[Paper Review] An Adversarial Approach to High-Quality, Sentiment-Controlled Neural Dialogue Generation

Xiang Kong, Bohan Li+3 more|arXiv (Cornell University)|Jan 22, 2019

Topic Modeling32 references20 citations

TL;DR

This paper proposes a conditional generative adversarial network (CGAN) framework for high-quality, sentiment-controlled neural dialogue generation, where a generator produces responses conditioned on dialogue history and sentiment labels, while a discriminator enforces fluency and sentiment accuracy. The method achieves state-of-the-art results in both automatic and human evaluation, significantly improving response quality and sentiment control over baseline models.

ABSTRACT

In this work, we propose a method for neural dialogue response generation that allows not only generating semantically reasonable responses according to the dialogue history, but also explicitly controlling the sentiment of the response via sentiment labels. Our proposed model is based on the paradigm of conditional adversarial learning; the training of a sentiment-controlled dialogue generator is assisted by an adversarial discriminator which assesses the fluency and feasibility of the response generating from the dialogue history and a given sentiment label. Because of the flexibility of our framework, the generator could be a standard sequence-to-sequence (SEQ2SEQ) model or a more complicated one such as a conditional variational autoencoder-based SEQ2SEQ model. Experimental results using automatic and human evaluation both demonstrate that our proposed framework is able to generate both semantically reasonable and sentiment-controlled dialogue responses.

Motivation & Objective

• To develop a neural dialogue generation system that produces high-quality, semantically appropriate responses while explicitly controlling sentiment.
• To address the limitations of existing methods that either lack sentiment control or suffer from data sparsity and high computational costs.
• To leverage conditional adversarial learning to jointly optimize response fluency and sentiment alignment.
• To enable flexible integration with various generator architectures, including standard seq2seq and CVAE-based models.
• To improve response diversity and quality through adversarial training with a policy gradient approach for discrete outputs.

Proposed method

• The framework employs a conditional GAN (CGAN) with a generator and a discriminator, where the generator produces responses conditioned on dialogue history and a sentiment label.
• The discriminator evaluates the authenticity of the triple (dialogue history, sentiment label, response), distinguishing real data from generated samples.
• The generator is trained via policy gradient to maximize the discriminator's confidence in classifying generated responses as real, improving fluency and sentiment alignment.
• The generator can be implemented as a standard sequence-to-sequence (SEQ2SEQ) model or a conditional variational autoencoder (CVAE)-based SEQ2SEQ model.
• The training process uses adversarial loss to encourage the generator to produce responses that are both fluent and consistent with the given sentiment label.
• The framework supports end-to-end training with reinforcement learning, enabling optimization over discrete response sequences.

Experimental results

Research questions

• RQ1 Can adversarial training effectively improve the fluency and sentiment consistency of neural dialogue responses?
• RQ2 Can the proposed CGAN framework enable explicit control over the sentiment of generated responses while maintaining semantic relevance to the dialogue history?
• RQ3 How does the proposed method compare to baseline models in terms of response quality and sentiment accuracy?
• RQ4 Does combining the CGAN framework with a CVAE-based generator further improve response diversity and quality?
• RQ5 Can the adversarial training objective generalize across different dialogue contexts and sentiment labels?

Key findings

• The CGAN-CVAE model achieved the lowest perplexity (69.54) and highest sentiment accuracy (78.8%) among all models in automatic evaluation.
• Human evaluation showed the CGAN-CVAE model achieved the highest quality score (3.9) and sentiment accuracy (78.9%), outperforming all baselines.
• The CGAN-CVAE model generated more diverse and contextually relevant responses, such as 'omg i love it' for positive sentiment and 'that’s the worst' for negative sentiment.
• The CGAN framework significantly improved response quality over standard SEQ2SEQ and CVAE models, with a 20% improvement in sentiment accuracy over the baseline SEQ2SEQ model.
• The discriminator effectively guided the generator to produce responses that were both fluent and aligned with the target sentiment, as confirmed by both automatic and human evaluation.
• The method demonstrated robustness across diverse dialogue contexts, generating appropriate responses even for ambiguous or emotionally charged inputs.