Semantic dependency network for lyrics generation from melody.

Melody-conditioned lyrics generation aims to create novel lyrics based on the melodies by learning the relationship between lyrics and melodies, which is an attractive topic in the music field. However, two serious issues, called deficiency of inter-dependency between melody attributes and text degeneration, degrade the quality of the lyrics generation. To solve these issues, this paper proposes a new model called semantic dependency network with two key components: (i) N-gram CNN block is used to compress the information from the single melody attribute and extract the inter-dependency from the multiple melody attributes. (ii) In lyrics, unlikelihood training is exploited to mitigate the syllables mismatching and logic missing and keep the intra-syllable integrity and logic by learning semantic dependency. Extensive evaluation experiments on a large-scale dataset demonstrate that our model can generate higher quality and more harmonic lyrics from the melodies compared with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]