This project aims to study the effects and optimization techniques of cross-domain sentiment analysis. While this NLP task is largely deemed as 'resolved' in academic studies, it isn't in the wild.
Since the most commom technique for sentiment analysis is training and testing in the same domain, two problems emerge: 1) you depend on having an annotated dataset from the same domain as the target, which is not true for most real-world cases and 2) there is no clear and unique definition of "domain" so its not possible to define for sure whether the target-domain is of the same domain as the source-domain you trained yourt model on.
For further information on methodology and results, please refeer to the paper and poster.