Maryland at FIRE 2011: Retrieval of OCRed Bengali

In this year's Forum for Information Retrieval Evaluation (FIRE), the University of Maryland participated in the Retrieval of Indic Script OCRed Text (RISOT) task to experiment with the retrieval of Bengali script OCR’d documents. The experiments focused on evaluating a retrieval strategy motivated by recent work on Cross-Language Information Retrieval (CLIR), but which makes use of OCR error modeling rather than parallel text alignment. The approach obtains a probability distribution over substitutions for the actual query terms that possibly correspond to terms in the document representation. The results reported indicate that this is a promising way of using OCR error modeling to improve CLIR.