MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in\n Conversations

Abstract

Emotion recognition in conversations is a challenging task that has recently\ngained popularity due to its potential applications. Until now, however, a\nlarge-scale multimodal multi-party emotional conversational database containing\nmore than two speakers per dialogue was missing. Thus, we propose the\nMultimodal EmotionLines Dataset (MELD), an extension and enhancement of\nEmotionLines. MELD contains about 13,000 utterances from 1,433 dialogues from\nthe TV-series Friends. Each utterance is annotated with emotion and sentiment\nlabels, and encompasses audio, visual and textual modalities. We propose\nseveral strong multimodal baselines and show the importance of contextual and\nmultimodal information for emotion recognition in conversations. The full\ndataset is available for use at http:// affective-meld.github.io.\n