Surface Form Competition: Why the Highest Probability Answer Isn't Always Right

TL;DR

This paper identifies the problem of surface form competition in large language models' multiple choice tasks and proposes a new scoring method that improves zero-shot performance by reweighing answer options.

Contribution

It introduces Domain Conditional Pointwise Mutual Information, a novel scoring function that mitigates surface form competition in zero-shot multiple choice tasks.

Findings

Consistent performance improvements across GPT-2 and GPT-3 models.

Effective in various multiple choice datasets.

Outperforms calibrated and uncalibrated scoring methods.

Abstract

Large language models have shown promising results in zero-shot settings (Brown et al.,2020; Radford et al., 2019). For example, they can perform multiple choice tasks simply by conditioning on a question and selecting the answer with the highest probability. However, ranking by string probability can be problematic due to surface form competition-wherein different surface forms compete for probability mass, even if they represent the same underlying concept, e.g. "computer" and "PC." Since probability mass is finite, this lowers the probability of the correct answer, due to competition from other strings that are valid answers (but not one of the multiple choice options). We introduce Domain Conditional Pointwise Mutual Information, an alternative scoring function that directly compensates for surface form competition by simply reweighing each option according to a term that is proportional to its a priori likelihood within the context of the specific zero-shot task. It achieves consistent gains in zero-shot performance over both calibrated (Zhao et al., 2021) and uncalibrated scoring functions on all GPT-2 and GPT-3 models over a variety of multiple choice datasets.