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

By Ari Holtzman, Peter West, Vered Schwartz, Yejin Choi, Luke S. Zettlemoyer. 2021

301 citations recorded by the source corpus.

Large language models have shown promising results in zero-shot settings. 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 in a given context, 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 its a priori likelihood within the context of a specific task. It achieves consistent gains in zero-shot performance over both calibrated and uncalibrated scoring functions on all GPT-2 and GPT-3 models on a variety of multiple choice datasets.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Finetuned Language Models Are Zero-Shot Learners (2021, background): “The GPT-3 paper does not show results on sentiment analysis, but Holtzman et al. (2021) find that zero-shot performance on SST-2 also gets worse from 13B to 175B parameters.”
Emergent Abilities of Large Language Models (2022, positive): “For instance, simple modiﬁcations such as calibrating output probabilities (Zhao et al., 2021; Holtzman et al., 2021) or using a noisy channel (Min et al., 2021) have improved performance on a range of tasks.”
Can language models learn from explanations in context? (2022, methodology): “We explored correcting for the prior probability of answers (Holtzman et al., 2021), but did not observe a benefit in a preliminary experiment, and so used direct likelihood comparisons for the main experiments.”
Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? (2022, methodology): “Prior work proposes better ways of formulating the problem (Zhao et al., 2021; Holtzman et al., 2021; Min et al., 2021a), better ways of choosing labeled examples for the demonstrations (Liu et al., 2021; Lu et al., 2021; Rubin et al., 2021), meta-training with an explicit in-context learning…”
Reframing Human-AI Collaboration for Generating Free-Text Explanations (2021, background): “Inspired by work in other generation tasks (Holtzman et al., 2020; Massarelli et al., 2020; Holtzman et al., 2021), we hypothesize that equally or more informative explanations can be generated by sampling stochastically.”
Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing (2021, background): “This is because that all surface forms of a same object will compete for inite probability mass [45].”
MetaICL: Learning to Learn In Context (2021, background): “…outperforms baselines including (1) a variety of LM in-context learning baselines without meta-training (Brown et al., 2020; Zhao et al., 2021; Holtzman et al., 2021; Min et al., 2022), and (2) multi-task learning followed by zero-shot transfer (Zhong et al., 2021; Wei et al., 2022; Sanh et…”
Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections (2021, background): “However, since these models are not adapted to answer those prompts (Holtzman et al., 2021), we suspect that its knowledge and true potential to perform few-shot learning is much higher than reported.”
An Explanation of In-context Learning as Implicit Bayesian Inference (2021, positive): “Zhao et al. (2021); Holtzman et al. (2021) improve in-context accuracy via calibration or renormalization, a form of adaptation to the prompt distribution.”
Learning To Retrieve Prompts for In-Context Learning (2021, methodology): “…prompt engineering through manual patterns (Petroni et al., 2019; Schick and Schütze, 2021), decoding meth-ods (Min et al., 2021; Zhao et al., 2021; Holtzman et al., 2021), and methods for extracting either hard (Shin et al., 2020; Haviv et al., 2021) or soft (Li and Liang, 2021; Zhong et al.,…”
Yuan 1.0: Large-Scale Pre-trained Language Model in Zero-Shot and Few-Shot Learning (2021, background): “Based on previous work on calibration[26, 27], the model’s bias can be ﬁxed with an empty text.”
Noisy Channel Language Model Prompting for Few-Shot Text Classification (2021, methodology): “Holtzman et al. (2021) explores zero-shot models that compute the probability of x given y based on Pointwise Mutual Information, but with a restriction that the input and the output are in-terchangeable.”
ERNIE 3.0 Titan: Exploring Larger-scale Knowledge Enhanced Pre-training for Language Understanding and Generation (2021, background): “In practice, it is better to restrain the P ( y ) in the target domain by prompting the text using a domain-speciﬁc prompt but taking as input an empty input which is used in [101].”
SimPO: Simple Preference Optimization with a Reference-Free Reward (2024, unknown): “To address this issue, we consider using the average log-likelihood as the implicit reward: This metric is commonly used for ranking options in beam search [33, 49] and multiple-choice tasks within language models [12, 37, 58].”
Demystifying Prompts in Language Models via Perplexity Estimation (2022, background): “For the word prediction tasks we only report scores, since accuracy in general is less stable, suffers more from the surface form competition (Holtzman et al., 2021), and is usually quite low for these tasks in our setting (the chances the model will generate an exact match of the label are low).”
Selective Annotation Makes Language Models Better Few-Shot Learners (2022, methodology): “…methods to improve in-context learning in many aspects: e.g., meta-training (Chen et al., 2022; Min et al., 2022b), task instructions (Efrat & Levy, 2020; Mishra et al., 2022; Wei et al., 2021; Sanh et al., 2022), or task formulation (Holtzman et al., 2021; Zhao et al., 2021; Min et al., 2022a).”
Automatic Chain of Thought Prompting in Large Language Models (2022, background): “…fine-grained information, such as incorporating task instruction [Mishra et al., 2022, Wei et al., 2022b, Sanh et al., 2022]; (iii) manipulating output probabilities of LLMs instead of directly computing the likelihood of target labels [Holtzman et al., 2021, Zhao et al., 2021, Min et al., 2022a].”
On the Effect of Pretraining Corpora on In-context Learning by a Large-scale Language Model (2022, background): “A few follow-up studies have tried to improve the in-context learning ability (Zhao et al., 2021; Holtzman et al., 2021).”
Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations (2022, background): “This is in part due to the broader limitations of generative LMs, which assign high probability to unlikely sentences (Welleck et al., 2020; Holtzman et al., 2021) and are sensitive to semantic perturbations (Elazar et al., 2021).”