Domain Conditional Pointwise Mutual Information (PMI_DC) Scoring

A zero-shot and few-shot multiple-choice scoring method for generative language models that compensates for surface form competition—the rationing of finite probability mass across distinct strings representing the same concept—by selecting the candidate hypothesis y_i that maximizes Domain Conditional Pointwise Mutual Information: arg max_i PMI_DC(x, y_i, domain) = arg max_i log [P(y_i | x) / P(y_i | x_domain)], where P(y_i | x_domain) is the a priori probability of hypothesis y_i within the task domain estimated by conditioning on a short task-specific domain premise string x_domain (typically the trailing cue of the prompt template x, such as 'because' or 'the answer is:').