Perspective matters: event framing in language and society

Did "a cyclist tragically die in traffic" or did "a car driver hit and killed a cyclist"? Do we say "three injuries during protest" or "riot police injures three protesters"? Different linguistic choices to describe the same event give rise to different framings of that event and can influence people's perception of what happened. While this is a well-known concept in social sciences and discourse studies, it so far has received little attention in computational linguistics. In this thesis, we develop SocioFillmore: an approach that integrates insights from cognitive linguistics and modern Natural Language Processing (NLP) models to automatically analyze how societally-important events are conceptualized in language. We draw extensively on Fillmore's theory of Frame Semantics and investigate end-to-end Frame Semantic Parsing systems for Dutch, Italian, and English. We use SocioFillmore to identify agent-backgrounding constructions in two large datasets of femicides in Italian news articles and traffic crashes in Dutch news articles, and build a web interface for finding patterns in these datasets. We also adapted the system for studying the framing of migration in Italian newspapers. Going beyond linguistic analysis, we also conducted two perception studies: in the first study, we show how different ways of describing a femicide lead to different perceptions as to who is held responsible, and show that NLP models are able to predict perception scores from raw text. In a second study, we provide a proof-of-concept that generative NLP models are also able to rewrite texts to suggest alternative framings of femicide news.