Improving NER State-of-the-art for Latvian with LLM and Few-Shot Learning

Developing Named Entity Recognition (NER) solutions for morphologically rich but low-resource languages like Latvian is complex.Most state-of-the-art methods rely on deep learning models like BERT (Bidirectional Encoder Representations from Transformers), which require substantial expertise in architectures, methods, and access to extensive computational resources and data.This study shows how to use popular large language models (LLMs) in a few-shot setting, without fine-tuning to surpass state-of-the-art results.We evaluate their performance on the publicly available Latvian dataset (Gruzitis, et al., 2018).We analyze the performance of LLMs on the recognition of different NER types.