ECO Decoding: Entropy-Based Control for Controllability and Fluency in Controllable Dialogue Generation

TL;DR

Controllable Dialogue Generation often trades fluency for controllability due to fixed control strengths in weighted decoding. ECO decoding introduces entropy-based, per-step dynamic control that jointly considers language model entropy and attribute entropy to adjust weighting between $P(r_i|r_{<i},h)$ and attribute signals, yielding $P(r_i|r_{<i},h,c) \propto P_{lm}(r_i|r_{<i},h)^{\alpha_{lm,i}} \times \prod_{c_j\in C} P_{c_j}(c_j|r_{\le i},h)^{\lambda \alpha_{c_j,i}}$, with $\alpha$ values derived from entropies. Across DailyDialog and MultiWOZ, using DialoGPT and Llama2-7B, ECO consistently improves controllability (attribute accuracy) while preserving grammar, fluency, and diversity, outperforming FUDGE, Director, and DASC in single- and multi-attribute scenarios and mitigating interpolation issues in multi-attribute generation. The approach is training-free, scalable, and generalizes across domains, with minimal inference overhead, suggesting practical impact for robust, attribute-aligned dialogue systems. The work also provides POS-based entropy insights and analyzes entropy smoothing, reinforcing the rationale for entropy-guided control in generative language models.

Abstract

Controllable Dialogue Generation (CDG) enables chatbots to generate responses with desired attributes, and weighted decoding methods have achieved significant success in the CDG task. However, using a fixed constant value to manage the bias of attribute probabilities makes it challenging to find an ideal control strength that satisfies both controllability and fluency. To address this issue, we propose ECO decoding (Entropy-based COntrol), which dynamically adjusts the control strength at each generation step according to the model's entropy in both the language model and attribute classifier probability distributions. Experiments on the DailyDialog and MultiWOZ datasets demonstrate that ECO decoding consistently improves controllability while maintaining fluency and grammaticality, outperforming prior decoding methods across various models and settings. Furthermore, ECO decoding alleviates probability interpolation issues in multi-attribute generation and consequently demonstrates strong performance in both single and multi-attribute scenarios.

Figures (6)

• Figure 1: An example of a controllable dialogue generation method based on dynamic weighting with ECO decoding. ECO decoding dynamically adjusts the weights between the language model probability distribution and the attribute control probability distribution, enabling attribute control while maintaining fluency.
• Figure 2: An illustration of controllable dialogue generation using the weighted decoding method, incorporating ECO decoding.
• Figure 3: The single attribute control performance of the existing weighted decoding method (red) and ECO decoding (green) with respect to changes in the control strength $\lambda$. The y-axis represents grammar, and the x-axis represents accuracy. The blue dot line represents uncontrolled DialoGPT's grammar score.
• Figure 4: The multi attribute control performance of the existing weighted decoding method (red) and ECO decoding (green) with respect to changes in the control strength $\lambda$. The y-axis represents grammar, and the x-axis represents accuracy. The blue dot line represents uncontrolled DialoGPT's grammar score.
• Figure 5: In case where the response fails to satisfy the desired attribute with the existing method but satisfies the desired attribute using ECO decoding.