Despite recent advancements in Large Language Models (LLMs), idiomaticity remains a significant obstacle to robust semantic representation (He et al, 2025). While there has been considerable work on the evaluation of idiomatic language using LLMs, there is currently less knowledge about idiomatic interpretation in a multimodal setting (i.e. combining image and text in so-called Vision-Language Models (VLMs)), in particular for underresourced languages, such as Norwegian.
The recent SemEval shared task AdMIRe (Pickard et al, 2024) released a dataset for evaluating multimodal interpretation of idiomatic expressions in English and Portuguese. The second iteration of the task extends the set of languages to over 30 languages, including a small sample for Norwegian. This thesis will investigate multimodal interpretation of Norwegian idioms and has several possible directions depending on the interests and background of the student, including:
- extending the Norwegian part of the XMPIE dataset
- assessing the effect of different image generation models for data creation
- evaluating multilingual VLMs on the dataset
- analyzing performance wrt linguistic properties of the idioms (frequency, part-of-speech, compound vs non-compound etc)