MR-Bench-H / docs /ANNOTATION_GUIDE.md
Bluexxx's picture
MR-Bench-H v2.0.1
e259ac7 verified
|
Raw History Blame Contribute Delete
3.18 kB

Blind human validation (stage 3)

Three annotators judge each clip from the transcript, the muted video and the full clip, in this order. They receive shuffled clips without labels (about 50 s per clip); some rows are ordinary control clips.

Guide given to the annotators

You receive annotation/<dataset>_to_annotate.csv. Each row is one short clip (clip_path, relative to the dataset's raw-video folder, e.g. SIMS_Raw/Raw/video_0024/0022.mp4) and its transcript. Labels are hidden and rows are shuffled; some rows are ordinary control clips. Annotate every row independently and do not discuss items with other annotators. Budget: about 50 s per clip.

Source-specific notes

  • IEMOCAP clips show both actors of the dyad. Judge the speaker only: the clip file name ends with _F<nnn> for the female actor or _M<nnn> for the male actor (e.g. Ses01F_impro05_M035.mp4 = the male actor speaks).
  • MELD clips are from the TV series Friends; judge the character who speaks the transcript line.
  • MUStARD++ clips are the final (target) utterance of a sarcasm scene; ignore the preceding context.

Order of work for each row (important — do not change the order)

  1. Transcript only. Read the transcript without playing the clip. Fill Q2_transcript_only: neg / neu / pos — the sentiment a reader would infer from the words alone.
  2. Muted video. Play the clip with the sound OFF. Fill Q5_muted_video: neg / neu / pos — the sentiment you read from the face, gaze and gestures alone.
  3. Full clip. Play the clip with sound. Fill Q1_full_clip: neg / neu / pos — the speaker's overall sentiment in this clip.
  4. Q3_decisive_channel: which channel mainly determines your Q1 answer: words (wording), voice (tone, pitch, laughter, sighs), face (expression, gaze, gestures), mixed.
  5. Q4_ambiguous: yes if you are unsure about Q1, the clip is too short or noisy, the speaker is not visible, or several sentiments are expressed; otherwise no.
  6. comment (optional): e.g. "sarcasm", "laughs while saying something negative", "audio missing".

Examples

  • "你还恭喜我嘛" ("You're still congratulating me?") said with a frown and a cold voice: Q2 pos, Q5 neg, Q1 neg, Q3 voice or face, Q4 no.
  • "I actually loved it" said with a grimace that turns into a laugh: Q2 pos, Q5 neg, Q1 pos, Q3 voice, Q4 no.
  • A neutral statement read flatly with no visible reaction: Q2 neu, Q5 neu, Q1 neu, Q3 words, Q4 no.

Retention

With three annotators per clip, retention is decided by majority vote:

  • LS: Q1 = label, Q2 = misleading polarity, Q3 in {voice, face, mixed}, and Q4 = no.
  • NS: Q1 = label, Q5 = misleading polarity, Q3 in {words, voice, mixed}, and Q4 = no.

Every retained LS and NS clip meets these criteria, so majority-vote accuracy on the released clips is fixed by the rule and is not reported as a result. Annotator agreement is Fleiss' κ = 0.75. Control clips are excluded from the release and are not selected by these criteria; the majority answer matches the label on 91.0% of them, which flags careless annotators.