lewiswatson/PAIR_Corpus

PAIR: A Pilot Dataset for Dual Perspective-based Video-Grounded Dialogue and Reconciliation

This repository contains the PAIR dialogue corpus for studying collaborative reconciliation under perceptual asymmetry (dual ego/exo video perspectives).

Links

• GitHub mirror: https://github.com/lewiswatson55/PAIR_Corpus
• Hugging Face dataset mirror: https://huggingface.co/datasets/lewiswatson/PAIR_Corpus/
• Paper (public link): Coming soon (the link will be added here once the paper is public)

Dataset at a glance

• 5 video scenarios (from Ego-Exo4D clips), each collected with 3 independent participant pairs
• 15 dialogues total
• 2,676 utterances in the released transcripts
• 42-label dialogue act schema (Stolcke et al., 2000); 38 labels occur in this release
• Transcripts are manually verified and include speaker labels plus dialogue-act tags

Repository structure

• JSON Version/pair_dataset.json
  • Full corpus in one JSON file
• CSV Version/<clip_id>/Dialogue_01.csv ... Dialogue_03.csv
  • Per-dialogue CSV exports for each clip ID
• Analysis/PAIR_Corpus_TBR_Analysis.csv
  • Turn-balance and dialogue-structure analysis table used in the paper
• PAIR_FIGURE.png
  • Figure used in this README

Data format

JSON format

pair_dataset.json is a dictionary:

• Key: clip/scenario ID (e.g., upenn_0630_Cooking_4_3, iiith_soccer_001_4)
• Value: list of 3 dialogues for that clip
• Each dialogue: list of utterance objects with:
  • Speaker
  • Text
  • Tag

CSV format

Each CSV file has the same schema:

• Speaker
• Text
• Tag

Source videos and redistribution

PAIR is derived from ego/exo clips in Ego-Exo4D (Grauman et al., 2023). This repository does not redistribute the original video stimuli. Only anonymised transcript and annotation data are included here.

Users wishing to access the original video material must obtain it directly from the official Ego-Exo4D release under its respective license.

Citation

Please cite PAIR as:

Watson, L., Strathearn, C., Mitchell, K., & Yu, Y. (2026, May 11-16). PAIR: A Pilot Dataset for Dual Perspective-based Video-Grounded Dialogue and Reconciliation. LREC 2026: Language Resources and Evaluation Conference, Palma, Mallorca. https://lrec2026.info/

@inproceedings{watson2026pair,
  title = {PAIR: A Pilot Dataset for Dual Perspective-based Video-Grounded Dialogue and Reconciliation},
  author = {Watson, Lewis and Strathearn, Carl and Mitchell, Kenny and Yu, Yanchao},
  booktitle = {LREC 2026: Language Resources and Evaluation Conference},
  year = {2026},
  month = {May},
  address = {Palma, Mallorca},
  publisher = {European Language Resources Association (ELRA)},
  url = {https://lrec2026.info/}
}

Citation metadata will be updated when the final publication page and DOI are available.