Natural Language Generation for Dialogue Systems

SIGDIAL 2017 Special Session

Natural Language Generation for Dialogue Systems

While natural language generation (NLG) for dialogue interaction has been a long-standing research topic, there has been a recent explosion of work due to industrial interest in conversational assistants such as Alexa, Siri, Cortana and Google Assistant. Recent advances in the field of deep learning together with the availability of large corpora of human dialogues from social media has also contributed to new research techniques for language generation. However, to date, neural methods have not been able to replicate much of the rich dialogue phenomena targeted by previous rule-based, statistical and data-driven approaches for language generation for dialogue.

Therefore we created this special session at SIGDIAL 2017.

Accepted Papers

We have accepted the following papers for presentation:

  1. "Redundancy Localization for the Conversationalization of Unstructured Responses", by Sebastian Krause, Mikhail Kozhevnikov, Eric Malmi, & Daniele Pighin
  2. "Neural Language Generation in Dialogue using RNN Encoder-Decoder with Semantic Aggregation", by Van-Khanh Tran & Le-Minh Nguyen
  3. "A surprisingly effective out-of-the-box char2char model on the E2E NLG Challenge dataset", by Shubham Agarwal & Marc Dymetman
  4. "The E2E Dataset: New Challenges For End-to-End Generation", Jekaterina Novikova, Ondřej Dušek, & Verena Rieser

After some brief opening remarks, the session will begin with 20 minute talks on papers (1) and (2) and 5 minute talks about papers (3) and (4) before the main event—a panel discussion with:

  • Verena Rieser (Heriot-Watt)
  • Hadar Shemtov (Google Research, Generation for Assistant)
  • Amanda Stent (Bloomberg)
  • Milica Gasic (Cambridge)

Questions for the Panel Discussion

  1. Does a dialogue system need to “know what it is talking about”, that is have the meaning of what it is saying grounded in some semantic representation? or are surface approaches such as Eliza-type chat or neural translation methods going to cover many cases of what is needed? What representations does an NLG for dialogue need to generate coherent utterances in context?
  2. How do you see recent neural generation work in relation to “traditional” approaches to NLG for dialogue? Where do you think the challenges/ limitations/ advantages lie?
  3. How do you evaluate your NLG systems? Do we have the right evaluation metrics?
  4. Data-driven approaches need in-domain data. What datasets are you using? Are they public? How challenging is it to create the data or get the data you need? Are data-driven approaches scalable/ applicable to industry?
  5. As we move beyond Siri / Google Assistant / Alexa style personal assistants, is there more room for traditional concept-to-speech NLG to contribute to dialogue systems, as the length and variety of the generated utterances grow?


Marilyn Walker <>


Marilyn Walker, University of California Santa Cruz (Chair)

Verena Rieser, Heriot-Watt University

Vera Demberg, Universität des Saarlandes

Dietrich Klakow, Universität des Saarlandes

Dilek Hakkani-Tur, Google Research

David M. Howcroft, Universität des Saarlandes

Shereen Oraby, University of California Santa Cruz