Part Three of this five part report on Salton's Magical Automatic Retriever of Texts (SMART) project contains four papers. The first: "Variations on the Query Splitting Technique with Relevance Feedback" by T. P. Baker discusses some experiments in relevance feedback performed with variations on the technique of query splitting. The results indicate that these variations, as tested, offer no significant improvement over previously tried methods of query splitting. Paper two: "Effectiveness of Feedback Strategies on Collections of Differing Generality" by B. Capps and M. Yin evaluates the comparative effectiveness of several feedback strategies on the Cranfield 200 and Cranfield 400 Collections. Results are assessed from both the user and the system viewpoint; some strategies appeared equally effective on both collections. The third paper: "Selective Negative Feedback Methods" by M. Kerchner deals with experiments performed with several new methods of using nonrelevant retrieval documents to modify queries which retrieved no relevant material in the first N documents retrieved. Paper four: "The Use of Past Relevance Decisions in Relevance Feedback" by L. Paavola investigates some possibilities for exploiting the potential similarity among documents judged relevant to the same query in relevance feedback. (See also LI 002 719-LI 002 721 and LI 002 723-LI 002 724.) (NH)