This paper presents a system for facial animation retargeting that al- lows learning a high-quality mapping between motion capture data and arbitrary target characters. We address one of the main chal- lenges of existing example-based retargeting methods, the need for a large number of accurate training examples to define the corre- spondence between source and target expression spaces. We show that this number can be significantly reduced by leveraging the in- formation contained in unlabeled data, i.e. facial expressions in the source or target space without corresponding poses. In contrast to labeled samples that require time-consuming and error-prone manual character posing, unlabeled samples are easily obtained as frames of motion capture recordings or existing animations of the target character. Our system exploits this information by learning a shared latent space between motion capture and character param- eters in a semi-supervised manner. We show that this approach is resilient to noisy input and missing data and significantly improves retargeting accuracy. To demonstrate its applicability, we integrate our algorithm in a performance-driven facial animation system.