Skip to content. | Skip to navigation

Personal tools


You are here: Home / Demos / Emotion-based analysis and recommendation of lectures

Emotion-based analysis and recommendation of lectures


We demonstrate the automatic emotion analysis of 1000 TED talks over 12 dimensions, as well as the recommendation of similar talks to a given one based on emotions. The input to our algorithms is the manual transcripts and the community emotion labels which are provided in TED API.

To view the demo, please use the following link: Emotion-based analysis and recommendation of lectures

Our model is based on multiple-instance learning applied to the prediction of ratings from texts such as lecture transcripts (as in this demo), comments on lectures, or product reviews. Each data point (text or transcript) is described by several independent feature vectors (one word vector per sentence). To learn from texts or transcripts of hyper-events with known aspect ratings (such as, here, the 12 emotion dimensions), the model performs multiple-instance regression and assigns importance weights to each of the sentences, uncovering their contribution to the ratings per aspect. Then, the model is used to predict aspect ratings in previously unseen texts or transcripts, demonstrating interpretability and explanatory power for its predictions.

For comparison we have two other options for recommendation presented previously (see here), namely text-based and audio-based.