1 / 21

Lyrics, Music, and Emotions

Rada Mihalcea Carlo Strapparava University of North Texas FBK- irst EMNLP 2012. Lyrics, Music, and Emotions. A Corpus of Music and Lyrics Annotated for Emotions. Corpus 內包含 100 首有名的英文歌曲 , 檔案格式為 MIDI Hotel California by Eagles, Let it Be by The Beatles…

Download Presentation

Lyrics, Music, and Emotions

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. RadaMihalcea Carlo Strapparava University of North Texas FBK-irst EMNLP 2012 Lyrics, Music, and Emotions

  2. A Corpus of Music and LyricsAnnotated for Emotions • Corpus內包含100首有名的英文歌曲,檔案格式為MIDI • Hotel California by Eagles,Let it Be by The Beatles… • MIDI(Musical Instrument Digital Interface)是一個工業標準的電子通訊協定 • MIDI不傳送聲音,只傳送音調,音樂強度的資料,音量,顫音,相位等參數的控制訊號,以及設定節奏的時鐘信號並與歌詞同步 • 不使用整個MIDI檔案,只提取出需要的部分

  3. 在音樂的部分 在歌詞的部分 在音符的部分 G大調:GABCDEF#G B->B 一度 B->C 二度 B->D 三度 …

  4. A Corpus of Music and LyricsAnnotated for Emotions

  5. A Corpus of Music and LyricsAnnotated for Emotions • 歌曲情緒標記 • AmazonMechanicalTurk service • 標記的類別分為六類,分數介於0~10分 • ANGER, DISGUST, FEAR, JOY, SADNESS, SURPRISE • 標記者被要求 • 要以作詞家的角度來看,不是以自己的角度 • 能夠解釋歌詞的意義 • 每一行歌詞都要標記六個類別的分數

  6. A Corpus of Music and LyricsAnnotated for Emotions • 為了避免spamming影響標記的品質, 以下列兩個方法解決 • 在歌詞中加入假的歌詞 • 例如:”Please enter 7 for each of the six emotions” • 計算標記者與其他標記者間的Pearson correlation, 如果低於0.4就不使用

  7. A Corpus of Music and LyricsAnnotated for Emotions 每一首歌曲作10次標記 經過spamming移除後 每首歌剩下2~5個標記 整體標記的correlation 係數為0.73

  8. Experiments and Evaluation • Experiments分為三組 • textual features • musical features • textual和musicalfeatures • Evaluation • Gold standard和分類器預測之間的Pearson correlation • 實驗使用linear regression(Weka machine learning toolkit)和Ten-fold cross-validation執行

  9. Experiments and Evaluation - Feature • Textual Features - Unigramfeatures(bag of word) • 先建一個詞彙表包含training set內所有出現過的單字(包含stop word), 將次數少於10次的單字去除, 剩下的單字當作unigram features

  10. Experiments and Evaluation - Feature • Textual Features - Lexiconfeatures(semantic class) • 利用LIWC(Linguistic Inquiry and Word Count)和 WA(WordNet Affect) • LIWC:包含約2200個單字,70個與心理歷程有關的類別 • WA:利用wordnet內情緒詞的synset所建成

  11. Experiments and Evaluation - Feature • Musical Features - Notes • 音符是用來表示一個聲音的音高和長短,以前七個大寫英文字母表示 (G-A-B-C-D-E-F) • 在聲音的部分有升記號#和降記號♭,表示升半音或降半音 • 在長短的部分有全音符,八分音符… • Musical Features - Key • Key是用來表示一首歌曲所使用的和絃或者音高集合, 例如C-major, F#, C-minor

  12. Evaluation – Textual & Musical feature 效果較好

  13. Evaluation – Joint Textual & Musical

  14. Discussion • Textual features和Musical features雖然都有用的,但是Textual features的效果較好 • 在實驗結果中,效能提升最多的三個類別分別為JOY, SADNESS, ANGER • 前兩者的提升是因為corpus中, 這兩類的歌詞較多 • 但ANGER與前兩者相比,corpus中的歌詞相對少,卻出乎意料的提升很多

  15. Discussion-Feature ablation

  16. Discussion-Coarse-grained classification • 將原本的task轉換成binary classification • Support vector machine(SVM) • Threshold設定為3 • Ten-fold cross-validation • 正確率(accuracy)為10次cross-validation的平均 • Baseline • 每一次的cross-validation,計算Trainingdata內資料量最多的類別的正確率 • 10次正確率的平均值當作baseline

  17. Discussion-Coarse-grained classification

  18. Discussion-Comparison to previous work • 因為先前沒有類似的task,沒有辦法直接做比較 • 挑選對1000則新聞頭條作情緒分類的task來做比較(分成相同的六類)

  19. Discussion-Comparison to previous work

  20. Conclusion • textual features和musicalfeatures對於歌曲的情緒分類上是有用的,而兩者都使用的效能是最好的

More Related