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

ACL ARR 2025 February Submission6892 Authors

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 a novel dynamic control strength method that considers the uncertainty of the model’s generation and classification probabilities. Specifically, we dynamically adjust the control strength at each generation step based on the entropy of the language model's next token probabilities and the entropy of the attribute classifier's probability estimates. Experimental results on various existing models demonstrate that our decoding method achieves high control performance while maintaining fluency compared to existing decoding strategies across all models. Additionally, our approach alleviates the probability interpolation issue in multi-attribute controlled generation, yielding superior performance.

Paper Type: Long

Research Area: Generation

Research Area Keywords: text-to-text generation, inference methods, interactive and collaborative generation

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 6892