Are SentiWordNet scores suited for multi-domain sentiment classification?

Motivated by the numerous applications of analysing opinions in multi-domain scenarios, this paper studies the potential of a still rarely considered approach to the problem of multi-domain sentiment analysis based on Senti-WordNet as lexical resource. SentiWordNet scores are exploited together with additional features to assign a polarity to a text using machine learning. On the other hand, a rule-based approach is studied based on sentiment scores. The introduced methods are tested on single domains of a real-world data set consisting of documents in six different domains, but also in cross-domain settings. The results show that for cross-domain sentiment analysis rule-based approaches with fix opinion lexica are unsuited. For machine-learning based sentiment classification a mixture of documents of different domains achieves good results.