Figurative language, such as metaphors and idioms, constitute a challenging testbed for Large Language Models and processing of these are a growing field in NLP. Their understanding requires language understanding capabilities that distinguish between figurative and non-figurative usages of one and the same expression and that are not based on compositional understanding. Furthermore, figurative language is highly language-specific, hence it will often be the case that idiomatic expressions in one language (e.g. “kick the bucket”, meaning “die” in English) does not have a corresponding expression in other languages, such as Norwegian.
LTG has recently released a number of Norwegian, generative LLMs and has a lot of activity related to benchmarking of these models. The goal of this project will be to gain a better understanding of the capabilities of generative LLMs for understanding of Norwegian figurative language. The thesis will involve the following steps:
- creating a benchmark dataset for understanding of figurative language such as idioms for Norwegian.
- comparing multilingual models and Norwegian models for figurative language understanding
- potentially also examine methods for improving figurative language understanding in LLMs
The thesis requirements are the following:
- Norwegian native speaker with a linguistic interest
- experience with neural NLP models (i.e. will take IN5550 in the spring)