Sentiment Analysis of Finnish Customer Reviews

This paper presents a hybrid algorithm that takes advantage of polarity lexicons along with state-of-the-art word embeddings with weighted average of reviews and headlines for predicting polarity of reviews. The algorithm gives a significantly better coverage compared to polarity lexicons and a better classification accuracy in comparison to the word embedding models. We implemented Finnish polarity lexicons with different construction techniques, and evaluated the quality of the lexicons with translated polarity lexicons. The polarity lexicon construction and the polarity classification involves a new dataset, which was extracted from online user reviews in Finnish. In general, using headlines instead of reviews as sentiment data source provides better results at the expense of coverage.