Sentiment analysis (or opinion mining) has been an active research area in Natural Language Processing (NLP). Nevertheless, little attention has been paid in sentiment analysis to other concepts such as suggestion and intention, which can be helpful in decision-making for businesses and individual consumers. We perform sentiment analysis using our proposed classification scheme for evaluative language, consisting of opinion (positive/negative/mixed), suggestion, intention, and description. We use different pre-trained language models developed for French: CamemBERT, FlauBERT, and multilingual BERT model. Their performances are analyzed and compared with the baseline (linear SVM), which integrates the linguistic features observed during the analysis of our corpus. Our results illustrate considerable improvement using BERT-based models, especially for the evaluation categories (i.e., negative/mixed opinion, description) that yield poor performances with the baseline. Nevertheless, we show that the identification of description remains challenging for both humans and automatic detection.