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

Construction and provenance

MR-Bench-H is built in three stages, as described in the paper (Section 2.2 and supplementary Section B): candidate mining, reader-panel screening and blind human validation. The released test set contains 900 clips: 620 LS and 280 NS; control clips are excluded from the release.

Stage LS NS Control Total
Candidate pool 1,474 380 186 2,040
Reader-panel pass 798 363 176 1,337
Final retained, human-validated test set 620 280 0 900
LS/NS items outside the final test set 178 83 0 261
Panel-passing controls tracked separately 0 0 176 176

Of 2,040 candidates from validation and test splits and unseen corpora, 1,337 pass the panel and 900 the annotators (620 LS, 280 NS).

Stage 1: candidate mining

Candidates are mined from unimodal judgments on the MMSA sources, explicit versus implicit emotion annotations on MUStARD++, and the text-only view of a trained fusion model on MELD and IEMOCAP. LS candidates have misleading linguistic evidence; NS candidates have misleading non-verbal evidence. Reference and distractor polarities are binary and opposite.

Stage 2: reader-panel screening

Four instruction-tuned text LLMs from three families (Qwen2.5-7B, Qwen3-8B, Yi-1.5-9B and Phi-3.5-mini) answer the binary sentiment question from the transcript alone. An LS candidate passes when at least three readers answer the distractor polarity. NS candidates and controls pass when at least three readers answer the reference polarity. This leaves 1,337 panel-passing items.

Stage 3: blind human validation

Three annotators independently assess each shuffled clip with its labels hidden. They answer Q2 from the transcript alone, Q5 from the muted video, then Q1 from the full clip; they also identify the decisive channel (Q3) and ambiguity (Q4).

  • LS retention: the majority gives Q1 = reference polarity, Q2 = distractor polarity, Q3 in {voice, face, mixed}, and Q4 = no ambiguity.
  • NS retention: the majority gives Q1 = reference polarity, Q5 = distractor polarity, Q3 in {words, voice, mixed}, and Q4 = no ambiguity.

These majority-vote criteria retain 620 LS and 280 NS clips. Control accuracy is used to flag careless annotators. The annotation workflow is described in ANNOTATION_GUIDE.md.

Agreement

Annotator agreement is Fleiss' κ = 0.75. Every retained LS and NS clip meets the criteria above, so majority-vote accuracy on the released clips is fixed by the rule and is not reported as a result. 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.

Every released row has validated: true and validation.n_annotators: 3. Individual annotator votes are not distributed; the item-level fields validation.agreement and validation.evidence_channel are null.

Per-source counts

Candidates after stage 1, items passing the reader panel (stage 2) and items retained after human validation (stage 3), as LS / NS / controls (Table S4 of the paper).

Source Language Candidates Reader panel Validated
CMU-MOSI EN 83 / 135 / 22 23 / 130 / 21 14 / 119 / 0
CMU-MOSEI EN 139 / 177 / 32 53 / 175 / 28 23 / 119 / 0
CH-SIMS ZH 48 / 26 / 7 38 / 22 / 7 24 / 16 / 0
CH-SIMSv2 ZH 96 / 42 / 14 74 / 36 / 14 49 / 26 / 0
MUStARD++ EN 101 / -- / 10 69 / -- / 10 53 / -- / 0
MELD EN 407 / -- / 41 176 / -- / 39 176 / -- / 0
IEMOCAP EN 600 / -- / 60 365 / -- / 57 281 / -- / 0
Total 1,474 / 380 / 186 798 / 363 / 176 620 / 280 / 0

Reproducibility

release_manifest.json records the SHA-256 hashes of the distributed files and the UID-set hash. JSONL and Parquet contain the same 900 records. Annotation returns, restricted IEMOCAP fields and original media are outside the public release.