Rhyme Word Embedding and Attentionfor Korean Rap Lyrics Generation

Rap lyrics require not only the meaning of the lyrics, but also the composition of the rhyme. In order to generate Korean rap lyrics composed of rhymes, it is necessary to study lyrics generation considering Korean syllables. Korean syllable consists of two or three combination of consonant and vowel characters. In this respect, Korean lyric generation model has a specific structure different from English-based lyric generation model. Furthermore, for Korean rap lyric generation model, rhyme information should be included in its structure because rap lyrics includes rhyme information as well as context information. In this paper, a rap lyrics generation model using the embedding model combining these two informations was proposed. To implement the embedding model, each syllable of a Korean word was split into initial consonant, medial (vowel), and final consonant. Then, they were rearranged into characters and again re-organized to subwords by grouping characters including rhyme information. To learn a seq2seq type of lyrics generation model, an encoder-decoder model was designed and attention mechanism was used to capture the specific word pair with a rhyme relationship between the input and output sentences. Finally, we performed performance evaluation of the proposed rap lyrics generation model and confirmed the better performance than that of the conventional model using only context information.