Overprotective Training Environments Fall Short at Testing Time: Let Models Contribute to Their Own Training
Not on the map. Why?
Abstract (arXiv)
Despite important progress, conversational systems often generate dialogues that sound unnatural to humans. We conjecture that the reason lies in their different training and testing conditions: agents are trained in a controlled "lab" setting but tested in the "wild". During training, they learn to generate an utterance given the human dialogue history. On the other hand, during testing, they must interact with each other, and hence deal with noisy data. We propose to fill this gap by training the model with mixed batches containing both samples of human and machine-generated dialogues. We assess the validity of the proposed method on GuessWhat?!, a visual referential game.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 12 of its 16 references that are in the map. First posted on arXiv 20 Mar 2021.
Builds on
12 of 16 references.
References · 12
- 2015Deep Residual Learning for Image Recognitioncited by 51,809
- 2014Microsoft COCO: Common Objects in Contextcited by 17,498
- 2014Sequence to Sequence Learning with Neural Networkscited by 2,974
- 2019The Curious Case of Neural Text Degenerationcited by 2,881
- 2016Visual Dialogcited by 377
- 2017Deal or No Deal? End-to-End Learning for Negotiation Dialoguescited by 202
- 2015A Neural Conversational Modelcited by 195
- 2016GuessWhat?! Visual object discovery through multi-modal dialoguecited by 112
- 2017Learning Cooperative Visual Dialog Agents with Deep Reinforcement Learningcited by 70
- 2017Image-Grounded Conversations: Multimodal Context for Natural Question and Response Generationcited by 59
- 2019Improving Generative Visual Dialog by Answering Diverse Questionscited by 21
- 2018Beyond task success: A closer look at jointly learning to see, ask, and GuessWhatcited by 15
Built on by
Not cited yet.