Named Entities Recognition (NER) has become one of the major issues in Information Retrieval (IR), knowledge extraction, and document classification. This paper addresses a particular case of NER, acronym expansion (or definition) when this expansion does not exist in the document using the acronym. Since acronyms may obviously expand into several distinct sets of words, this paper provides nine quality measures of the relevant definition prediction based on mutual information (MI), cubic MI (MI3), and Dice's coefficient. A combinaison of these statistical measures with the cosine approach is proposed. Experiments have been run on biomedical domain where acronyms are numerous. The results on our biomedical corpus showed that the proposed measures were accurate devices to predict relevant definitions.