Karma in websites is not necessarily about accuratly reflecting some ground truth upvote probability, mean chance of liking a comment, expected future vote ratio of the commenter, etc. It's an incentive-design mechanism, in that a karma system is good if and only if it leads to the desired behavior when people use the website. When you ask yourself "how should I compare 5 upvotes and 5 downvotes versus 1 upvote versus 1 downvote versus no action at all", the answer is not weighting one of these situations higher/lower because it will better approximate one of the criteria above in expectation or something like this, but instead weighting these high/low depending on, for example, if you want to encourage activity, agreement, controversy, etc.
Ideally, you should have some other behavioral metric in mind (say, mean comment quality, top comment quality, bottom comment quality, engagement, etc) and try to tune the voting system to maximize this quality over time. (this tuning can either be done intuitively, as pg tries to do, or algorithmically, using something like the technique behind Gmail's priority inbox or bandit algorithms)
Do not "define away" a social problem with mathematics, use the mathematics to help you solve the actual social problem.