Here it is.It fails routinely on topics with vastly more training data than ASL topics. It does particularly bad on things with minimal training data like ASL. There was a thread a month or so ago that covered this topic in depth that you might want to look up.