Towards a Hybrid Rule-based and Statistical Arabic-French Machine Translation System

Proceedings of Recent Advances in Natural Language Processing, pages 579–583, Hissar, Bulgaria, 7-13 September 2013. Towards a Hybrid Rule-based and Statistical Arabic-French Machine Translation System Fatiha Sadat University of Quebec in Montréal 201 President Kennedy, Montréal QC, Canada, H2X 3Y7 Sadat.fatiha@uqam.ca Abstract Arabic is a morphologically rich and complex language, which presents signifi- cant challenges for natural language pro- cessing and machine translation. In this paper, we describe an ongoing effort to build our first Arabic-French phrase– based machine translation system using the Moses decoder among other linguistic tools. The results show an improvement in the quality of translation and a gain in terms of Bleu score after introducing a pre-processing scheme for Arabic and ap- plying some rules based on morphological variations of the source language. The proposed approach is completed without increasing the amount of training data or changing radically the algorithms that can affect the translation or training engines. 1 Introduction Arabic is a morphologically rich and complex language, in which a word carries not only in- flections but also clitics, such as pronouns, con- junctions, and prepositions. It is a highly inflec- tional language, which makes the morphological analysis complicated. In Arabic, many coordinat- ing conjunctions, the definite article, many prep- ositions and particles, and a class of pronouns are all clitics that attach themselves either to the start or the end of words (Attia, 2008). This morpho- logical complexity has consequences on NLP applications, such as machine translation and information retrieval. One the one hand, developing an Arabic-French machine translation system is not an easy task, although there is a vast amount of training data nowadays. On the other hand, dealing with the complexity and ambiguity of the source language plays a major role in boosting the efficiency of the translation system. In previous research, it was shown that morpho- logical pre-processing of a morphologically rich language, such as Arabic does provide a benefit, especially in the case of limited volume of train- ing data (Goldwater and McClosky, 2005), (Sa- dat and Habash, 2006), (Lee, 2004), (El Ishibani et al., 2006), (Hasan et al., 2003). In Statistical Machine Translation (SMT) con- text, Habash et Sadat (Habash et Sadat, 2006) pre-processed Arabic texts using different seg- mentation schemes for translation into English and showed that the quality of translation is gen- erally better than the baseline. Similar findings were reported by (El Ishibani et al., 2006) on Arabic-English SMT. In relation to Arabic- French SMT, few research and evaluations were reported, compared to Arabic-English SMT among other pairs of languages. One of the first statistically-driven machine translation systems for Arabic-French was reported by Hasan et al (Hasan et al., 2006) during the second Cesta evaluation campaign1. The proposed SMT sys- tem used a simple stemming algorithm based on finite-state automata to split Arabic words into prefixes, stem and suffixes. Nevertheless, this simple segmentation method showed a reduced OOV rate from 8.2% to 2.6% for the test data and thus a better quality of translation in terms …

Content truncated to 500 words