| ## LCHAIM: A Hebrew Natural Language Inference Dataset | |
| ### Overview | |
| LCHAIM is a dataset designed to evaluate Natural Language Inference (NLI) models in Hebrew. Unlike English, Hebrew is a Morphologically Rich Language (MRL), requiring more research to develop robust NLI models. LCHAIM provides a benchmark for models that need to handle long premises and complex reasoning in Hebrew. | |
| ### Dataset Description | |
| LCHAIM was created by translating and validating the English ConTRoL dataset into Hebrew. It consists of 8,325 context-hypothesis pairs that require various types of reasoning, including: | |
| * Coreferential reasoning | |
| * Temporal reasoning | |
| * Logical reasoning | |
| * Analytical reasoning | |
| ### Performance Benchmarks | |
| Experiments with LCHAIM highlight the challenges of contextual reasoning in Hebrew. Key results include: | |
| Fine-tuning the LongHero model on both Hebrew NLI datasets and LCHAIM yielded a mean accuracy of 52%, which is 35% (absolute) lower than human performance. | |
| Large Language Models (LLMs) in a few-shot setting achieved the following top mean accuracies: | |
| * Gemma-9B | |
| * Dicta-LM-2.0-7B | |
| * GPT-4o | |
| Top performance: 60.12% mean accuracy | |
| ### Citation | |
| If you use LCHAIM in your research, please cite our work, which should be published in the next month | |
| ### License | |
| LCHAIM is released under the mit license. | |
| ### Contact | |
| For questions or feedback, please contact orielpe@post.bgu.ac.il | |