Date/Time: Friday 3/10 from 2-3:30pm
Location: 32-D831 and on Zoom
Speaker: Hayley Ross (Harvard)
Title: Adjective-noun compositionality in humans and language models
Abstract: A key component of human language is compositionality: the idea that we assemble the meaning of a sentence or phrase in a structured process from the meanings of its parts. When language models encounter a phrase like blue cup or fake gun, do they engage in any compositional process? Do they know that a blue cup is a thing which is blue and which is a cup – but that a fake gun is not a gun? I focus on adjective-noun compositionality, specifically with so-called privative adjectives, to try to get insight into whether neural language models learn about compositionality from their next word prediction objective, or whether they are merely memorising specific phrases. For that matter, do humans treat rare but plausible phrases like counterfeit scarf compositionally, or do they actually also rely on conventionalised meaning to know how to interpret such combinations? What is the relationship between frequency and ease of understanding? In this talk, I’ll present a series of experiments designed to tackle these questions.