The simple trick is to take pairs of bits. If you get 01, output 0. If you get 10, output 1. Otherwise, output nothing. Even if the source is biased (as long as

The simple trick is to take pairs of bits. If you get 01, output 0. If you get 10, output 1. Otherwise, output nothing. Even if the source is biased (as long as successive bits are uncorrelated), you get an unbiased 50:50 output. But obviously you get fewer bits out than you put in. The construction in the paper is basically a more efficient version of this.