VD-BERT: A Unified Vision and Dialog Transformer with BERT

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3325–3338, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3325 VD-BERT: A Unified Vision and Dialog Transformer with BERT Yue Wang1∗, Shafiq Joty2, Michael R. Lyu1, Irwin King1, Caiming Xiong2, and Steven C.H. Hoi2 1 Department of Computer Science and Engineering The Chinese University of Hong Kong, HKSAR, China 2Salesforce Research 1{yuewang,lyu,king}@cse.cuhk.edu.hk 2{sjoty,cxiong,shoi}@salesforce.com Abstract Visual dialog is a challenging vision-language task, where a dialog agent needs to answer a series of questions through reasoning on the image content and dialog history. Prior work has mostly focused on various attention mech- anisms to model such intricate interactions. By contrast, in this work, we propose VD-BERT, a simple yet effective framework of unified vision-dialog Transformer that leverages the pretrained BERT language models for Visual Dialog tasks. The model is unified in that (1) it captures all the interactions between the image and the multi-turn dialog using a single-stream Transformer encoder, and (2) it supports both answer ranking and answer generation seam- lessly through the same architecture. More crucially, we adapt BERT for the effective fu- sion of vision and dialog contents via visually grounded training. Without the need of pre- training on external vision-language data, our model yields new state of the art, achieving the top position in both single-model and ensem- ble settings (74.54 and 75.35 NDCG scores) on the visual dialog leaderboard. Our code and pretrained models are released at https: //github.com/salesforce/VD-BERT. 1 Introduction Visual Dialog (or VisDial) aims to build an AI agent that can answer a human’s questions about visual content in a natural conversational setting (Das et al., 2017). Unlike the traditional single-turn Visual Question Answering (VQA) (Antol et al., 2015), the agent in VisDial requires to answer ques- tions through multiple rounds of interactions to- gether with visual content understanding. The primary research direction in VisDial has been mostly focusing on developing various atten- tion mechanisms (Bahdanau et al., 2015) for a bet- *This work was mainly done when Yue Wang was an intern at Salesforce Research Asia, Singapore. V Q A (a) Most VQA V H Q A (b) Most VisDial V H Q A (c) Our VD-BERT Figure 1: Attention flow direction illustration. V: vi- sion, H: dialog history, Q: question, A: answer. The ar- row denotes the attention flow direction and the dashed line represents an optional connection. ter fusion of vision and dialog contents. Compared to VQA that predicts an answer based only on the question about the image (Figure 1(a)), VisDial needs to additionally consider the dialog history. Typically, most of previous work (Niu et al., 2019; Gan et al., 2019; Kang et al., 2019) uses the ques- tion as a query to attend to relevant image regions and dialog history, where their interactions are usu- ally exploited to obtain better visual-historical cues for predicting the answer. In other words, the atten- tion flow in these methods is unidirectional – from question to the other components (Figure 1(b)). …

Content truncated to 500 words