Abstract

Lyrics parody swaps one set of words that accompany a melody with a new setof words, preserving the number of syllables per line and the rhyme scheme.Lyrics parody generation is a challenge for controllable text generation. Weshow how a specialized sampling procedure, combined with backward textgeneration with XLNet can produce parody lyrics that reliably meet the syllableand rhyme scheme constraints.We introduce the Weird AI Yankovic system andprovide a case study evaluation. We conclude with societal implications ofneural lyric parody generation.

 

Quick Read (beta)