MELD

Why we included this project

MELD is one of the reference datasets for emotion recognition in conversation: a large share of published work in the area is trained and evaluated against it. It takes the EmotionLines corpus and expands it into more than 1,400 dialogues and roughly 13,000 utterances from the Friends TV series, each tagged with one of seven emotions, and it adds audio and visual tracks alongside the text so models can be judged across modalities at once. Teams validating a sentiment or emotion model against a widely cited public benchmark will find it a convenient starting point, and researchers new to the field can use the README's collection of related papers and updated baselines as a map of the area. Keep in mind this is a dataset, not a deployable emotion detection service; the downloading and preprocessing are yours to do, and you set up your own evaluation runs.