Honorary doctor Chris Manning: “Language models write credibly, but not necessarily the truth”
On the occasion of its 391st dies natalis, the UvA awarded two honorary doctorates today, one of them to Australian Christopher Manning, professor of Computer Science and Linguistics at Stanford University. Folia spoke with him about how to learn a computer language and the dangers of Artificial Intelligence.
A conversation with Christopher Manning (57) may seem easy, but it is anything but, he tells himself via video link from Stanford, just before his flight to Amsterdam. He is referring to people's ability to use language to exchange thoughts, feelings, and experiences on a detailed level.
“It seems so easy, but think about it. Language moves very quickly, I talk and you understand. While most words are very ambiguous, they have lots of meanings. And often people refer to things that have happened before or are going to happen in the future. All that happens so fast. It's unimaginable and still completely misunderstood. Understanding how people learn language is one of the challenges to better understanding the world around us. I strongly believe that language is, par excellence, the human trait that sets us apart from other animals.”
Manning is receiving his first honorary doctorate for his key role in the field of natural language processing by computers, also known as Natural Language Processing (NLP). Since early 2010, his research group at Stanford has been a world leader in the revolutionary developments of neural networks for NLP. His book, Foundations of Statistical Natural Language Processing, is the standard work in the field and required reading for many generations of students.
The connection between Manning and the Institute for Logic, Language and Computation (ILLC) at the UvA traces back to Johan van Benthem, emeritus university professor of logic and founder of the ILLC institute. Van Benthem spent an academic quarter each year at Stanford. Manning, still in college at the time, took several courses with him.
Were you surprised by the honorary doctorate from the UvA?
“Yes, I was surprised; it was unexpected. The ILLC Institute is at the forefront of linguistics, semantics, and the logic of language. It does a lot of work that overlaps with mine. And the Amsterdam Machine Learning Lab (AMLab) is leading Europe in the use of neural networks in machine learning. This also involves my honorary supervisor, Professor of Machine Learning Max Welling.”
You have been doing research on NLP for more than 30 years. Can you take us through the developments of how computers can understand language?
“In my student days, in the late 1980s, people tried to teach language to computers by introducing rules, grammar, and logic into those computers. People believed that the more knowledge you fed into the computer, the smarter it became. This was before the introduction of the World Wide Web.”
“While I was getting my PhD, digital texts became available, such as court proceedings and newspaper archives. Then the approach changed. Namely, in those millions of words of text, you could find patterns using statistics. That seemed like a fascinating way to look at language. We call that statistical NLP. Then that merged with machine learning, a form of artificial intelligence that can discover patterns on its own in a huge amount of data.”
“And then there's another term, deep learning, a part of machine learning that relies on neural networks. A neural network learns by itself, based on examples. The first attempts to get neural networks working took place back in the 1950s, the second round came in the 1980s. Then it was sidelined by mathematically clear, static machine learning techniques. Still, a few people continued to believe in neural networks and work on them. Around 2010, interesting results came out. From then on, I also joined the research, even before the real breakthrough in 2012-2013. That combination of NLP specialist and researcher in neural networks provided the second shift in my career.
What did the big breakthrough in neural networks entail?
“In 2016, neural machine learning worked so much better than static machine learning that all the tech companies switched to it en masse. And that was just the beginning. In 2018, they discovered that if you take a huge amount of human language and train a giant neural network to predict words, you get a model that can do much more. GPT-3 is an AI model trained with language. It can write paragraphs indistinguishable from human writing. But it can also, for example, summarize a news article in three sentences, or predict the folding of a protein. It blows my mind that that works.”
Can computers ever fully understand human language?
“There are some deep philosophical questions about that, but I think we will eventually get to the point where computers fully understand language. Right now, computers learn language by reading a huge amount of text and learning to predict what people are saying or writing. This gives computers a much broader understanding of text than some might imagine. Yet there are limitations to this. For a full understanding of language, computers or robots will have to connect with the world much more actively. One of the philosophical questions is whether a computer can ever truly understand happiness or despair without an emotional response system as humans have. I'm not sure of the answer to that question.”
In an earlier interview in Folia, UvA researcher Jelle Zuidema warned of the dangers of language models that we do not yet fully understand. How do you view this?
“On the one hand, language models will inevitably be used; they are already being used. On the other hand, there are legitimate reasons for concern. Misinformation and alternative facts may increase with these tools. After all, language models also make things up. They write credibly, but not necessarily the truth. The language model ChatGPT-3 is very good at answering short essay questions, such as: write four paragraphs about the origins of World War I. If you ask the model that 20 times, you get 20 different stories. So a lot of people worry when students start taking exams with ChatGPT-3.”
Do you worry about that too?
“Yes, I think it's very legitimate to worry. My fellow Stanford researcher Chris Potts already made a good suggestion, though. He's going to ask questions of ChatGPT-3 himself this quarter. He will then select the answers with inaccuracies and students will have to identify the errors. Students will have to separate the sense from the nonsense.”
According to Elon Musk, AI is the greatest existential threat to humanity, even more dangerous than nuclear weapons. What do you think about that?
“I'm not convinced of that. There are several prominent people, all white men by the way, who want to talk about the existential consequences of AI and the possibility of AI taking over our planet. I see no reason why this would happen. Besides, I think the obsession with the existential threat of super-intelligent AI is also a way to spend your energy on abstract future problems, rather than the problems that exist in our society in 2023. Our time is better spent on that.”
