In a conversation, feedback is mostly performed through short utterances produced by another participant than the main current speaker. These utterances are among the most frequent in conversational data. They are also considered as crucial communicative tools for achieving coordination in dialogue. They have been the topic of various descriptive studies and often given a central role in applications such as dialogue systems. The present project addresses this issue from a linguistic viewpoint and combines fine-grained corpus analyses of semi-controlled data with formal and statistical modeling. At the formal level, the dynamic turn of semantics has laid the ground for rich formal approaches of discourse in which the semantic/pragmatic interface can play its role. Such models allow for a rich yet precise characterization of dialogue communicative functions. The impoverished aspect of the linguistic material in these utterances allows for a truly multi-dimensional analysis that can unveil how different linguistic domains (morpho-syntax, prosody, visual channel) combine to convey meaning and achieve communicative goals. Statistical model will go in hand with the formal model giving it more empirical support and allowing it to focus on truly representative and reliable phenomena. The statistical model will be adapted to create a classification system determining the main communicative functions of a feedback item based on its properties on the different observable dimensions and contextual information. Human-systems interfaces as well systems processing audio and video data have reached interesting achievements in the domain of feedback but is waiting for more linguistic models to continue to improve. A better understanding of the form/function relation of feedback can help extracting a wide range of elements from conversational data such as decisions taken, information accepted, opinion changes or humor-involving sequences.