Modeling score distributions in information retrieval

摘要: We review the history of modeling score distributions, focusing on mixture normal-exponential by investigating theoretical as well empirical evidence supporting its use. discuss previously suggested conditions which valid binary models should satisfy, such Recall-Fallout Convexity Hypothesis, and formulate two new hypotheses considering component individually in pairs, under some limiting parameter values. From all mixtures past, current argument points to gamma most-likely universal model, with being a usable approximation. Beyond contribution, we provide experimental showing vector space or geometric models, BM25, `friendly' normal-exponential, that non-convexity problem possesses is practically not severe. Furthermore, recent non-binary speculate graded relevance, consider methods logistic regression for calibration.