Sentiment analysis (SA) is the task of detecting positive and negative opinions expressed in text and is one of the applications of Natural Language Processing that has found the most widespread use. The expression of sentiment interacts in important ways with various contextual and compositional phenomena in language. One of the most important such phenomena is that of negation.
Consider the following example sentence where the polar expression `recommended' (anbefales), with its corresponding target expression `the food' (maten), is modified by the negation cue `in no way' (på ingen måte), with the result that the polarity value is flipped from positive to negative.
The food can in no way be recommended.
Maten kan på ingen måte anbefales.
This project will explore on the impact of negation in sentiment analysis, and how it can be best handled in modeling.
Focusing on sentiment analysis for Norwegian text, the SANT project has created annotated data for both fine-grained sentiment analysis (NoReCfine) and negation resolution (NoReCneg). These data sets will form the starting point of the project. The texts themselves are arts- and consumer-reviews gathered from online news sources, as compiled in the Norwegian Review Corpus (NoReC).
Examples of relevant research questions include: To what degree do annotations of negation and sentiment overlap? Among cases of overlap, what is the proporttion of cases where negation actually affects the polarity? Among the incorrect predictions made by SA models, what is the proportion related to negation? How can we best incorporate information about negation in a SA model? This latter point could include both a literature survey and practical experiments that compare different approaches (e.g., joint prediction vs. multi-task learning vs. a pipeline-approach with negation features, etc.).
The project requires a balance of technical and linguistic expertise. Good programming skills, experience with machine learning and a solid background in NLP are relevant qualifications.