TR2017-180

Sequence Adversarial Training and Minimum Bayes Risk Decoding for End-to-end Neural Conversation Models


    •  Wang, W., Koji, Y., Harsham, B.A., Hori, T., Hershey, J.R., "Sequence Adversarial Training and Minimum Bayes Risk Decoding for End-to-end Neural Conversation Models", Dialog System Technology Challenges, December 2017.
      BibTeX TR2017-180 PDF
      • @inproceedings{Wang2017dec,
      • author = {Wang, Wen and Koji, Yusuke and Harsham, Bret A. and Hori, Takaaki and Hershey, John R.},
      • title = {Sequence Adversarial Training and Minimum Bayes Risk Decoding for End-to-end Neural Conversation Models},
      • booktitle = {Dialog System Technology Challenges},
      • year = 2017,
      • month = dec,
      • url = {https://www.merl.com/publications/TR2017-180}
      • }
  • MERL Contacts:
  • Research Areas:

    Artificial Intelligence, Speech & Audio

We present a neural conversation system that incorporates multiple sequence-to-sequence models, sequence adversarial training, example-based response selection, and BLEU-based Minimum Bayes Risk (MBR) decoding. The system was trained and tested using the 6th Dialog System Technology Challenges (DSTC6) Twitter help-desk dialog task. Experimental results demonstrate that adversarial training and the example-based method are effective in improving human rating score while system combination with MBR decoding improves objective measures such as BLEU and METEOR scores. Moreover, we investigate extension of the reward function for sequence adversarial training in order to balance subjective and objective scores.

 

  • Related News & Events

    •  EVENT   MERL leads organization of dialog technology challenges and associated workshop
      Date: Sunday, December 10, 2017
      MERL Contacts: Bret Harsham; Chiori Hori; Takaaki Hori
      Location: Hyatt Regency, Long Beach, CA
      Research Area: Speech & Audio
      Brief
      • MERL researcher Chiori Hori led the organization of the 6th edition of the Dialog System Technology Challenges (DSTC6). This year's edition of DSTC is split into three tracks: End-to-End Goal Oriented Dialog Learning, End-to-End Conversation Modeling, and Dialogue Breakdown Detection. A total of 23 teams from all over the world competed in the various tracks, and will meet at the Hyatt Regency in Long Beach, CA, USA on December 10 to present their results at a dedicated workshop colocated with NIPS 2017.

        MERL's Speech and Audio Team and Mitsubishi Electric Corporation jointly submitted a set of systems to the End-to-End Conversation Modeling Track, obtaining the best rank among 19 submissions in terms of objective metrics.
    •