Sentiment analysis is getting quite popular recently. Its goal is to determine sentiment of natural language. This popularity can be explained with the fast growth of available information or with the growth of available computational capacity. The state of the art on English text achieves higher than 90 % accuracy regarding sentiments. However, less resource is available for Hungarian text, the highest methods reach up to 80 % accuracy.
The goal of this thesis is to develop a method, which can predict sentiment of sentences accurately. The implementation of the method is mostly done with convolutional, recurrent and fully-connected neural networks. Some of the models use sentiment lexicons, which tells if a specific word is negative or positive. Complex models are also implemented, which combine different approached or different kinds of networks. The models are trained on the OpinHuBank corpus, which classifies each sentence regarding an entity. The best displayed model achieves around 80 % accuracy on the positive-negative labelled task, while the accuracy of the 3-label task is over 60 %.