Lexicon-based sentiment analysis by mapping conveyed sentiment to intended sentiment

As consumers nowadays generate increasingly more content describing their experiences with, e.g., products and brands in various languages, information systems monitoring a universal, language-independent measure of people's intended sentiment are crucial for today's businesses. In order to facilitate sentiment analysis of user-generated content, we propose to map sentiment conveyed by unstructured natural language text to universal star ratings, capturing intended sentiment. For these mappings, we consider a monotonically increasing step function, a naive Bayes method, and a support vector machine. We demonstrate that the way in which natural language reveals intended sentiment differs across our datasets of Dutch and English texts. Additionally, the results of our experiments on modelling the relation between conveyed sentiment and intended sentiment suggest that language-specific sentiment scores can separate universal classes of intended sentiment from one another to a limited extent.