CroSloMet

Recent advancements in large language models (LLMs) have opened new avenues for processing figurative language, yet their performance in metaphor interpretation continues to fall short of human-level understanding. One limitation lies in the inadequacy of existing metaphor datasets, which often lack explicit connections to conceptual metaphors and are predominantly monolingual. In this paper, we present CroSloMet, a novel dataset of over 1,120 metaphorical and 1,120 literal sentences in Croatian and Slovene, grounded in the MetaNet.HR framework. Each example is annotated with the corresponding conceptual metaphor, linguistic multi-word expression (MWE), canonical forms, and literal usage, enabling both metaphor identification and explanation tasks. We present preliminary evaluations of the dataset through two experiments: metaphor classification using CroSloEngual BERT, achieving 88.5% accuracy, and metaphor explanation generation with LLama 3-8B, where strict exact-match evaluation yielded low scores despite semantically valid outputs. To address this, we propose a multi-level validation framework combining manual annotation, natural language inference, semantic similarity, and LLM-based judgment. Our findings highlight the importance of capturing generality and specificity in metaphor mappings and call for more nuanced evaluation methods. CroSloMet provides a resource for advancing metaphor understanding in LLMs and contributes to cross-linguistic and cognitively informed metaphor research.