This thesis develops document-level sentiment analysis for Norwegian. First we present the Norwegian Review Corpus (NoReC) -- the first publicly available sentiment dataset for Norwegian, consisting of over 35,000 reviews rated on a six-point scale across multiple domains including films, music, restaurants and products. In addition to describing each step of the dataset creation, we perform extensive data exploration and analysis. Using ratings as a proxy to the overall sentiment of a review, we run a large number of rating inference experiments, first using traditional machine learning methods, then using convolutional neural networks and pre-trained word embeddings. We analyze the performance of the models with regards to ratings, categories, language standard and training set size, and perform a thorough hyperparameter search on the convolutional neural network architecture using Bayesian hyperparameter optimization. We demonstrate that our convolution architecture outperforms the traditional machine learning methods, and show that task-specific tuning can be necessary in order to train high performing models.