Abstract
In this work, we tackled the task of identifying sentiment bearing sentences for product reviews in Norwegian. We have created a set of automatically labeled datasets that classify sentences in terms of how relevant they are to the reviews' overall sentiment and also in terms of their sentiment polarity. We leveraged authors' annotations in the form of positive and negative keyphrases, called pros and cons, to provide distant supervision. Then, we used the created datasets to train a sentence identification system using both feed-forward and convolutional neural network models, and pre-trained word embeddings. We also performed a detailed hyperparameter search for our convolutional architecture. The performance of the models was analyzed with regards to product categories and a thorough manual error analysis was performed on the system's output. Our results demonstrate the usefulness of pros and cons to capture the overall sentiment of a review and our convolutional model outperformed all baselines. Our analysis illustrates how task-specific hyperparameter tuning is beneficial for training high performing models for sentence classification.