Intrinsic and Extrinsic evaluation of biases in large language models – InExtenso

Large Language Models (LLM) are the Swiss army knife of today’s Natural Language Processing (NLP). They often outperform the state-of-the-art on benchmarks commonly used in the field for tasks such as part-of-speech tagging, text classification and named-entity recognition, thus paving the way to a myriad of end-user applications. However, it has been shown that LLM exhibit major ethical issues including significant environmental impact, mirroring and amplification of stereotyped biases, which in turn have a disproportionate impact on historically disadvantaged social groups. It is urgent to address the social impact of NLP as the applications we develop, such as chatGPT, are now directly made available to end-users. The detection and mitigation of biases has therefore become an active area of research in the past few years, focusing mainly on Masked Language Models (MLM) such as BERT in English and the north American social context. Several sources of bias were identified in the NLP pipeline, however the inter-connection between sources and overall impact of each source on downstream applications remains unclear. In this project, we want to observe the entire pipeline, from the intrinsic point of view (within the model itself), to the pre-training task point of view (in the case of auto-regressive LLM, text generation), on to some real-world downstream applications. We chose to focus on two types of medical applications: mental illness diagnosis help and information extraction from clinical records for public health purposes such as patient enrollment into clinical trials. The project will provide corpora and methods for a global evaluation of bias
in LLM in French as well as studies to further the understanding of biases in clinical NLP pipelines and the environmental impact of the integration of these models in digital health.