for the computational reasons we used the log Preplrexity measure as the measure of fittness for each language model. 
This choice make since the log function is a increasing function and we can use an increasing fuction of each performance measure
as new performance measure. 

Using un-smoothed Language Model for predicting the perplexity we will encounter computational error. since there is a three gram 
"af'" in the test data but there is not such thing for examlpe for the english train data the perplexity of Englsih Language model will
be equal to infinity.