Biases and opportunities in using Large Language Models
- Date
- Friday 16 May 2025, 10:00-14:00
- Location
- Parkinson 2.27
You're warmly invited to a presentation reporting the outcomes of our study of biases and opportunities in using Large Language Models and to a discussion session afterwards.
The event will take place on Friday 16 May 2025 10:00-14:00 in Parkinson 2.27.
Topic:
Large Language Models, such as ChatGPT, have already started making an impact on many areas of our lives, and their influence is likely to expand. However, despite their empirical success, these models can also introduce biases and hallucinations caused by the nature of their training process. We will present the results of a study we conducted with Key Topics funding with the aim of investigating the decision-making processes of Large Language Models by examining (1) what they successfully extract from legal and medical texts, what they fail to extract, and the factually incorrect predictions they make, and (2) how the quality of these extractions varies across different languages, specifically English, French, Igbo, and Nigerian Pidgin.
Schedule:
10:00-10:45, Serge Sharoff, Biases and opportunities in using Large Language Models
10:45-11:30, Tony Cohn, Can Large Language Models reason about Spatial Information?
11:30-12:15, Adaeze Ohuoba, Law and Disorder: When LLMs Produce Hallucinations in Multiple Languages
12:15-13:00, Discussion
13:00-14:00, Lunch
In this programme, we will also provide space for discussing the challenges and opportunities in the use of LLMs. Please share your experience on how the current models deal with critically important domains and with lesser-resourced languages.