological Modeling for Machine Translation of English-Iraqi Arabic Spoken Dialogs

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 995–1000, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Morphological Modeling for Machine Translation of English-Iraqi Arabic Spoken Dialogs Katrin Kirchhoff Department of Electrical Engineering University of Washington Seattle, WA, USA kk2@u.washington.edu Wilson Tam Microsoft Corporation wilson.yctam@gmail.com Colleen Richey, Wen Wang SRI International Menlo Park, CA, USA colleen@speech.sri.com wwang@speech.sri.com Abstract This paper addresses the problem of mor- phological modeling in statistical speech-to- speech translation for English to Iraqi Ara- bic. An analysis of user data from a real-time MT-based dialog system showed that generat- ing correct verbal inflections is a key problem for this language pair. We approach this prob- lem by enriching the training data with mor- phological information derived from source- side dependency parses. We analyze the per- formance of several parsers as well as the ef- fect on different types of translation models. Our method achieves an improvement of more than a full BLEU point and a significant in- crease in verbal inflection accuracy; at the same time, it is computationally inexpensive and does not rely on target-language linguistic tools. 1 Introduction SMT from a morphologically poor language like En- glish into a language with richer morphology con- tinues to be a problem, in particular when training data is sparse and/or the SMT system has insufficient modeling capabilities for morphological variation in the target language. Most previous approaches to this problem have utilized a translate-and-inflect method, where a first-pass SMT system is trained on lemmatized forms, and the correct inflection for every word is predicted in a second pass by statis- tical classifiers trained on a combination of source and target language features. This paper looks at morphological modeling from a different perspec- tive, namely to improve SMT in a real-time speech- to-speech translation system. Our focus is on resolv- ing those morphological translation errors that are most likely to cause confusions and misunderstand- ings in machine-translation mediated human-human dialogs. Due to the constraints imposed by a real- time system, previous approaches that rely on elabo- rate feature sets and multi-pass processing strategies are unsuitable for this problem. The language pair of interest in this study is English and Iraqi Arabic (IA). The latter is a spoken dialect of Arabic with few existing linguistic resources. We therefore de- velop a low-resource approach that relies on source- side dependency parses only. We analyze its perfor- mance in combination with different types of parsers and different translation models. Results show a sig- nificant improvement in translation performance in both automatic and manual evaluations. Moreover, the proposed method is sufficiently fast for a real- time system. 2 Prior Work Much work in SMT has addressed the issue of translating from morphologically-rich languages by preprocessing the source and/or target data by e.g., stemming and morphological decomposition (Popovic and Ney, 2004; Goldwater and McClosky, 2005), compound splitting (Koehn and Knight, 2003), or various forms of tokenization (Lee, 2004; Habash and Sadat, …

Content truncated to 500 words