What Kind Of Language Statistics Must Be In Long-Term Memory To Make Language Understanding Possible? A Computational Perspective