In language modeling, L(lambda1, lambda2, lambda3) is defined as:
Sum(count of trigram(u,v,w) x q(w|u,v))
where u, v, w are words in the corpus and perplexity is defined as:
2^-l
where
l = (1/M)Sum(log(q(w|u,v)).
where M is the total no. of words in the corpus.
Also,
q(w|u,v) = lambda1*q(w|u,v) + lambda2*q(w|v) + lamdba3*q(w)
The q-values on the right hand side are maximum likelihood estimates. Some of the materials for Natural Language Processing state that maximizing L is same as minimizing perplexity. I don't see how that is true or can be mathematically proved.
