YouTube
Morency et al. [7] presented a tri-modal approach for classifying
sentiment in 30 second excerpts from YouTube videos.
Over 10,000 new videos are posted to YouTube every day.
The selected videos consisted of users addressing the camera
and giving their opinions on a particular product or topic,
and were classied as either `positive', `negative' or `neutral'
at the le level. Their work used linguistic, visual, and
acoustic features combined in early fusion and were classied
using a HMM. Good performance was demonstrated when
all three data modes were in use. However, their manually
annotated dataset comprised of only 47 videos