Probing neural language models for understanding of words of estimative probability