Improving Generative Visual Dialog by Answering Diverse Questions

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1449–1454, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1449 Improving Generative Visual Dialog by Answering Diverse Questions Vishvak Murahari1 Prithvijit Chattopadhyay1 Dhruv Batra1,2 Devi Parikh1,2 Abhishek Das1 1Georgia Tech 2Facebook AI Research Abstract Prior work on training generative Visual Dia- log models with reinforcement learning (Das et al., 2017b) has explored a Q-BOT-A- BOT image-guessing game and shown that this ‘self-talk’ approach can lead to im- proved performance at the downstream dialog- conditioned image-guessing task. However, this improvement saturates and starts degrad- ing after a few rounds of interaction, and does not lead to a better Visual Dialog model. We find that this is due in part to repeated interac- tions between Q-BOT and A-BOT during self- talk, which are not informative with respect to the image. To improve this, we devise a sim- ple auxiliary objective that incentivizes Q-BOT to ask diverse questions, thus reducing repeti- tions and in turn enabling A-BOT to explore a larger state space during RL i.e. be exposed to more visual concepts to talk about, and varied questions to answer. We evaluate our approach via a host of automatic metrics and human studies, and demonstrate that it leads to bet- ter dialog, i.e. dialog that is more diverse (i.e. less repetitive), consistent (i.e. has fewer con- flicting exchanges), fluent (i.e. more human- like), and detailed, while still being compara- bly image-relevant as prior work and ablations. 1 Introduction Our goal is to build agents that can see and talk i.e. agents that can perceive the visual world and com- municate this understanding in natural language conversations in English. To this end, Das et al. (2017a); de Vries et al. (2017) proposed the task of Visual Dialog – given an image, dialog history consisting of a sequence of question-answer pairs, and a follow-up question about the image, predict a free-form natural language answer to the ques- tion – along with a dataset and evaluation metrics1. Posing Visual Dialog as a supervised learning problem is unnatural. This is because at every 1visualdialog.org, guesswhat.ai round of dialog, the agent’s answer prediction is thrown away and it gets access to ground-truth di- alog history, thus disabling it from steering conver- sations during training. This leads to compound- ing errors over long-range sequences at test time, a problem also common in training recurrent neu- ral networks for language modeling (Bengio et al., 2015; Ross et al., 2011; Ranzato et al., 2016). To overcome this, Das et al. (2017b) devised a goal-driven approach for training Visual Dialog agents with deep reinforcement learning. This is formulated as a game between two agents – Q- BOT (that asks questions) and A-BOT (that an- swers questions). Q-BOT is shown a one-line de- scription of an unseen image that A-BOT has ac- cess to and Q-BOT is allowed to ask questions (in natural language) to A-BOT for …

Content truncated to 500 words