Can Graph Neural Networks Learn Language with Extremely Weak Text Supervision?