This is my thesis work at Linnéuniversity in Kalmar. The research problem was to investigate if deeplearning and a convolutional neural network could be used to analyse sentiment in Swedish reviews. The result was very good on just 2 categories (positive and negative) with 95% accuracy but only 80% accuracy as best on 3 categories (positive, negative and neutral).

Link to thesis: Sentiment Analysis With Convolutional Neural Networks: Classifying sentiment in Swedish reviews

The two webscrapers scrapes Swedish reviews from www.reco.se and se.trustpilor.com to create training datasets for the neural networks. You can either scrape for 3 categories (positive, negative and neutral) or just for 2 categories (positive and negative).

In the CNN-model folder is the thesis works research problem. There is a playground file for testing TFlearn and its functionality and a working class to import and use for training a convolutional neural network and to predict swedish reviews. There are also some existing CNN-models in the folder that could be loaded and used.