RSA-Control: A Pragmatics-Grounded Lightweight Controllable Text Generation Framework

Yifan Wang

Saarland University

The NLP Reading Group is excited to host Yifan Wang, from Saarland University, who will be discussing his recent work on controllable text generation using the Rational Speech Acts (RSA) framework. The talk will happen on Zoom in A14 on Friday April 4th at 1PM.

Talk Description

Controlling pre-trained language models to follow specific attributes during generation without additional training remains a challenge, largely due to the lack of modeling how generated text is perceived by listeners. To address this, we introduce RSA-Control, a training-free controllable text generation framework grounded in pragmatics. RSA-Control directs the generation process by recursively reasoning between imaginary speakers and listeners, enhancing the likelihood that target attributes are correctly interpreted by listeners amidst distractors. Additionally, we introduce a self-adjustable rationality parameter, which allows for automatic adjustment of control strength based on context. Our experiments, conducted on both open-ended generation and input-output tasks, across foundation models and instruction-tuned models, show that RSA-Control achieves strong attribute control while maintaining language fluency and content consistency.

Speaker Bio

Yifan Wang (he/him) is a PhD student at Saarland University, co-supervised by Prof. Dr. Vera Demberg and Prof. Dr. Isabel Valera. His research focuses on identifying and mitigating undesirable behaviors in NLP systems, including toxicity, stereotypes, and social biases. He is also passionate about making deep neural networks more interpretable and transparent.

Logistics

Date: April 4th
Time: 1PM
Location: A14 or Zoom (See email)