Due to the presence of long-range correlations in language [21], [22] it is not possible to compute accurate measures of the entropy by estimating block probabilities directly. More efficient nonparametric methods that work even in the presence of long-range correlations are based on the property that the entropy of a sequence is a lower bound to any lossless compressed version of it [15]. Thus, in principle, it is possible to estimate the entropy of a sequence by finding its length after being compressed by an optimal algorithm. In our analysis, we used an efficient entropy estimator derived from the Lempel-Ziv compression algorithm that converges to the entropy [19], [23], [24], and shows a robust performance when applied to correlated sequences 25.
