Rethinking the Idiomaticity Decomposability Hypothesis: Evidence from Distributional Learning
Researchers have investigated the concept of idiomaticity decomposability in language models, exploring how constituent meanings contribute to figurative language. Their findings suggest that while decomposability is thought to predict syntactic flexibility, usage-based accounts emphasizing distributional experience are more relevant. The study utilized contextualized language models as distributional learners and developed a model-internal measure of decomposability, finding it correlates weakly with human judgments and has a negative relationship with syntactic flexibility. Analysis of model pretraining revealed that idiom representation stabilization is influenced by surprisal, decomposability, and frequency, with decomposability showing the most significant training-dependent effect. AI
IMPACT This research provides insights into how language models learn and represent idiomatic expressions, potentially informing future model development and evaluation.