Novel score normalization methods for keyword search

摘要

In Keyword Search, the system scores belonging to different keywords vary in range due to the characteristics of the keyword and the audio that we search in. However, system decision of a given hit being relevant or irrelevant is made using the same threshold for all keywords. Hence the normalization of the scores of different keywords to the same scale is important. In this paper, we propose novel normalization techniques and test them on outputs of 4 different systems working on 307 Turkish keywords. We show that, the proposed methods outperform the commonly used known normalization techniques.