[Paper Review] Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders

Tiancheng Zhao, Ran Zhao+1 more|arXiv (Cornell University)|Mar 31, 2017

Topic ModelingComputer Science30 references165 citations

TL;DR

The paper introduces CVAE-based dialog models to capture discourse-level diversity in open-domain conversations, including a knowledge-guided variant (kgCVAE) and a bag-of-words loss to stabilize training.

ABSTRACT

While recent neural encoder-decoder models have shown great promise in modeling open-domain conversations, they often generate dull and generic responses. Unlike past work that has focused on diversifying the output of the decoder at word-level to alleviate this problem, we present a novel framework based on conditional variational autoencoders that captures the discourse-level diversity in the encoder. Our model uses latent variables to learn a distribution over potential conversational intents and generates diverse responses using only greedy decoders. We have further developed a novel variant that is integrated with linguistic prior knowledge for better performance. Finally, the training procedure is improved by introducing a bag-of-word loss. Our proposed models have been validated to generate significantly more diverse responses than baseline approaches and exhibit competence in discourse-level decision-making.

Motivation & Objective

• Motivate the one-to-many nature of open-domain dialogue and the need for discourse-level diversity.
• Develop a CVAE-based framework to model latent discourse factors for diverse responses.
• Incorporate linguistic knowledge via kgCVAE to improve performance and interpretability.
• Improve CVAE training with a bag-of-words loss to mitigate vanishing latent variables.
• Demonstrate that discourse-level diversity can outperform word-level diversity in generating appropriate responses.

Proposed method

• Represent each dialog as context c, response x, and latent z with p(x|z,c)p(z|c).
• Use a prior network p(z|c) and a recognition network q(z|x,c) to optimize a variational lower bound.
• Encode utterances with BRNN-GRU and context with a GRU to form c, then generate x with a GRU conditioned on z and c.
• Introduce kgCVAE by incorporating linguistic features y into generation and prediction of y' from z,c.
• Mitigate vanishing latent variables with a bag-of-word (BOW) auxiliary loss that predicts the BOW of x from z and c.

Experimental results

Research questions

• RQ1 Can a conditional VAE capture discourse-level diversity in open-domain dialogue?
• RQ2 Does incorporating latent discourse factors improve diversity and coverage of valid responses compared to baselines?
• RQ3 Does knowledge guidance via linguistic features (kgCVAE) improve performance and interpretability?
• RQ4 Does bag-of-words loss stabilize CVAE training and enhance latent usage?

Key findings

• CVAE and kgCVAE generate more diverse responses than a strong baseline, with higher recall across metrics.
• kgCVAE achieves the highest precision and recall across BLEU-1 to BLEU-4 and A-BOW metrics.
• Across entropy contexts, CVAE/kgCVAE show higher recall than baseline, with kgCVAE maintaining higher precision.
• Bag-of-words loss effectively mitigates vanishing latent variables and improves training stability without relying solely on KL annealing.
• t-SNE visualization indicates the learned z-space clusters in ways correlated with dialog acts and response length.