Revisiting Known-Item Retrieval in Degraded Document Collections

Abstract

Optical character recognition software converts an image of text to a text document but typically degrades the document's contents. Correcting such degradation to enable the document set to be queried effectively is the focus of this work. The described approach uses a fusion of substring generation rules and context aware analysis to correct these errors. Evaluation was facilitated by two publicly available datasets from TREC-5's Confusion Track containing estimated error rates of 5% and 20%. On the 5% dataset, we demonstrate a statistically significant improvement over the prior art and Solr's mean reciprocal rank (MRR). On the 20% dataset, we demonstrate a statistically significant improvement over Solr, and have similar performance to the prior art. The described approach achieves an MRR of 0.6627 and 0.4924 on collections with error rates of approximately 5% and 20% respectively.

Searching Corrupted Document Collections

“…This suggests that 2 may be the best candidate vector size for most applications. This is supported by previous research ( [24], [19]) which show that a balance of context-free and context-dependent candidates perform best. By using a fusion method with a candidate vector size of 2, we select 2 candidates based on context (selected using bigrams) and 2 candidates not based on context (selected using Segments' substrings rules).…”

“…For example, we don't include symbols as white space when tokenizing because of the increased probability of OCR systems to replace characters with punctuation. A detailed explanation of the filtering process is described in [19]. We then iterate over all the filtered terms, tasking each evaluated approach to generate substitution candidates for the unrecognized terms.…”

“…Over the past years, we evaluated methods for reliably correcting phase one errors via post-processing using our method called Segments [17], [18], [19]. Segments differs from previous research in that it is an unsupervised approach, which makes minimal assumptions about resource availability, and has no dependence on language within the algorithm.…”

“…It uses substring rules to correct phase one errors. More recent research has focused on the role of balance between context, such as word-level bi-grams and tri-grams, and term focused substrings, finding that a fusion between the two improves performance [19].…”