Exploring and Improving the Treatment of Metaphor in Large Language Models (METALLM)

The project’s objective is to explore and improve how large language models interpret and produce metaphors by clarifying their relationship with cognition, language, and culture. METALLM will use eye-tracking patterns as a window to human cognition and combine them with linguistic features across languages and cultures.

Metaphors are pervasive in our daily life, not limited to language, but also intimately involved in our cognition. They offer cognitive benefits, improve communication and elicit strong emotional responses. The use of metaphors in communication is however no longer exclusive to humans. Recent advancements in AI have resulted in language models capable of generating and interpreting metaphors, although often with limited precision and strong language bias. 

We explore and improve how models interpret and produce metaphors by clarifying their relationship with cognition, language, and culture. We take an innovative approach by using eye-tracking patterns as a window to human cognition and combine them with syntactic and semantic features across languages and cultures. The project increases the transparency of language models in the area and paves the way to improve figurative language processing, leading to models that have a better understanding of the linguistic and cultural diversity of metaphors.