On the Investigation of Evolutionary Multi-Objective Optimization for Discrete Prompt Search

Anonymous

Abstract: Discrete prompt search (DPS) aims to automatically find high-performing prompts that yield top accuracy in interactions with a pretrained language model. In the context of few-shot learning, evaluations of candidate prompts can only be done via a limited number of labelled examples. The search is often formulated as an optimization problem where prediction accuracy, F1 score, or cross-entropy loss is used as the objective function. While resulting prompts achieve top performance, they are mostly unreadable and uninterpretable, i.e., unlike natural languages. In this paper, we formulate DPS as a true multi-objective optimization (MOO) problem considering simultaneously both prompt performance and readability as separate objectives. We show that there exist certain degrees of conflict between the objectives, making the search for human-readable and highly-accurate prompts a challenging problem. We then propose the Multi-objective Evolutionary Algorithm for Predictive Probability guided Prompting (MoEAP3) to address the problem. Our MoEAP3 returns not a single final prompt as in conventional methods but a whole front of multiple candidate prompts, each representing an efficient trade-off between the objectives. Decision makers can straightforwardly investigate this front and intuitively select the prompt that yields the desired trade-off. Experimental results exhibit the superiority of MoEAP3 over state-of-the-art baselines.

Paper Type: long

Research Area: Efficient/Low-Resource Methods for NLP

Contribution Types: Model analysis & interpretability

Languages Studied: English