From Trial by Fire To Sleep Like a Baby: A Lexicon of Anxiety Associations for 20k English Multiword Expressions

Researchers at Canada's National Research Council (NRC) have published a groundbreaking AI resource: the first large-scale lexicon mapping anxiety associations for over 20,000 English multiword expressions (MWEs). The work, led by Dr. Saif M. Mohammad and detailed in a paper submitted to LREC 2026, systematically quantifies how common phrases like 'trial by fire' or 'sleep like a baby' are perceived in terms of anxiety or calmness.

The technical achievement lies in creating descriptive norms—reliable human judgments—for these complex linguistic units. The researchers analyzed how anxiety associations vary across two-, three-, and four-word sequences, and investigated the extent to which a phrase's emotional weight is 'compositional' (derived from its individual words) versus an emergent property of the phrase itself. This addresses a significant gap in computational linguistics and affective computing, where most emotion lexicons focus on single words.

This resource enables a wide variety of new research. In psychology and public health, it allows for large-scale analysis of anxiety-laden language in social media, literature, or clinical texts. For NLP and AI, it provides crucial training data to build models that better understand nuanced human emotion in longer text segments, improving everything from mental health chatbots to content moderation systems that can detect subtle distress signals. The lexicon is freely available, lowering the barrier for interdisciplinary studies on language and emotion.