Abstract Unsupervised word representations have demonstrated improvements in predictive generalization on various NLP tasks. Much effort has been devoted to effectively learning word embeddings, but little attention has been given to distributed character representations, although such character-level representations could be very useful for a variety of NLP applications in intrinsically “character-based” languages (e.g. Chinese and Japanese). On the other hand, most of existing models create a singleprototype representation per word, which is problematic because many words are in fact polysemous, and a single-prototype model is incapable of capturing phenomena of homonymy and polysemy. We present a neural network architecture to jointly learn character embeddings and induce context representations from large data sets. The explicitly produced context representations are further used to learn context-specific and multipleprototype character embeddings, particularly capturing their polysemous variants. Our character embeddings were evaluated on three NLP tasks of character similarity, word segmentation and named entity recognition, and the experimental results demonstrated the proposed method outperformed other competing ones on all the three tasks.