Probing Neural Language Models for Human Tacit Assumptions

TL;DR

This paper investigates whether neural language models inherently encode human stereotypic tacit assumptions by evaluating their ability to retrieve concepts based on associated properties, using prompts inspired by psychological studies.

Contribution

The study introduces a diagnostic prompt set to assess if neural models capture human-like stereotypic assumptions about concepts, providing empirical evidence of such representations.

Findings

Models effectively retrieve concepts from associated properties.

Neural models encode stereotypic conceptual associations.

Empirical evidence supports the presence of tacit assumptions in models.

Abstract

Humans carry stereotypic tacit assumptions (STAs) (Prince, 1978), or propositional beliefs about generic concepts. Such associations are crucial for understanding natural language. We construct a diagnostic set of word prediction prompts to evaluate whether recent neural contextualized language models trained on large text corpora capture STAs. Our prompts are based on human responses in a psychological study of conceptual associations. We find models to be profoundly effective at retrieving concepts given associated properties. Our results demonstrate empirical evidence that stereotypic conceptual representations are captured in neural models derived from semi-supervised linguistic exposure.