Very interestring. I might be actually wrong here. So I assumed the problem is not cultural bias, but a tokenizer which does not have the concept of open compo

Very interestring. I might be actually wrong here. So I assumed the problem is not cultural bias, but a tokenizer which does not have the concept of open compound words. For example a naive implementation which is just splitting the strings into chunks of printable characters would be suspectible to this problem. To test this I have tried it with various open compound words. My theory was that it will reject all of them. Turns out it is not the case! The site rejected “sweet tooth”, “apple tree”, “chow mein”, “banana split”, “school bus”, “life jacket”, “mobile phone”, but curriously accepted “rain forest”. This means that it has the ability to sometimes recognise open compound words. Curiouser and curiouser!