Human language inspired AI – and now we can use that AI to learn about language
PhD candidate Yuchen Lian (LIACS) wants to understand why human languages look the way they do – and find inspiration to improve AI along the way. She defended her thesis on 12 December.
‘Languages change all the time. Think about ancient Chinese and modern Mandarin. There is a huge difference between them,’ Lian starts talking about her research enthusiastically. ‘And even more so between different languages.’
At the same time, there are universal characteristics, the PhD candidate explains. ‘Linguists want to understand which features are common and why they appear. With the increase in computational power of the last years, we can now learn about language evolution through computer models in increasingly realistic settings.’
‘I believe it's crucial to combine computer science and linguistics to reach a full understanding of language evolution.’
Artificial miniature languages
Lian completed her PhD under the supervision of Tessa Verhoef and Arianna Bisazza, and with support from the China Scholarship Council. As part of the programme, she spent three years at Leiden University before returning to China to continue her doctoral work. ‘I missed Leiden quite a lot, the atmosphere here is really good. I’m happy to return for my PhD defence.’
In her research, the PhD candidate is inspired by miniature languages used in linguistics experiments with people. Participants are given three words: a subject, object and verb. For example, cat, mouse, and chasing. In various assignments, researchers monitor how they use these words. In English, for example, the word order is fixed, explains Lian. “The cat chases the mouse” is the only correct sentence. But in Japanese, there are markers after a word to indicate its function. The order of words is therefore more flexible.
A conversation between AI agents
Lian models two or more AI agents that can communicate with each other. These agents are first trained on a pre-defined artificial language, like the experiments with human participants. ‘We basically give them a vocabulary at the start,’ says Lian. ‘Then we let the agents interact in pairs or even groups through interactive language games. When they accomplish a task successfully, all agents get a reward. That's how they learn.’
The different agents can have the same vocabulary, but different grammar rules to analyse various scenarios. The results show that these agents can replicate the trade-off between word order flexibility and the use of markers. ‘So we know the model works. Moreover, it allows us to extend the results with humans to a larger scale. This complements experiments with humans nicely.’
Human language inspires AI chatbots
Lian opened a new avenue to learn about language evolution. At the same time, her work has the potential to improve AI. ‘We can take inspiration from human language and use that to improve the outcomes of an AI chatbot, for example. Currently, they are trained predominantly through passive exposure to enormous amounts of data. But humans clearly acquire language in a much more interactive way. My simulations demonstrate that repeated communication can result in more efficient interactions and spontaneous emergence of human-like patterns.’
A computer scientist in the field of linguistics
‘My topic extends over the borders of computer science,’ reflects Lian on her PhD. ‘It's quite a challenge to get into this interdisciplinary field. We speak different languages and have different standards in publishing, for example. It was sometimes hard for me that language evolution has no ground truths, propositions cannot be proved in a mathematical sense.’
At the same time, she also found the combination to be fulfilling. ‘Luckily, my promotors had expertise in both fields. I believe it's crucial to combine computer science and linguistics to reach a full understanding of language evolution.’
Lian defended her thesis successfully on 12 December. Aske Plaat was the promotor, Tessa Verhoef and Arianna Bisazza the co-promotors.