MERCI — Multimodal Emotionally-aware Robot Dialogue Corpus

Real-robot open-domain dialogues collected with the PERCY conversational system on the ARI social robot.

Each session includes synchronized video (3 views), audio, transcripts, visual emotion (FER), text sentiment (VADER), and persona metadata. This release adds a turn-level cross-modal audit and two indexing benchmarks from our journal extension of CBMI 2025 MERCI.

	Count
Sessions	41
Dialogue turns (user + robot)	2,432
User turns	1,205 (all with FER + sentiment)
Valence-conflict user turns	151 (12.5% of labelled user turns)
Aligned audio–video	~12.0 hours

Counts match metadata/manifest.json (export v0.2.2, 2026-05-31). Four incomplete source sessions remain excluded (see DATA_STANDARD.md).

Links: Paper scripts · PERCY code · Full usage guide · Schema reference · Benchmark specs

---

Who is this dataset for?

You want to…	Start with…
Explore real robot dialogue quickly	Quick start below
Run the paper benchmarks	benchmarks/README.md
Inspect cross-modal disagreement	data/conflict_stratum.jsonl
Replay video/audio for one session	media/<session_id>/ + data/sessions/<id>/dialogue.json
Reproduce human validation (§4.1)	annotation/section41/

---

Quick start (3 steps)

Step 1 — Download

pip install -U huggingface_hub datasets pandas
hf download zhijin-meng/MERCI --repo-type dataset --local-dir merci_hf
cd merci_hf

Step 2 — Load user turns

from datasets import load_dataset

users = load_dataset("json", data_files="data/turns_user.jsonl", split="train")
print(users[0]["content"])
print(users[0]["emotion_visual"], users[0]["sentiment"])
print("conflict?", users[0]["cross_modal_valence_conflict"])

Step 3 — Get conflict turns only

conflicts = users.filter(lambda x: x["cross_modal_valence_conflict"])
print(len(conflicts), "conflict turns")
print(conflicts[0])

That is enough to start exploring. For session-level media and persona fields, see USAGE.md.

---

Core idea: when face and text disagree

At each user turn, MERCI logs:

• Visual channel: 7-class facial emotion (emotion_visual)
• Text channel: sentiment polarity (sentiment: positive / neutral / negative)

These channels often agree — but not always. Example conflict: happy face + negative text.

Turn class	Meaning	Share (1,205 labelled user turns)
consistent	Face valence and text polarity agree	17.3%
neutral	At least one channel is neutral	68.3%
conflict	Face and text disagree in valence	12.5%
other	Surprise + non-neutral text	1.8%

The field cross_modal_valence_conflict marks the conflict class. The file data/conflict_stratum.jsonl lists all 151 conflict turns for benchmarking.

---

Benchmark tasks (Section 4)

MERCI ships two main downstream tasks aligned with the paper. The checked-in split in benchmarks/README.md is 24 / 3 / 3 sessions (30-session release used in early paper baselines). For this 41-session release, always split by session_id (never by turn) and re-derive train/val/test lists before comparing to published baseline numbers.

Benchmark A — Text-only conflict retrieval

Goal: Given only the user transcript, predict whether the turn is in the valence-conflict stratum.

Input	User-turn text (text only — no video/audio labels at test time)
Label	cross_modal_valence_conflict
Metrics	Macro-F1, conflict-class F1, PR-AUC
Gold file	data/conflict_stratum.jsonl or turns_user.jsonl

Paper baselines (v0.2.1 held-out test: 96 turns, 8 conflict turns — re-run on v0.2.2 recommended):

Model	Macro-F1	Conflict-F1	PR-AUC
Majority baseline	0.478	0.000	0.083
LogReg + TF-IDF	0.523	0.125	0.158
DistilRoBERTa FT	0.478	0.000	0.164

Benchmark B — Conflict-stratified affect indexing

Goal: Predict 7-class visual emotion from text, and report performance on all turns vs conflict turns only.

