Answer‐providing documents: Some inference descriptions and text‐searching retrieval results

Document‐answer inference descriptions were written for 20 question‐ (answer‐providing) document pairs randomly selected from 66 found in an earlier study. The first premise of each inference consisted of one or more passages quoted from the document. Some of the quotations were accompanied by bracketed additions, justified by other passages in the document, to make meanings explicit. Each inference included one or more background knowledge premises. In three cases, an inference also required an “author intent” premise because (e.g.) a document gave a reason for negotiating requests but did not say that it was giving a reason; and the question asked, “What reasons have been given … ?” Using the 20 question‐document pairs and the document corpus that had been searched to find them, a small‐scale study was mode of how effectively answer‐providing documents can be retrieved by text searching. Only rather simple procedures were considered. Specifically, answer‐providing document passages, their near contexts, and titles and headings regarded as adjacent were examined for stem and thesaurus matches to question words. Proximity of matching words in a document was used as a measure of relation. Matches to question words probably occurring infrequently in the collection were weighted more heavily. A procedure was formulated for ranking question‐document pairs for strength of match on the basis of these factors. As a test, the procedure was applied to a random 30 question‐document passage pairs in which the documents were not answer providing. One of these false pairs tied or ranked above 9 of the 20 correct question‐document pairs. Seven other false pairs each tied or outranked 1 or 2 of the correct pairs. These results indicate that the text‐searching procedures used are probably inadequate for high‐recall retrieval, but might be satisfactory for high‐precision retrieval.