Predicting the sense of a discourse relation is particularly challenging when connective markers are missing. To address this challenge, we propose a simple deep neural network approach that replaces manual feature extraction by introducing event vectors as an alternative representation, which can be pre-trained using a very large corpus, without explicit annotation. We model discourse arguments as a combination of word and event vectors. Event information is aggregated with word vectors and a Multi-Layer Neural Network is used to classify discourse senses. This work was submitted as part of the CoNLL 2016 shared task on Discourse Parsing. We obtain competitive results, reaching an accuracy of 38%, 34% and 34% for the development, test and blind test datasets, competitive with the best performing system on CoNLL 2015.