運用序列學習為基礎之中文歌詞產生系統

Abstract

隨著AI發展，人工智慧作曲日益興盛，在未來人工智慧作詞也應將於人工智慧音樂扮演重要角色，然而現今仍鮮少對於中文歌詞的研究，因此我們在這篇論文提出中文歌詞產生系統，也希望能為這塊領域吸引更多研究。本研究之中文產生系統使用序列學習，不同於常見的使用詞嵌入作輸入輸出，我們還導入了拼音向量，由於現下許多中文歌詞參雜英文，所以我們的拼音向量中文部分為注音，英文部分則轉換成國際音標。最後由實驗結果可以看到，同時使用拼音向量、詞向量和詞嵌入的序列模型可以產生相較於原本只使用詞嵌入的序列模型，明顯順暢的歌詞。

With the development of AI, composing music by AI keeps prospering. However, sometimes we struggle with the lyrics, which play an important role to form more delicate expressions. Unfortunately, there are limited researches on lyrics generation. Among these researches, there is only one for Chinese lyrics generation. Since Chinese is the second most used language in the world, there should be a huge demand for Chinese lyrics composition. In our work, we demonstrate how a Chinese lyrics generator based on Sequence to Sequence can support those who are out of ideas for writing Chinese lyrics. Other than typically used word embedding and word2vec, we also introduced a pronunciation vector in our model. Our pronunciation vector consists of Chinese Bopomofo phonetic spelling for Chinese words and International Phonetic Alphabet for English words appearing sometimes in Chinese compositions. Demonstrated by the experimental results, our model with word2vec and pronunciation vector can generate catchier lyrics with fluency.