Learning Easily Updated General Purpose Text Representations with Adaptable Task-Specific Prefixes

TL;DR

This paper introduces a prefix-based method for learning fixed, general text representations that adapt to multiple tasks efficiently, reducing computational costs compared to traditional multi-task fine-tuning.

Contribution

The proposed prefix-based approach improves the generalizability of fixed text representations and is more computationally efficient than multi-task training.

Findings

Prefix-based training outperforms multi-tasking in generalization.

The method requires less computation for updates.

It effectively adapts to unseen downstream tasks.

Abstract

Many real-world applications require making multiple predictions from the same text. Fine-tuning a large pre-trained language model for each downstream task causes computational burdens in the inference time due to several times of forward passes. To amortize the computational cost, freezing the language model and building lightweight models for downstream tasks based on fixed text representations are common solutions. Accordingly, how to learn fixed but general text representations that can generalize well to unseen downstream tasks becomes a challenge. Previous works have shown that the generalizability of representations can be improved by fine-tuning the pre-trained language model with some source tasks in a multi-tasking way. In this work, we propose a prefix-based method to learn the fixed text representations with source tasks. We learn a task-specific prefix for each source task independently and combine them to get the final representations. Our experimental results show that prefix-based training performs better than multi-tasking training and can update the text representations at a smaller computational cost than multi-tasking training.