Grounding Conversations with Improvised Dialogues (ACL2020)

Paper » Dataset » Demo »

SPOLIN: a collection of improvised "Yes, and" dialogue turns

Effective dialogue involves grounding, the process of collecting mutual knowledge and assumptions that is essential for communication between people. Modern dialogue systems are not explicitly trained to build common ground, and therefore overlook an important aspect of communication. Improvisational theater (improv) intrinsically contains a high proportion of dialogue focused on building common ground, and make use of the "Yes, and" principle, a strong grounding speech act, to establish coherence and an actionable objective reality.

We collect a corpus of more than 68,000 "Yes, and" turns, called Selected Pairs of Learnable ImprovisatioN (SPOLIN). We transcribe them from improv dialogues and extracting them from larger, but more sparsely populated movie script dialogue corpora and subtitles corpora, via a bootstrapped classifier. We fine-tune chit-chat dialogue systems with our corpus to encourage more grounded, relevant conversation and confirm these findings with human evaluations.

What is a "Yes, and..."?

Yes-and examples

The yes-and principle is a rule-of-thumb of improvisational theatre that suggests that a participant should accept the reality of what the other participant has said (“Yes”) and expand or refine that reality with additional information(“and”). It does not require the response to explicitly contain the phrase "Yes, and".

Both explicit (top) and implicit (bottom) examples of "Yes, and" responses in the SPOLIN corpus are shown in the figure. The text highlighted in light blue reflects acceptance of the context established in the prompt ("yes") and the text highlighted in orange initiates a new relevant contribution to the dialogue ("and"). Note that in the implicit example that acceptance of the reality and adding more relevant information can be achieved in many different ways, which makes the task of determining "Yes, and" dialogue turns and generating them a non-trivial task.

Demo

Chat with SpolinBot, a DialoGPT model fine-tuned with the latest SPOLIN corpus to generate "Yes, and" responses. Each response from the bot contains five options to choose from, which are ordered based on maximum mutual information (MMI) scoring. For the full demo, visit spolin.isi.edu. Not sure how to start? Check out some examples.

Examples

Here are some fun example dialogues with SpolinBot that people have submitted. You can also submit yours at the full demo!

Paper

If our work inspires you, please cite us:
@inproceedings{cho2020spolin,
    title={Grounding Conversations with Improvised Dialogues},
    author={Cho, Hyundong and May, Jonathan},
    booktitle ={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
    year={2020}
}                    

Disclaimer

This material is based on research sponsored by the AFRL and DARPA under agreement number FA8650-18-C-7878. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the AFRL, DARPA, or the U.S. Government.

Contact

If you have any questions about the dataset, feel free to email us or raise an issue on the Github page.