Large Language Models Don't Make Sense of Word Problems