In her lecture Vanja Subotić will deal with the methodological and ethical challenges that arise in the application of large language models in clinical neuroscience and psychiatry. The lecture is scheduled for December 20 at 11:00 a.m. and will be live in the Great Hall of IDN. The broadcast will be via Zoom.
Natural language processing is a technology that has revolutionized cognitive models of language competence based on the artificial neural network research program. This technology analyzes a large amount of textual data in a statistical way, combining it with a specific type of neural network and an algorithm for training that network. The most recent models, the so-called large language models, implement artificial neural network transformers and deep learning and reinforcement learning algorithms. With such models, which are present in popular chatbots, such as ChatGPT, it is possible to process part of the text data at once, that is, to associate with each token all the contexts in which that token occurs.
Methodological challenges concerning the use of models based on natural language processing concern the explanation of human language competence. These challenges were mostly related to the model’s ability to provide explanations at all or that they were not sufficiently biologically and psychologically plausible. Closely related to the methodological challenges are the ethical challenges: datasets for model training omit the cultural and ethnic diversity of idiolects, sociolects and dialects, and we cannot even assume that they are sufficiently inclusive in terms of linguistic behaviour. On the other hand, the specific application of large language models in cognitive and clinical neuroscience and psychiatry rarely intersects with these challenges. These models are understood as neutral diagnostic tools. For example, people’s daily records reflect various patterns of thoughts and emotions, which, through natural language processing, can be observed and analyzed as indicators that would lead to early diagnosis of a specific mental disorder more efficiently than in the case of human expert knowledge. In her presentation, the author will critically review the assumed neutrality of this methodological tool, contrasting this assumption with the aforementioned methodological and ethical challenges.
Vanja Subotić is a researcher-associate at the Institute of Philosophy of the Faculty of Philosophy of the University of Belgrade and an external collaborator on the TechEthos project within Horizon Europe. Her research areas are philosophy of language and linguistics, philosophy of cognitive science and experimental philosophy. She received her PhD on the role of natural language processing models based on deep learning algorithms in explaining human language competence. He is actively engaged in scientific communication and popularization of ethical challenges of natural language processing technology.
https://zoom.us/j/98089508962?pwd=Q1Rsb2lQMFVMUWNvZUozU3ppSlRSZz09
Meeting ID: 980 8950 8962
Passcode: 410968
