Improving the Accuracy of English-Arabic Statistical Sentence Alignment

Abstract

Multilingual natural language processing systems are increasingly relying on parallel corpus to ameliorate their output. Parallel corpora constitute the basic block for training a statistical natural language processing system and creating translation and language models. Several systems have been devised that automatically align words of a pair of sentences, each in a language. Such systems have been used successfully with European languages. In this paper, one such system is used to align sentences in an English-Arabic corpus. The system works poorly given raw unaligned sentence English-Arabic sentence pairs. This prompted the development of a preprocessing step to be applied to the Arabic sentences. The same corpus was then preprocessed and a significant improvement is reported when alignment is attempted using the preprocessed unaligned sentences.