Seminar logo

  • Photo of Ayla Rigouts Terryn
    Multilinguality in Large Language Models
    My focus on the multilingual capabilities and limitations of large language models started with a pilot study, where I established how English bias can manifest in texts generated by LLMs in Dutch or French. A significant number of the linguistic errors and peculiarities found in these texts can be traced directly to English, often reading like overly literal translations. I intend to further pursue this line of research and combine it with an investigation of the translation capabilities of LLMs. Beyond English bias, I'm intrigued by the role of prompt language on cultural bias in LLM outputs. How significantly do LLM responses vary when the same question is asked in different languages? For instance, when seeking a comforting dish for a cold winter evening, an English prompt might yield a suggestion for beef stew, while the same query in French could result in recommendations like pot-au-feu, onion soup, or gratin dauphinois. Similarly, German and Dutch prompts tend to elicit traditional German, Dutch, or Belgian recipes. This culinary example raises broader questions about cultural biases in LLMs. How do responses differ when asking about values rather than recipes? Is the perceived importance of work-life balance consistent across prompts in various languages? How do LLMs represent the significance of family, friends, work, leisure time, politics, and religion in different linguistic and cultural contexts? Ultimately, my research aims to use my strengths as a trained translator and researcher in the humanities to improve our understanding of these extremely influential black boxes.
    Calendar icon Friday, October 04, 2024 @13h00 3195, Pav. André-Aisenstadt