Exploring Neural Entity Linking With Pre-Trained Transformer Models For Turkish

Official URL: https://risc01.sabanciuniv.edu/record=b3205849

Entity Linking, a vital component of Natural Language Processing (NLP), aims to link named entities in a given text to their corresponding real-world entities in a knowledge base. This work presents an exploration of transformer-based Neural Entity Linking models adapted to the Turkish language, focusing on their robustness across genre and domain shifts. We take two advanced Neural Entity Linking models originally designed for English and adapt them to Turkish. We then thoroughly assess how well these adapted models perform on different Turkish datasets, along with a new method we developed called EntityBERT, which serves as a reference point for comparison. EntityBERT is a simple Neural Entity Linking model which exploits pretrained Turkish transformer model and contextualized learning capabilities of transformer models. The evaluation was conducted on three distinct datasets, including one newly created dataset, publicly available for further research. The findings revealed that Neural Entity Linking models exhibited robust performance across language and genre shifts, demonstrating their adaptability to Turkish and diverse textual genres. Nonetheless, our investigation also highlights a noteworthy limitation: the susceptibility of Neural Entity Linking models to domain shift challenges. Despite their favorable performance in general settings, adapting to domains with distinctive characteristics poses considerable difficulties. Overall, this study sheds light on the potential and limitations of Neural Entity Linking models in Turkish, provides an evaluation dataset of Turkish tweets, and finally delivers valuable insights for advancing the field of natural language processing in non-English languages.