Prof. Reut Tsarfaty

Generative AI in general, and specifically generative Large Language Models (LLMs), have changed the game for many tasks and domains in the field of natural language processing. Indeed, the ability to comprehend and generate natural language in a human-like way is instrumental in many language technology applications – from sentiment analysis and information extraction to dialogue systems.

However, much of this success has been demonstrated in English. For a resource-scarce language such as Hebrew, with its profound linguistic structure and morphological intricacies, do LLMs hold the same promise? Will Hebrew speakers be able to use generative AI in their native tongue?

In this talk, I will present three different threads from my research lab, focusing on Hebrew LLMs. One concerns the LLM architecture and, in particular, inquires about the correlation of LLMs’ performance with properties of the tokenizer and the LLM’s internal lexicon. The second concerns the use of LLMs, particularly how to develop good prompts for Hebrew tasks, in monolingual and multilingual settings. The last one concerns the evaluation of LLMs on real downstream tasks, where I will present benchmarks for challenging applications such as summarization and credibility (fake news), and news detection in Hebrew.

All in all, I will show that on the path to LLM Hebrew fluency, interesting research questions emerge, providing useful answers and further insights into this groundbreaking technology.