The aim of this study is to investigate conversational feedback that contains smiles and laughter. Firstly, we propose a statistical analysis of smiles and laughter used as generic and specific feedback in a corpus of French talk-in-interaction. Our results show that smiles of low intensity are preferentially used to produce generic feedback while high intensity smiles and laughter are preferentially used to produce specific feedback. Secondly, based on a machine learning approach, we propose a two-stage classification of feedback to automatically predict not only the presence/absence of a smile but, also the type of smile according to an intensity-scale (low or high).