Controlling the Focus of Pretrained Language Generation Models

TL;DR

This paper introduces a method to control the focus of pretrained language models by using trainable focus vectors, enabling targeted content generation based on user-selected highlights.

Contribution

It proposes a novel mechanism with trainable focus vectors applied to fixed pretrained models, allowing explicit control over the model's attention to specific input spans.

Findings

Focus vectors effectively steer model output towards user-selected highlights.

The approach improves relevance in dialogue response generation.

The method enhances controllability without retraining the entire model.

Abstract

The finetuning of pretrained transformer-based language generation models are typically conducted in an end-to-end manner, where the model learns to attend to relevant parts of the input by itself. However, there does not exist a mechanism to directly control the model's focus. This work aims to develop a control mechanism by which a user can select spans of context as "highlights" for the model to focus on, and generate relevant output. To achieve this goal, we augment a pretrained model with trainable "focus vectors" that are directly applied to the model's embeddings, while the model itself is kept fixed. These vectors, trained on automatic annotations derived from attribution methods, act as indicators for context importance. We test our approach on two core generation tasks: dialogue response generation and abstractive summarization. We also collect evaluation data where the highlight-generation pairs are annotated by humans. Our experiments show that the trained focus vectors are effective in steering the model to generate outputs that are relevant to user-selected highlights.