Ad hoc Retrieval Using Thresholds, WSTs for French Mono-lingual Retrieval, Document-at-a-Glance for High Precision and Triphone Windows for Spoken Documents

This paper describes work done by a team from Dublin City University as part of TREC-6. In this TREC exercise we completed series of runs in 4 categories. The first was the mainline ad hoc retrieval task in which we repeated our entry for TREC-5, without modification. This is based on applying various thresholds to processing a query including query term and posting list thresholds, in order to improve retrieval efficiency. As our previous work has shown, this can be done without any loss in retrieval effectiveness. Our second set of submitted runs were as part of the crosslingual retrieval track where we ran French topics against French texts, effectively mono-lingual retrieval. What is novel about our approach is that it is based upon matching word shape tokens derived from character shape codes, rather than matching word stems or base forms. This technique is useful for retrieving from scanned document images rather than full texts and is something we are currently refining for English texts (and English queries). With those other experiments we have obtained surprisingly effective retrieval and this venture in TREC-6 was to see how effective WST-based retrieval could be for French. The third series of experiments we submitted were based on the high precision track in which we used a graphical representation of a ranked list of documents and the positional occurrences of search terms within those top-ranked documents, relative to each other. Our final experiments were as part of the spoken document retrieval track in which we removed the tags used for story bounds, turned transcripts and topics into a phonetic representation using a phoneme dictionary and we then retrieved story identifiers based on a triphone match between topic and fixed-width windows of triphones in the transcripts. We also applied a weighting function to triphones as they occurred in story “windows” based on their offset within those windows.