HDRS-Corpus

Contact: Shrikant Malviya ([email protected]), Rohit Mishra ([email protected])

"HDRS- Hindi Dialogue Restaurant Search", is a fully-labelled collection of human-human written conversations in an Indic language Hindi on the restaurant domain. This is the first attempt to provide a dialogue corpus on the Hindi language.

• This corpus is collected to promote research and development in the field of the Hindi dialogue system.
• Corpus consists of 1400 dialogues collected using the Wizard-of-Oz paradigm, divided into Training (800), Development (200) and Testing (400) parts.
• The corpus uses the details of 118 Indian restaurants in a city named Allahabad.

For more details, follow HDRS: Hindi Dialogue Restaurant Search Corpus for Dialogue State Tracking in Task-Oriented Environment (Special issue DSTC-8, IEEE/ACM Transactions on Audio, Speech, and Language Processing)

A sample dialogue form the corpus :

Comparison of the DST models on Joint-Goal accuracy :

• Category-1 = Static domain ontology based DST,
• Category-2 = BERT-Based DST

References

If you use the HDRS dataset in your work, please cite the below paper:

[Malviya et al. 2021]
@article{malviya2021hdrs,
  author={Malviya, Shrikant and Mishra, Rohit and Barnwal, Santosh Kumar and Tiwary, Uma Shankar},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  title={HDRS: Hindi Dialogue Restaurant Search Corpus for Dialogue State Tracking in Task-Oriented Environment},
  year={2021}
}