Enabling Natural Zero-Shot Prompting on Encoder Models via Statement-Tuning

TL;DR

This paper introduces Statement-Tuning, a method that enables smaller encoder models to perform zero-shot and few-shot tasks by training them to discriminate between finite statements, achieving competitive results with fewer parameters.

Contribution

It presents Statement-Tuning, a novel approach that allows encoder models to generalize to zero-shot and few-shot tasks by modeling discriminative tasks as statement sets.

Findings

Statement-Tuning achieves competitive performance with fewer parameters.

Task and statement diversity improve zero-shot generalization.

Strong performance with modest training data.

Abstract

While Large Language Models (LLMs) exhibit remarkable capabilities in zero-shot and few-shot scenarios, they often require computationally prohibitive sizes. Conversely, smaller Masked Language Models (MLMs) like BERT and RoBERTa achieve state-of-the-art results through fine-tuning but struggle with extending to few-shot and zero-shot settings due to their architectural constraints. Hence, we propose Statement-Tuning, a technique that models discriminative tasks as a set of finite statements and trains an encoder model to discriminate between the potential statements to determine the label. We do Statement-Tuning on multiple tasks to enable cross-task generalization. Experimental results demonstrate that Statement-Tuning achieves competitive performance compared to state-of-the-art LLMs with significantly fewer parameters. Moreover, the study investigates the impact of several design choices on few-shot and zero-shot generalization, revealing that Statement-Tuning can achieve strong performance with modest training data and benefits from task and statement diversity for unseen task generalizability.