Retrieval of answer‐providing documents

(I) Better understanding of subject document retrieval might result if different functions of subject document retrieval systems are studied separately. This paper is concerned with retrieval of documents, in response to a question, from which answers to that question can be inferred (“answer‐providing documents”). “Answer can be inferred from document” has many possible meanings, one of which must be selected (an “inference specification”). Inasmuch as scientists in a field sometimes disagree about the correctness of inferences, have somewhat different background knowledge, etc., any inference specification can only approximate scientific inference practices. Two sources of systematic knowledge of document‐statement inference practices in a scientific field are described. (II) If a content word occurs in a question, then it occurs in any answer to that question (with some apparently tractable exceptions). An indexing procedure based on that fact is described which would permit retrieval of all answer‐providing documents for a question. However, because the indexing is “nonrelational,” it could cause false retrievals as well. Various ways of dealing with such false retrievals are briefly indicated, and a study is sketched that would provide data for helping selection among them. Two special points concerning indexing for retrieval of answer‐providing documents are discussed separately.