Source-linked AI summary
Are You Talking to Me? Reasoned Visual Dialog Generation through Adversarial Learning
Qi Wu, Peng Wang, Chunhua Shen, Ian Reid, Anton van den Hengel
TL;DR
Visual Dialogue asks agents to answer image questions in a conversation where earlier turns may shape later answers. This paper combines sequential co-attention generation with an adversarial discriminator that judges answers using attention-based reasons, and reports state-of-the-art VisDial performance alongside more human-like responses.
Problem
Visual Dialogue requires consistent, natural answers to image questions that may depend on previous dialogue turns.
Source: §1 · p. 1Method
A sequential co-attention generator and adversarial discriminator jointly use image, question, dialogue history, and attention-based reasons to train more human-like responses.
Source: §1 · p. 2Results
Takeaways & Limitations
For visual-dialogue interfaces, human-likeness is a significant response-quality index alongside correctness.
Source: §5 · p. 8
Abstract
from arXiv · showhide
The Visual Dialogue task requires an agent to engage in a conversation about an image with a human. It represents an extension of the Visual Question Answering task in that the agent needs to answer a question about an image, but it needs to do so in light of the previous dialogue that has taken place. The key challenge in Visual Dialogue is thus maintaining a consistent, and natural dialogue while continuing to answer questions correctly. We present a novel approach that combines Reinforcement Learning and Generative Adversarial Networks (GANs) to generate more human-like responses to questions. The GAN helps overcome the relative paucity of training data, and the tendency of the typical MLE-based approach to generate overly terse answers. Critically, the GAN is tightly integrated into the attention mechanism that generates human-interpretable reasons for each answer. This means that the discriminative model of the GAN has the task of assessing whether a candidate answer is generated by a human or not, given the provided reason. This is significant because it drives the generative model to produce high quality answers that are well supported by the associated reasoning. The method also generates the state-of-the-art results on the primary benchmark.
1. Introduction
Visual Dialogue extends image question answering into a history-dependent conversation, where agents must stay consistent while answering successive questions. The paper combines co-attention generation with adversarial reinforcement to encourage human-like, reason-supported replies.
1. Introduction
- Visual Dialogue requires successive answers that may depend on earlier questions and answers, making consistent, meaningful conversation central to the task.The paper presents this as a challenge for agents that collaborate with people through dialogue.Source: §1 · p. 1
- Human-like answers are longer and include useful unasked-for reasoning, whereas word-level MLE tends to favor generic, repetitive replies.Simple answers such as “yes,” “no,” or “I don’t know” can score highly under MLE.Source: §1 · p. 2
- The approach jointly trains a response generator and discriminator with reinforcement learning, using the discriminator’s human-versus-machine judgment as a reward.The generator conditions on image content and dialogue history and aims to produce responses that fool the discriminator.Source: §1 · p. 2
- The discriminator evaluates answers using the generator’s attention weights, which provide interpretable clues about the image regions and dialogue turns informing each response.It can assess the answer in the context of the dialogue so far.Source: §1 · p. 2
- Monte Carlo search estimates intermediate rewards because response quality is assessed over the whole generated sequence.The policy gradient is updated using REINFORCE.Source: §1 · p. 2
- The generative model exceeds its strong MLE baseline by 1.87% on recall@5 and previous best reported results by 2.14% on recall@5 and 2.50% on recall@10.Qualitative evaluation reports more informative responses, and 49% of responses pass the Turing Test.Source: §1 · p. 2
2. Related work
Prior work treats visual dialogue as a vision-language task and explores text dialogue generation and adversarial learning. This paper distinguishes its approach by integrating visual context and attention-based reasoning into both adversarial components.
2. Related work
- VisDial pairs people to chat about images and formulates the task as multi-round VQA evaluated through retrieval or multiple-choice responses.Other cited visual-dialogue datasets include Image Grounded Conversation [20] and GuessWhat.Source: §2 · p. 2
- Unlike Lu et al. [18], whose discriminator ranks a provided list of candidate responses, this work evaluates answers using attention that reflects reasoning and dialogue history.The authors report that this procedure results in more suitable generated responses.Source: §2 · pp. 2–3
- Text dialogue research includes sequence-to-sequence generation and reinforcement-learning systems trained with rewards such as informativity, coherence, and ease of answering.The cited work also explores combining reinforcement learning with neural generation.Source: §2 · p. 3
- Earlier dialogue GANs distinguish human from machine responses using text, while this work includes visual content and attention-based reasoning in both adversarial components.The discriminator assesses whether an answer is human-generated given its attention-based reason.Source: §2 · p. 3, Figure 2 · p. 3
- Building on adversarial sequence-generation methods, this work redesigns the generator and discriminator for visual content and dialogue history and uses intermediate rewards during generation.The authors cite policy-gradient sequence generation and prior applications to image captioning and dialogue.Source: §2 · pp. 3–4
3. Adversarial Learning for Visual Dialog Generation
The section presents a generative answerer that conditions on the image, question, and dialogue history, and trains it adversarially using a discriminator that scores visual-dialogue fit.
