“We are trying to understand how these systems deal with information. What is the thought process inside these systems when it comes to answering factual questions, says Richard Johansson, who was promoted to full professor in April.
His field of study, natural language processing, concerns how computers deal with human language. This ranges from grammar checking and spam filters in emails to, as in Richard Johansson’s current project, trying to understand complex large language models (LLM).
Large language models are the foundations of tools like Chat-GPT, and they generate answers based on probabilities. When these systems are unsure of an answer, they tend to fabricate responses they find likely. Richard Johansson and his team are now trying to find out how these systems decide on an answer.
“Our goal is to find a method of understanding how the algorithm acts when it knows something compared to when it does not. To see if we can distinguish cases when fact is present in the model from when it is not.”
A changed research landscape
The introduction of these AI-based language models worldwide has significantly changed Richard Johanssons’s field. While it has become easier to explain what he does to family and friends, there are now also huge companies competing in creating the best models.
“The type of research that these companies do requires a lot of resources and infrastructure and is not something that Academia can, or should, compete with”, Richard Johansson argues.
Instead, he suggests that researchers should approach this phenomenon from a different angle.
“We need to try to understand this development, how these models function and why they act in a certain way. For me as a researcher the biggest challenge is to find relevant, important work that is feasible on a smaller scale.”
Curiosity-driven research
For Richard Johansson, it was the combination of language and computers, his two main interests, that led him into the area of natural language processing. Much of Richard Johansson’s research is so-called basic research, meaning research aimed at creating new knowledge rather than research with a practical application.
“As a researcher I am driven by curiosity. I find an interesting research problem fascinating in itself. While there of course could be practical applications of my research, my primary motivation stems from trying to understand”, Richard Johansson explains.
In the case of large language models, however, he sees a direct societal need for better understanding these systems.
“These systems will obviously have a great societal impact and therefore need to be reliable – or at least we need to understand when they are not reliable. Knowing when to trust these systems allows us to use them safely.”
Richard Johansson and his three PhD students Lovisa Hagström, Mehrdad Farahani, och Nicolas Audinet de Pieuchon are part of the research group NLP@DSAI, where they investigate methods to understand langugage models and how to apply them to political and social sciences.
Text Natalija Sako
