In the fields of opinion mining and sentiment analysis, Pang et al. (2002), Turney (2002) and Liu (2012, 2015), among others, have focused on extracting positive and negative opinions expressed in the text and the targets of these opinions. In contrast, beyond the opinion polarity and its target, we propose a corpus-based model that detects different evaluative language. Based on this model, we classify sentences into one of the evaluation type which is composed of four classes: (1) the reviewer’s view or judgment about the restaurant (positive, negative, mixed opinion); (2) the reviewer’s suggestion, advice and warning to readers, i.e., potential customers and restaurant (suggestion); (3) the reviewer’s intention whether to revisit the restaurant (intention); and (4) the reviewer’s neutral statement about the experience (description). Moreover, previous works assume that positive and negative classes are evenly distributed, whereas in real time application, classes are highly imbalanced (Gopalakrishnan & Ramaswamy, 2014). Similary, in our work, the number of observations per evaluation type were unequal in our work, that is 68% of positive opinions. We chose a dataset of restaurant online reviews written in French. We used, on one hand, resampling and algorithmic approaches to deal with class imbalance problem and on the other hand, supervised machine learning methods to detect and classify evaluative language. We obtained the best macro-average F1-score of 0.79 with SVM classifier and ADASYN resampling method.