MELD dataset

MELD, introduced by Poria and colleagues at ACL 2019, is a multimodal emotion corpus built from the television series Friends. It contains roughly 13,000 utterances drawn from about 1,400 dialogues, each utterance carrying one of seven emotion labels — anger, disgust, fear, joy, neutral, sadness and surprise — and a coarser sentiment label of positive, negative or neutral. It extends the earlier text-only EmotionLines dataset by adding the aligned audio and video, so the same utterances can be studied across modalities.

Its defining property is conversational context. Unlike acted corpora where each clip stands alone, MELD utterances sit inside multiparty dialogues with several speakers taking turns, and the emotional reading of a line frequently depends on what came before it. A flat "great" is sarcastic or sincere depending on the preceding turn. This makes MELD a benchmark for context-aware emotion recognition specifically, and models that classify each utterance in isolation leave a substantial amount of performance on the table.

It is also closer to natural speech than the acted corpora. Utterances are short, often a few words, and they contain interruptions, overlapping speech, laughter and background noise. That realism is the point, and it is why reported scores on MELD are markedly lower than on studio corpora such as RAVDESS or CREMA-D — a gap that reflects the difficulty of the material rather than a defect in the models.

The label distribution is heavily skewed. Neutral accounts for a large share of utterances while fear and disgust are rare, so accuracy is a poor summary and results should be read as macro-averaged F1. Papers reporting only accuracy on MELD are usually reporting the neutral class.

The source material carries its own problems. The audio is broadcast television with a laugh track, music cues and post-production compression, all of which are correlated with emotional content in ways that have nothing to do with the speaker — a model can learn to hear the laugh track rather than the delivery. Actors performing to a studio audience are not speaking naturally. And because the cast is fixed and small, speaker leakage across splits is a real risk. The same caution applies to MUStARD, the sarcasm corpus drawn from similar sources, where we report cross-show splits precisely so that show-specific cues cannot be mistaken for a general result.