Large language models are demonstrating increasing capabilities, excelling at benchmarks once considered very difficult. As their capabilities grow, there is a need for more challenging evaluations that go beyond surface-level linguistic competence. The latter involves not only syntax and semantics but also pragmatics, i.e., understanding situational meaning shaped by context and linguistic and cultural norms. To contribute to this line of research, we introduce SloPragEval and SloPragMega, the first pragmatics understanding benchmarks for Slovene, comprising 405 multiple-choice questions. We discuss the difficulties of translation, describe the campaign to establish a human baseline, and report pilot evaluations with LLMs. Our results indicate that current models have substantially improved in their understanding of nuanced language but may still fail to infer implied speaker meaning in non-literal utterances, especially those that are culture-specific. We also observe a significant gap between proprietary and open-source models. Finally, we argue that benchmarks targeting nuanced language understanding and knowledge of the target culture must be designed with care, preferably constructed from native data, and validated with human responses.

While new benchmarks for large language models (LLMs) are being developed continuously to catch up with the growing capabilities of new models and AI in general, using and evaluating LLMs in non-English languages remains a poorly-charted landscape. We give a concise overview of recent developments in LLM benchmarking, and then propose a new taxonomy for the categorization of benchmarks that is tailored to multilingual or non-English use scenarios. We further propose a registry of benchmarks implementing the new categorization and documenting benchmarks with a rich set of metadescriptors. While still at a pilot stage, such a registry can lead to a more coordinated development of benchmarks for European languages. We conclude with a review of current trends and advocate for a higher language and culture sensitivity of evaluation methods.

Dehumanisation involves the perception and/or treatment of a social group’s members as less than human. This phenomenon is rarely addressed with computational linguistic techniques. We adapt a recently proposed approach for English, making it easier to transfer to other languages and to evaluate, introducing a new sentiment resource, the use of zero-shot cross-lingual valence and arousal detection, and a new method for statistical significance testing. We then apply it to study attitudes to migration expressed in Slovene newspapers, to examine changes in the Slovene discourse on migration between the 2015-16 migration crisis following the war in Syria and the 2022-23 period following the war in Ukraine. We find that while this discourse became more negative and more intense over time, it is less dehumanising when specifically addressing Ukrainian migrants compared to others.

Large language models (LLMs) show extraordinary performance in a broad range of cognitive tasks, yet their capability to reproduce human semantic similarity judgements remains disputed. We report an experiment in which we fine-tune two LLMs for Slovene, a monolingual SloT5 and a multilingual mT5, as well as an mT5 for English, to generate word associations. The models are fine-tuned on human word association norms created within the Small World of Words project, which recently started to collect data for Slovene. Since our aim was to explore differences between human and model-generated outputs, the model parameters were minimally adjusted to fit the association task. We perform automatic evaluation using a set of methods to measure the overlap and ranking, and in addition a subset of human and model-generated responses were manually classified into four categories (meaning-, positionand form-based, and erratic). Results show that human-machine overlap is very small, but that the models produce a similar distribution of association categories as humans.

Word embeddings and pre-trained language models have achieved great performance in many tasks due to their ability to capture both syntactic and semantic information in their representations. The vector space representations have also been used to identify figurative language shifts such as metaphors, however, the more recent contextualized models have mostly been evaluated via their performance on downstream tasks. In this article, we evaluate static and contextualized word embeddings in terms of their representation and unsupervised identification of relation-level (ADJ-NOUN, NOUN-NOUN) metaphors in Slovene on a set of 24 literal and 24 metaphorical phrases. Our experiments show very promising results for both embedding methods, however, the performance in contextual embeddings notably depends on the layer involved and the input provided to the model.

The study of metaphors in media discourse is an increasingly researched topic as media are an important shaper of social reality and metaphors are an indicator of how we think about certain issues through references to other things. We present a neural transfer learning method for detecting metaphorical sentences in Slovene and evaluate its performance on a gold standard corpus of metaphors (classification accuracy of 0.725), as well as on a sample of a domain specific corpus of migrations (precision of 0.40 for extracting domain metaphors and 0.74 if evaluated only on a set of migration related sentences). Based on empirical results and findings of our analysis, we propose a novel metaphor annotation scheme containing linguistic level, conceptual level, and stance information. The new scheme can be used for future metaphor annotations of other socially relevant topics.