The algorithm starts with a large, untagged corpus, in which it identifies examples of the given polysemous word, and stores all the relevant sentences as lines.
In this case, the words "life" and "manufacturing" are chosen as initial seed collocations for senses A and B respectively.
The algorithm should initially choose seed collocations representative that will distinguish sense A and B accurately and productively.
Add those examples in the residual that are tagged as A or B with probability above a reasonable threshold to the seed sets.
At the end of each iteration, the "one sense per discourse" property can be used to help preventing initially mistagged collocates and hence improving the purity of the seed sets.
In order to avoid strong collocates becoming indicators for the wrong class, the class-inclusion threshold needs to be randomly altered.
When the algorithm converges on a stable residual set, a final decision list of the target word is obtained.