MELD raw videos and annotations

This is a mirror of the locally available MELD 1.0 raw release. MELD (Multimodal EmotionLines Dataset) contains conversation clips from the television series Friends, with one of seven emotion labels and a three-way sentiment label per annotated utterance. The creators publish the official dataset and source repository. Please cite Poria et al., ACL 2019 when using MELD.

Files

• train/train_sent_emo.csv and train/train_splits/*.mp4
• dev_sent_emo.csv and dev_splits_complete/*.mp4
• test_sent_emo.csv and output_repeated_splits_test/*.mp4
• README.txt: documentation bundled with the raw release

The CSV files have 9,989 train, 1,109 dev, and 2,610 test rows (13,708 total). This copy has 13,848 MP4 files. Every annotated train and test clip is present. The annotated dev clip dia110_utt7.mp4 is absent from the source release; there are also 4 dev and 137 test MP4 files without matching annotation rows. The authors' issue #22 documents this source-data discrepancy. Consumers should join CSV rows to video filenames rather than assume every MP4 has a label.

The upstream Hugging Face repository marks MELD as GPL-3.0. The source footage comes from Friends; the upstream repository does not state that the television footage itself is GPL-licensed. Consult the creators' release and applicable rights before further redistribution or reuse.