Fusion of Multiple Corrupted Transmissions and its effect on Information Retrieval

Much previous work has focused on correction of OCR degraded text with little work addressing the possibility of fusing the generated text from different OCR systems, which are assumed to produce different types of errors. This paper explores text fusion, which involves the use of language modeling to determine which OCR system (if any) properly recognized individual words. The technique was applied on Arabic text that was synthetically degraded using different models of OCR degradation. The different degraded versions were consequently fused leading to a new version with significantly fewer errors than any of the original versions. Also, the effect of fusion on retrieval was examined. Index Terms – Text Fusion, OCR, Language Modeling, and Information Retrieval