數位圖書館中的音樂心情偵測之研究

Abstract

現今，我們的生活充斥著音樂，但是每當我們要運用它在特殊用途時，資源非常難取得。 舉例來說，若我們想要找浪漫的音樂來放在婚禮上；放鬆的音樂來治療病人，我們該去哪裡尋 找? 此外目前只有專家能進行音樂心情的判斷。 雖然許多專家依他們的專業知識建立像 Allmusic.com 的網站來標示哪些歌曲是屬於哪些心情的，但是目前並沒有一個有效率的分類音 樂心情系統。事實上，音樂無所不在，不只作曲家在作音樂，大自然也無時無刻的在作音樂。 我們該如何做才能應付這龐大的音樂資訊呢？ 本研究會以預測使用者聽完音樂後的心情為目標，使用Allmusic.com 已定義好的音樂心情 分類，從中擷取英文歌曲屬於此心情的英文歌詞。在斷詞切字、移除停用字和透過英文單字正 規化(Porter Stemming)的處理後，計算出每個詞彙的Log Likelihood Ratio 值，找出優秀的 特徵值作為分類依據，最後再放進LibSVM 分類建構出模型，預測使用者聽完新進音樂後可能呈 現的心情狀況。研究結果發現，單單只靠歌詞來預測的效果並沒有很好，正確率約在五成。此 外，我們也發現在歌詞心情的偵測中，開心、激烈和戲劇性心情較容易偵測到。可能原因是這 三個心情的用字明確；反觀荒謬、文學、冷嘲和愚笨的心情在偵測上有難度。

Nowadays, there is much music in our daily life, but it is hard to find suitable ones as we would like to make use of them. For example, if we try to find romantic music for wedding, relieved music for therapy, where we can find it? Moreover, music mood only is judged by human experts in present. Although these experts develop many digital music libraries websites such as Allmusic.com to store music by moods, there has not been any efficiency system to classify music mood in automatic classifier so far. In fact, music is everywhere, not only human produces them but also Mother Nature generates them as well. Therefore, how can we deal with such great amount of music information? In this thesis, we regard predicting music mood as our goal. With the digital music library - Allmusic.com already classified music mood, we extracted English song lyrics in it. We calculate every token’s Log Likelihood Ratio after tokenization, removing stop-words and porter stemming. Further, we select out feature tokens as our classifying rule. Finally, we put them into LibSVM to build model, which tries to predict music mood of new added in data. Findings show that the accuracy only depends on lyrics is not good enough, only 50% predicting correct. Besides, we find mood Happy, Fiery and Drama are easy to detect, it perhaps that their emotional terms are clear. On the other hand, mood Wry, Literate, Ironic and Silly are hard to detect.