sheacon/song_lyrics
收藏资源简介:
Creation Steps - Downloaded [5 Million Song Dataset](https://www.kaggle.com/datasets/nikhilnayak123/5-million-song-lyrics-dataset) from Kaggle - Selected quality artists, as defined by me - Remove songs featuring any [profanity](https://github.com/surge-ai/profanity) - Added normalized version of lyrics (used for GloVe embedding only) - lower case, remove punctuation, remove stopwords, lemmatize) - Computed four sets of embeddings using all-MiniLM-L12-v2, all-distilroberta-v1, text-embedding-ada-002, and average_word_embeddings_glove.840B.300d
数据集创建步骤 - 从Kaggle平台下载[500万首歌曲歌词数据集(5 Million Song Dataset)](https://www.kaggle.com/datasets/nikhilnayak123/5-million-song-lyrics-dataset) - 按照自定义标准筛选优质艺术家 - 移除包含任何[冒犯性词汇(profanity)](https://github.com/surge-ai/profanity)的歌曲 - 新增歌词归一化版本(仅用于GloVe词嵌入),具体处理方式为小写转换、去除标点符号、移除停用词及词形还原 - 使用all-MiniLM-L12-v2、all-distilroberta-v1、text-embedding-ada-002以及average_word_embeddings_glove.840B.300d生成四组词嵌入向量
数据集创建步骤
- 数据来源:从Kaggle下载了5 Million Song Dataset。
- 艺术家筛选:根据个人标准选择了高质量的艺术家。
- 内容净化:移除了包含profanity的歌曲。
- 歌词标准化:
- 转换为小写
- 移除标点符号
- 移除停用词
- 进行词形还原
- 嵌入计算:使用以下四种模型计算了歌词的嵌入:
- all-MiniLM-L12-v2
- all-distilroberta-v1
- text-embedding-ada-002
- average_word_embeddings_glove.840B.300d