Input	User-turn text
Label	emotion_visual (7 classes)
Key metric	Gap Δ = full-corpus Macro-F1 − conflict-subset Macro-F1
Why it matters	Shows where text-only indexing breaks under cross-modal disagreement

Paper baselines:

Model	Full Macro-F1	Conflict-subset F1	Gap Δ
Majority baseline	0.130	0.000	0.130
LogReg + TF-IDF	0.109	0.000	0.109
DistilRoBERTa emotion head	0.098	0.000	0.098

Full task definitions, split lists, and reproduction commands: benchmarks/README.md.

Extended probes (Appendix B): label-level conflict prediction (B.1), persona-conditioned generation (B.2), deployment affect classification (B.3).

---

What's in the box?

merci_hf/
├── README.md                     ← you are here
├── USAGE.md                      ← copy-paste recipes
├── DATA_STANDARD.md              ← full field definitions
├── benchmarks/                   ← task specs + baseline results
├── data/
│   ├── turns.jsonl               ← all messages (user + robot)
│   ├── turns_user.jsonl          ← user turns only  ★ start here
│   ├── conflict_stratum.jsonl    ← 151 conflict turns  ★ Benchmark A gold
│   └── sessions/<session_id>/    ← per-session JSON bundle (7 files)
├── metadata/
│   ├── sessions.jsonl            ← one row per session
│   └── manifest.json             ← release counts
├── media/<session_id>/           ← whole_video.mp4, side_video.mp4, audio.wav
└── annotation/section41/         ← human validation pack (§4.1)

Dataset Viewer tabs: use the config dropdown to preview turns, turns_user, sessions, or conflict_stratum. Video/audio files are under Files → media/.

---

Key fields (cheat sheet)

Field	What it is
session_id	Unique session identifier (7-part demographic code)
role	user or assistant
content	Transcript text
emotion_visual	7-class face emotion (user turns only)
sentiment	Text polarity: positive / neutral / negative
cross_modal_valence_conflict	true if face valence ≠ text polarity (Benchmark A label)
cross_modal_turn_class	consistent / neutral / conflict / other
t_start_sec, t_end_sec	Turn timing on session timeline (seconds)
cohort	ari_robot (original CBMI) or benchmark_rerecord

Persona fields live in data/sessions/<id>/profile.json and pre_survey.json.

---

Tips for beginners

1. Always split by session, not by individual turns — otherwise the same participant leaks into train and test.
2. Start with turns_user.jsonl — robot turns have no emotion/sentiment labels.
3. Conflict turns are sparse (~13%) — report conflict-class F1 and PR-AUC, not accuracy alone.
4. Benchmark A is text-only — do not pass emotion_visual or sentiment as model inputs at test time.
5. Need help? See USAGE.md for pandas examples, media clipping, and session replay.

---

Human validation (§4.1)

We release a stratified annotation pack under annotation/section41/:

• 176 turns (88 auto-conflict + 88 matched controls)
• Binary question: do face and text feel contradictory?
• Includes annotation UI, merge scripts, and per-annotator exports

See annotation/section41/ANNOTATION_USAGE.md for the full workflow.

---

Privacy & ethics

• Ethics approval: iRECS4630; all participants gave informed consent
• No email, phone, IP, or raw Qualtrics exports in the public release
• Surveys are redacted; profile.json contains persona Q&A only

---

Citation

@inproceedings{althubyani2025merci,
  title={MERCI: A multimodal dataset for personalised and emotionally-aware dialogues},
  author={Althubyani, Mohammed and Meng, Zhijin and others},
  booktitle={CBMI},
  year={2025}
}

Please also cite the journal extension when using the cross-modal audit strata and benchmarks (Meng et al., MTAP extension of CBMI 2025).

---

Changelog

Version	What changed
v0.2.2	41 active sessions (+106/111/112, +120–123/130–132); media/ and text/metadata aligned; counts in metadata/manifest.json
v0.2.1	30 sessions (+105/107/108/109/110); benchmark package; redesigned dataset card
v0.2	Unified turns.jsonl schema; conflict stratum; dual-camera media
v0.1	25 sessions, text-only release (CBMI 2025)