以對抗式生成網路建立中文歌詞風格分類器

詞曲詩詞等文本生成類型的應用在近期越來越多人開始研究，然而關於這方面研究通常使用人力評估，這樣的做法難免會出現量化標準不一情形，因為當使用同一方法研究卻用不同人力評估時，結果很可能不盡相同，就算是同一批人的調查結果，根據調查時的各種變因都有可能會影響結果。本論文國內流行歌曲歌詞網中，幾位風格較具特色的作詞家之歌詞作品，將其作品資料使用SeqGAN進行訓練，並於訓練過程中，取出訓練程度不同的判別器，並將五個不同訓練次數(20, 40, 60, 80, 100)組合建立文本相似度評分器，以此作為歌詞與歌詞之間的評量方式。實驗結果顯示，詞風相似的作詞家，有較高的評分。

關鍵字

GAN ； SeqGAN ；文本生成；詞曲生成

並列摘要

The research of text generative network has been widely studied recently. However, the evaluation of this kind of research always adopts human assessment. In generally, the human assessment is determined by the people we choose, and the assessments may lack of fairness and objectivity. Furthermore, the same group of people may have different opinion under various conditions. Therefore, a fair lyrics discriminator merits our study. This thesis collects several representatively pop song writes’ works in Taiwan. The W. S. Fang’s lyrics are trained to build the discriminator in SeqGAN, which is a sequence generation framework of Generative Adversarial Nets (GAN). The discriminators trained under different epochs, which include 20, 40, 60, 80, and 100, to form the five-degree discriminator. Experimental results show that the proposed discriminator can efficiently distinguish different types of lyric writers’ works.