Neural language models represent an important advance in natural language processing. Tools like ChatGPT and Gemini can hold coherent conversations and seem to be able to reason intelligently. A key factor in their success appears to be the availability of large quantities of data, without the need for architectures that incorporate language specific mechanisms. 

But are they comparable to humans in the way that they learn and process languages? Some experiments find that they handle complex phenomena appropriately (e.g. Wilcox et al., 2023), whereas others find they fail on simpler tasks when the test examples are novel (e.g. Wilson et al., 2023; Berglund et al., 2023). A long-standing criticism of neural approaches is that they struggle to generalise outside their training data (e.g. Marcus, 2018). Bender et al. (2021) attacked the reliance on huge training datasets directly, accusing the resulting models of being “stochastic parrots” which simply regurgitate patterns they have already seen. This criticism raises the question of whether current language models are too general (e.g. Mitchell & Bowers, 2020 vs. Kallini et al., 2024) rather than being language specific and so lack the biases that are needed to learn as efficiently as humans. 

Are there insights from cognitive science and psychology that will improve their ability to mimic human language users and to make better use of their training data?  

Some researchers have pursued architectures that are more closely related to linguistic models of the structures of language (e.g. Sartran et al., 2022). Others have evaluated the performance of various approaches when they are trained on the same data as human children (e.g. Warstadt et al., 2023). 

Can you produce a research proposal outlining the directions of research you think would be most fruitful for answering these questions and some specific examples of computational experiments that you might run? 

Get in touch (jm2136@sussex.ac.uk) if any of this seems interesting and you want to discuss this opportunity further. 

This scholarship is available to UK, EU and overseas applicants. 

Eligible candidates will have an upper second-class (2:1) undergraduate honours degree (or equivalent qualification) in a related field. 

The University of Sussex believes that the diversity of its staff and student community is fundamental to creative thinking, pedagogic innovation, intellectual challenge, and the interdisciplinary approach to research and learning. We celebrate and promote diversity, equality and inclusion amongst our staff and students. As such, we welcome applicants from all backgrounds.

Please clearly state on your application that you are applying for the Cognitive Plausibility of Neural Language Models scholarship under the supervision of Dr Jeff Mitchell.

Please ensure you application includes each of the following: 

• A research proposal. For guidance on preparing your proposal, please see here. 
• Degree certificates and transcripts. 
• 2 references, including a minimum of 1 from any institution studied at within the last 5 years.
• If your first language is not English you will need to demonstrate that you meet the University’s English language requirements, see here for details of our accepted documentation.