One problem of machine-learning-based prediction of structure-odor relationship is that odorant molecules are usually labeled with ambiguous descriptors when they are collected from different sources. This study focused on the clustering of the odor descriptors by text mining approaches as well as the prediction of newly established labels from physicochemical parameters of the classified odorant molecules. An odor database was established by web scraping and transferred to a document-Term matrix including 4011 odorants and 100 odor descriptors. The clustering of the odor descriptors was carried out by using different co-occurrence matrix and clustering approaches. A hierarchical cluster analysis combined with a co-occurrence probability distribution matrix has shown good results in the descriptor clustering. The attribute labels of each class were established and then predicted from physicochemical parameters of the classified odorants by using random forest model. An average accuracy higher than 82.42% was obtained, indicating the effectiveness of the proposed approaches for predicting structure-odor relationship.