Retrieving Images of Scanned Text Documents

Information retrieval is the task of nding documents, usually text, which are relevant to a user's information need. A conventional approach to information management of paper documents is normally based on classifying them into a hierarchical classiication structure. More recently we have seen electronic document management systems which manage scanned images of documents in the same way as paper, or which do some OCR to determine machine-readable document content which may then be used for content-based information retrieval. In this paper we present an alternative to full-scale OCR of scanned document images in which words in scanned images of documents are represented as word shape tokens (WSTs) representing the approximate shape of words in text. We describe an approach to WST-based information retrieval that can be entirely automatic or can involve the user in reening the retrieval process. To measure the eeectiveness of WST-based retrieval we have implemented and reened it on two collections of documents/queries/relevance judgments, one in English, the other in French. Each of these has over 250 Mbytes of ASCII text and in indexing documents we have faithfully recreated the kind and frequency of errors expected in a WST recognition process. When compared to word-based retrieval, the eeectiveness of WST-based retrieval is surprisingly good, though erratic across diierent queries. We outline some of the directions we are pursuing in order to address this inconsistency of performance across queries and our results show real potential for WST-based retrieval of scanned document images.