No language model has yet answered my opening sentence correctly. I always ask that the lyrics ‘Crisp morning in Paris’ be completed correctly. The language models, no matter how advanced, come up with flowery scenes in the French capital with the Eiffel Tower and croissants, but they overlook something that I expect a person will quickly see; those words form the beginning of the song Paris by Kenny B. There is a reason that I think of that song, because the chorus starts with ‘talk Dutch with me’, which language models can do, but do to a relatively limited extent.
That is logical, because many language models are trained mainly on English texts. Take Meta’s relatively new Llama 3, which will become the basis for AI in WhatsApp and Instagram. About 95 percent of the training data is in English. So it is logical that Llama 3 is not as good in Dutch as in English.
Language models are often good at translating things, but what about talking from the culture? It is easy for an AI chatbot to translate English texts, but it only becomes truly Dutch or Flemish if it has been trained on texts from the Low Countries. After all, that training data not only contains language, but also culture is secretly hidden.
So it’s time to get started with some variants of Dutch models. Is it better and more fun to chat with a Dutch-speaking model than to chat with an American one? large language model?
Image at top and on front page: Dall-E with prompt ‘a Dutch language model with a mix of typically Dutch and Flemish elements’
