While it does appear rather arbitrary, it wasn't a homework problem. I work at the Panasonic Speech Technology Research Lab in Santa Barbara (now PDNL for a number of asinine and annoying reasons) and we were in the process of developing a japanese recognizer (hence the Japanese exaMPLE), however, there was a problem with the input lexicon we were using to train the recognizer. The transcription lexicon did not reflect several important phonological idiosyncracies of japanese, which happen not to be predictable. The best solution for us was to tag the phonemes that were susceptible to the idiosyncracies and then produce all possible variations, as this ensured that the correct model would get into the recognizer. Unfortunately nobody had the perl expertise to do this, but once we realized this we thought we'd like to find out what clever solution might be waiting out there...