Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders

Citation

• Authors: Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi
• Year: 2017
• Venue: ACL 2017
• arXiv: arXiv:1703.10960
• URL: https://aclanthology.org/P17-1061/

One-Line Takeaway

This paper is one of the clearest bridges from classical CVAEs to text generation: it treats dialogue response generation as a conditional one-to-many problem, uses a latent variable for discourse-level variation, and adds practical losses to keep the latent from being ignored.

Technical Core

• Let $c$ denote dialogue context, $x$ the target response, and $z$ a latent discourse variable. The core model is $$p_\theta(x, z \mid c) = p_\theta(x \mid z, c)\, p_\theta(z \mid c).$$
• Training introduces a recognition model $$q_\phi(z \mid x, c)$$ and optimizes the conditional ELBO $$\mathcal{L}_{\mathrm{CVAE}} = \mathbb{E}_{q_\phi(z \mid x,c)}[\log p_\theta(x \mid z,c)] - \mathrm{KL}\!\left(q_\phi(z \mid x,c)\,\|\,p_\theta(z \mid c)\right).$$
• In the knowledge-guided version, the model also predicts a linguistic label $y$ such as a dialogue act, adding an auxiliary term $$\mathbb{E}_{q_\phi(z \mid x,c,y)}[\log p_\theta(y \mid z,c)].$$
• To prevent latent underuse, the paper augments training with a bag-of-words reconstruction term $$\mathbb{E}_{q_\phi(z \mid x,c,y)}[\log p_\theta(x_{\mathrm{bow}} \mid z,c)],$$ which forces $z$ to carry response-level content not reducible to local autoregressive token prediction.

Method

• Encode dialogue history with utterance encoders plus a recurrent context encoder to obtain $c$.
• Parameterize both the prior $p_\theta(z \mid c)$ and posterior $q_\phi(z \mid x,c)$ as Gaussian networks.
• Decode the response with a GRU conditioned on both $c$ and sampled $z$.
• In kgCVAE, inject explicit linguistic features such as dialogue acts and train a network to predict them from $(z,c)$.
• Use KL annealing and the bag-of-words loss to reduce the vanishing-latent problem that otherwise appears with the recurrent decoder.

Empirical Evaluation

• Experiments are on the Switchboard dialogue corpus with multiple-reference evaluation assembled to reflect the one-to-many nature of dialogue.
• The main empirical claim is that CVAE and especially kgCVAE increase discourse-level response diversity while keeping outputs relevant and coherent.
• The bag-of-words loss is also evaluated separately as a training device for keeping the latent informative.

Assumptions And Limits

The learned latent is still an unconstrained continuous discourse variable, not an identifiable or task-grounded strategy label. The setting is dialogue generation rather than general language-model factorization, and the paper depends on practical heuristics such as KL annealing and auxiliary losses rather than a clean theorem for latent semantics.

Relevance To This Project

• Similarities:
  • The paper uses the same conditional latent decomposition that underlies the repo's CVAE methodology: $$p(z \mid x), \qquad p(y \mid x, z), \qquad q(z \mid x, y).$$
  • It explicitly treats conditional text generation as a one-to-many mapping and uses $z$ to capture high-level variation beyond what the condition already determines.
  • It directly grapples with the latent-underuse problem in a sequence decoder, which is central to the repo's methodology.
• Differences:
  • The latent here is mainly a discourse-intent or response-style variable, not a latent reasoning strategy attached to problem solving.
  • The model is trained as a dialogue generator from data, whereas the repo often starts from an already-strong entangled generator and asks how to factorize or post-train it.
  • The paper's main anti-collapse tools are KL annealing, bag-of-words prediction, and optional dialogue-act supervision; the repo studies a broader family of ELBO-adjacent objectives aimed at making latent modes semantically sharper and more reusable.

Comparison Notes

Sohn et al. (2015) is the clean conditional-VAE predecessor in structured prediction, while this paper transports the same idea into autoregressive text generation. Bowman et al. (2016) is the unconditional sentence-VAE ancestor that makes the sequence-collapse problem vivid. Fu et al. (2019) later studies a cleaner scheduling fix for the same latent-vanishing issue in text VAEs.

Related Notes

• Learning Structured Output Representation using Deep Conditional Generative Models
• Generating Sentences from a Continuous Space
• CVAE post-training methodology