minsik-oh/template-dialogue-pairs

Template Dialogue Pairs

English utterance-template pairs for dialogue representation learning, combining original and synthetically augmented examples from five datasets. This is the training corpus used by Template Dialogue Embedding.

Load

from datasets import load_dataset

pairs = load_dataset(
    "minsik-oh/template-dialogue-pairs",
    "combined",
    split="train",
)
print(pairs[0])

Use config snips, atis, massive, hwu or clinc for one source. Only the train split is included. Pin a Hub revision for reproducibility.

Contents

The corpus combines SNIPS (all seven intents), ATIS, MASSIVE (en-US), HWU and CLINC, including original and synthetically augmented examples.

Columns

Column	Meaning
utterance	Original or augmented utterance used for training
pattern	Corresponding template, with generic {SLOT} markers
slot_count	Number of {SLOT} markers in the pattern
source	snips, atis, massive, hwu or clinc
source_row	Zero-based position within that source's training CSV

Intent labels and original-versus-synthetic flags are not included. source_row preserves record identity when text pairs repeat. The combined configuration concatenates SNIPS, ATIS, MASSIVE, HWU and CLINC in that order.

Preparation and verification

Templates use supplied slot annotations for SNIPS, ATIS, MASSIVE and HWU, and cached spaCy NER annotations for CLINC. The release retains all seven SNIPS intents. Packaging preserves the training text, encoding, embedded newlines, duplicates and order without further generation or normalization. Existing source overlap is retained; this is not a newly decontaminated dataset.

Every Parquet row was checked against the training CSVs. The complete dataset was also loaded from Hugging Face and checked for matching content and order. provenance.json records generation settings, input hashes and per-intent counts; verification.json records packaging checks. These files preserve detailed provenance beyond this release overview.

Combined training CSV SHA256: 9b8404042073814afbf0a63670e8ef9919d722cf8bee1b7aea208ecdad696dd4.

Sources and attribution

The source datasets retain their respective licenses and attribution requirements; this package does not assert a new blanket license. SNIPS attribution: Coucke, Alice, et al. (2018), Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces, arXiv:1805.10190.

Paper: Template-assisted Contrastive Learning of Task-oriented Dialogue Sentence Embeddings

Source code: https://github.com/minsik-ai/Template-Contrastive-Embedding

Citation

If you use this dataset in your research, please cite our ACL 2026 Oral paper:

@inproceedings{oh-etal-2026-template,
    title = "Template-assisted Contrastive Learning of Task-oriented Dialogue Sentence Embeddings",
    author = "Oh, Minsik  and
      Li, Jiwei  and
      Wang, Guoyin",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1015/",
    doi = "10.18653/v1/2026.acl-long.1015",
    pages = "22181--22198",
    ISBN = "979-8-89176-390-6"
}