Fundamentals for reliability of automatic ranking

Statistical properties of population to be retrieved and index terms are demonstrated experimentally with a retrieval target of 367,363 items for 5.5 years in INIS data base and 6 query formulas for a search topic. Occurrence of hits (r) in the population is assured to be a stochastic event following very close approximation to a binominal distribution function. Confidence interval for precision of the population is estimated roughly to be 0.2 -- 0.3 with a sample size of 10 < np' = r < 40 and bellow 0.2 with 40 < r < 150. Weak dependencies are observed on occurrence of index terms in the population-, hit- and noise-sets respectively by both of correlation coefficient and expected mutual information measure. It is confirmed for this weakness that the dependency is right in principle for the occurrence but independency is also a justifiable assumption. Ranking performance of the population is affected by a sample size for weight generation. While a size of np' = r > 10 has a favourable influence upon the performance, a size of r < 10 doesn't produce any performance. Those data should be beneficial to creation and evaluation of various ranking schemes. (author)