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

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7038–7051 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7038 Surface Form Competition: Why the Highest Probability Answer Isn’t Always Right Ari Holtzman1* Peter West1,2* Vered Shwartz 1,2 Yejin Choi 1,2 Luke Zettlemoyer1 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Allen Institute for Artificial Intelligence {ahai,pawest}@cs.washington.edu Abstract Large language models have shown promis- ing 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 compe- tition—wherein different surface forms com- pete for probability mass, even if they repre- sent the same underlying concept in a given context, e.g. “computer” and “PC.” Since probability mass is finite, this lowers the prob- ability of the correct answer, due to competi- tion 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 sur- face form competition by simply reweighing each option according to its a priori likeli- hood within the context of a specific task. It achieves consistent gains in zero-shot perfor- mance over both calibrated (Zhao et al., 2021) and uncalibrated scoring functions on all GPT- 2 and GPT-3 models on a variety of multiple choice datasets. * 1 Introduction Despite the impressive results large pretrained lan- guage models have achieved in zero-shot settings (Brown et al., 2020; Radford et al., 2019), we argue that current work underestimates the zero- shot capabilities of these models on classification tasks. This is in large part due to surface form competition—a property of generative models that causes probability to be rationed between different valid strings, even ones that differ trivially, e.g., by capitalization alone. Such competition can be largely removed by scoring choices according to *Code is available at https://github.com/ peterwestuw/surface-form-competition Figure 1: While humans select from given options, lan- guage models implicitly assign probability to every pos- sible string. This creates surface form competition be- tween different strings that represent the same concept. Example from CommonsenseQA (Talmor et al., 2019). Domain Conditional Pointwise Mutual Informa- tion (PMIDC), which reweighs scores by how much more likely a hypothesis (answer) becomes given a premise (question) within the specific task domain. Specifically, consider the example question (shown in Figure 1): “A human wants to sub- merge himself in water, what should he use?” with multiple choice options “Coffee cup”, “Whirlpool bath”, “Cup”, and “Puddle.” From the given op- tions, “Whirlpool bath” is the only one that makes sense. Yet, other answers are valid and easier for a language model to generate, e.g., “Bathtub” and “A bathtub.” Since all surface forms compete for finite 7039 probability mass, allocating significant probability mass to “Bathtub” decreases the amount of prob- ability mass assigned to “Whirlpool bath.” …

Content truncated to 500 words