An uncertain approach to sentiment prediction in text-based context
In this thesis, we adopt an uncertainty approach to sentiment prediction (Sentiment includes both opinion and emotion). We perform sentiment prediction in two steps: (1) Sentiment analysis for evoking the extractable sentiments and (2) prediction of the non-extracted sentiments. In the first step, we analyze the users’ written text to extract their expressed sentiments (or sentiment of whom they write about) on different entities/topics and insert the information into a dataset. But, unlike the existing methods in state of the art, we do not represent users’ sentiments as numbers, but as a probability/possibility distribution function (PDF).
Regarding the sufficiency of distribution parameters for representing a PDF, we use only 1 or 2 numbers to represent each sentiment PDF. For the second phase, utilizing collaborative filtering on the created dataset, we predict the non-extracted sentiments. It is notable that, meanwhile we also utilize some simpler models as preliminary phases of proving the main contributions.
Last published articles