What’s the data set from? It will at least be skewed towards people most posting to the social media networks scraped for training data. I wouldn’t be surprised if foreign language social media was substantially underrepresented in the data set because the programmers putting it together weren’t as familiar with it.