A Fuzzy Training Framework for Controllable Sequence-to-Sequence Generation

The generation of music lyrics by artificial intelligence (AI) is frequently modeled as a language-targeted sequence-to-sequence task. Formally, if we convert the melody into word sequence, can consider task to be machine translation Traditional tasks involve translating between cross-lingual sequences, whereas and natural language sequences. theme or key words generated are usually limited meet needs users when they generated. This requirement thought restricted problem. In this paper, propose fuzzy training framework that allows model simultaneously support both unrestricted adopting an additional auxiliary process without constraining decoding process. maintains benefits but greatly reduces extra time overhead constrained decoding, thus improving its practicality. experimental results show our well suited tasks, it also generate sequence under condition given multiple models, thereby achieving goal green AI.