Semi-Automatic Parallel Corpora Extraction from Comparable News Corpora

The parallel corpus is a necessary resource in many multi/cross lingual natural language processing applications that include Machine Translation and Cross Lingual Information Retreival. Preparation of large scale parallel corpus takes time and also demands the linguistics skill. In the present work, a technique has been developed that extracts parallel corpus between Manipuri, a morphologically rich and resource constrained Indian language and English from a comparable news corpora collected from the web. A medium sized Manipuri-English bilingual lexicon and another list of Manipuri-English transliterated entities have been developed and used in the present work. Using morphological information for the agglutinative and inflective Manipuri language, the alignment quality based on similarity measure is further improved. A high level of performance is desirable since errors in sentence alignment cause further errors in systems that use the aligned text. The system has been evaluated and error analysis has also been carried out. The technique shows its effectiveness in Manipuri-English language pair and is extendable to other resource constrained, agglutinative and inflective Indian languages.