The Holistic Storage of Verb+Up Phrases in Text-based and Audio-based Language Models

This article presents research on the holistic storage of verb+up phrases in both text-based language models and an automatic speech recognition (ASR) model, revealing that these models develop distinct representations for such phrasal verbs influenced by their frequency and predictability. The findings provide empirical support for usage-based theories of language, highlighting the importance of considering multi-word units in model training and evaluation. This research is significant for practitioners as it suggests that integrating frequency and predictability into training data could enhance the performance of LLMs and ASR systems in understanding and generating natural language.