Emotion annotation in texts remains a challenging task in the field of Natural Language Processing (NLP), as, unlike voice or images, texts might not only contain peculiar cues to express emotions. Methods for emotion annotation are based on lexicons or on machine learning techniques which are based on the use of manually annotated corpora. This paper aims to explore if and how the combination of these two types of methods might be useful for the annotation of emotions in texts. Four data sets are used for comparison of the two approaches, and then to investigate to what extent the results are distinct or complementary on three aspects: (i) identification of emotional sentences; (ii) identification of emotion categories; (iii) identification of one specific mode of expression of emotions called "behavioral emotions" (e.g. shout, cry). Findings show that not all emotions are equally easy to annotate, and, most specifically, the learning-based approach tends to over detect Admiration.