Performance improvement of Machine Translation system using LID and post-editing

Abstract

Machine Translation systems are fast replacing human translators to aid us in translation between any pair of natural languages. Conversational sentences form the fundamental structure of a natural language. These sentences help in understanding the basic conversations in a language and hence improving translation of these sentences is considered more essential to improve the quality of a good translation systems in various real-life applications. This paper focuses on machine translation for travel conversations between any pair of languages namely, English, Hindi, and Tamil, with minimal intervention of linguistics. The translation system is trained using various text corpus sizes (1K, 5K, 13K sentences) of conversational and complex sentences for comparison. To improve the performance of a machine translation system a language identification system coupled with post-editing approach that applies n-best translation list analysis and language models are used. An improvement of translation performance by 3-5% in bilingual evaluation understudy (BLEU) score is observed without incorporating linguistic information.