3. Adversarial Learning for Visual Dialog Generation
- The authors choose a generative decoder to avoid fixed-vocabulary response limits and use discriminator rewards to encourage answers that fit the visual dialogue.The discriminator's reward is used to update the generator with REINFORCE.Source: §3 · p. 4, §3.3 · p. 5
3.1. A sequential co-attention generative model
- Sequential co-attention passes guidance among image regions, question features, and dialogue utterances before combining the attended features for response generation.The question first guides image attention; image and question features then guide history attention, and attended history and image features guide question attention.Source: §3.1 · pp. 4–5
3.2. A discriminative model with attention memories
- The discriminator classifies whether a visual dialogue is human-generated using attended image and history features, the question, and the generated answer.The question-answer pair is encoded, then the features are combined for a two-way classification.Source: §3.2 · p. 5
3.3. Adversarial REINFORCE with an intermediate reward
- Monte Carlo rollouts complete each response prefix, and the average discriminator score becomes the reward for the next token's generation action.The method uses these intermediate rewards to address vanilla REINFORCE's sequence-level reward shared across all tokens.Source: §3.3 · p. 6
- Training alternates discriminator-reward policy updates with teacher-forcing updates on real dialogue using a maximum-likelihood objective.The discriminator uses human answers as positive examples and generated responses as negative examples.Source: §3.3 · pp. 5–6
4. Experiments
Experiments on VisDial v0.9 evaluate response retrieval and human judgments. The final generative model leads the reported generative metrics, while ablations, a discriminative setting, and a human study examine its components and outputs.
4. Experiments
- VisDial v0.9 contains image-grounded conversations of up to 10 rounds; the experiments use 80k dialogs for training, 3k for validation, and 40k for testing.The dialogs pair AMT questioners, who cannot see the image, with answerers who can.Source: §4 · p. 6
4.1. Evaluation Metrics
- At test time, the model ranks 100 candidate answers using generative log-likelihood, evaluated by human-response rank, recall@k, and mean reciprocal rank.Recall@k measures whether the human response appears among the top k ranked candidates.Source: §4.1 · p. 6
4.2. Implementation Details
- The model uses 512-dimensional learned word embeddings shared across question, history, and decoder LSTMs, with five Monte Carlo searches per token.The co-attention generator and discriminator are each pretrained for 30 epochs.Source: §4.2 · p. 7
4.3. Experiment results
- Compared methods include answer-frequency ranking, nearest-neighbor retrieval, late fusion, hierarchical recurrent encoders, memory networks, and HCIAE.The baselines encode image, question, and dialogue history in different combinations.Source: §4.3 · p. 7
- CoAtt-GAN-w/ Rinte-TF performs best on all generative evaluation metrics and beats MN by 3.81% on R@1.Table 1 reports generative-method performance on VisDial v0.9.Source: §4.3 · p. 7, Table 1 · p. 7
- Intermediate rewards improve the baseline by an average 1%, and teacher forcing adds a further average 0.5%, achieving the best results.Adversarial learning with global reward alone yields only a slight improvement over the baseline.Source: §4.3 · p. 8
- In the discriminative setting, the model outperforms prior baseline and state-of-the-art models on all evaluation metrics.The response-generation LSTM is replaced with a single-step classifier over a predefined answer set.Source: §4.3 · p. 8, Table 2 · p. 8
4.4. Human study
- In a study of 1,000 sampled responses, three raters judged whether each final response was human-generated, with agreement by at least two counting as a pass.The study also found that 45% of responses were rated better than or equal to human responses.Source: §4.4 · p. 8, Table 3 · p. 8
5. Conclusion
The proposed adversarial approach combines a sequential co-attention generative model with a discriminator that accesses attention memories. The final model achieves state-of-the-art results on VisDial, and a Turing Test study finds its responses more human-like.
5. Conclusion
- The paper frames human-like dialogue as significant for human-computer interaction in robotics and AI, beyond response correctness alone.Source: §5 · p. 8
- The final model achieves state-of-the-art results on VisDial, and a Turing Test study finds its responses more human-like.Source: §5 · p. 9
- The approach combines a sequential co-attention generative model that jointly reasons over the image, dialogue history, and question with a discriminator that dynamically accesses attention memories.Source: §5 · pp. 8–9