Makalenin niyetini yanlış anladığınızdan korkuyorum. Biraz belirsiz yazılmış olduğu için bu büyük bir sürpriz değil. İki farklı şey oluyor.
Birincisi, günlük ölçeğinde çalışmaktır.
pAB=pA⋅pBlog(pAB)=log(pA)+log(pB)". If you need the actual probability, you can exponentiate at the end to get back pAB: pAB=elog(pA)+log(pB), but if needed at all, the exponentiation would normally be left to the last possible step. So far so good.
The second part is replacing logp with −logp. This is so that we work with positive values.
Personally, I don't really see much value in this, especially since it reverses the direction of any ordering (log is monotonic increasing, so if p1<p2, then log(pA)<log(p2); this order is reversed with −logp).
This reversal seems to concern you, but it's a direct consequence of the negation - it should happen with negative log probabilities. Think of negative log probability as a scale of "rarity" - the larger the number, the rarer the event is (the article refers to it as 'surprise value', or surprisal, which is another way to think about it). If you don't like that reversal, work with logp instead.
To convert negative-log-probabilities back to probabilities, you must negate before exponentiating. If we say si=−log(pi) (s for 'surprise value'), then pAB=e−[sA+sB]. As you see, that reverses direction a second time, giving us back what we need.