A Danish semantic reasoning benchmark compiled from lexical semantic resources This benchmark is the first version of a semantic reasoning benchmark for Danish compiled semi-automatically from a number of human-curated lexical-semantic resources, which function as our gold standard. The datasets constitute a benchmark for assessing selected language understanding capacities of large language models (LLMs) for Danish.
Current version is a beta version and is not yet fully curated. Compared to the first version, errors in the “semantic inference” dataset have been corrected. Furthermore, the data in “inference-point-in-time.txt” was substituted with more concise terms for “points in time”.
This version comprises 26 datasets across 7 different tasks and include 4,200 test instances. Although still limited in size, we go beyond comparative evaluation datasets for Danish by including both negative and contrastive examples as well as low-frequent vocabulary; aspects which tend to challenge current LLMs when based substantially on language transfer. The datasets focus on
- semantic inference
- entailment
- similarity
- relatedness,
- disambiguation of words in context.
- sentiment
We have used ChatGPT to assess to which degree our datasets challenge the ceiling performance of state-of-the-art LLMs, average performance being relatively high with an average accuracy of 0.6 on ChatGPT 3.5 turbo and 0.8 on ChatGPT 4.0.
We hope to extend the datasets in the near future depending on funding.
Bolette Pedersen, Nathalie Sørensen, Sussi Olsen, Sanni Nimb, and Simon Gray. 2024. Towards a Danish Semantic Reasoning Benchmark - Compiled from Lexical-Semantic Resources for Assessing Selected Language Understanding Capabilities of Large Language Models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 16353–16363, Torino, Italia. ELRA and ICCL.
and