Negative values can reflect reverse scored questions (although I can't see how this could occur in your example) or unusual variation. I have never heard it suggested that only two observers influence alpha. Any negative results make no substantive sense. I don't understand how you can have no "absolute answer" for eye contact, but I am sure the more vague answers could be the more difficult reliability would be. I am not sure how you even score results that have to be interpreted this way.
Why not create a result that does have a clear, absolute answer for the raters?