共起語の特異性と繰り返しに着目した歌詞からの印象的フレーズ抽出

抄録

This paper proposes a method to detect attractive phrases from musical lyrics. The method considers not only acoustic but also lyrical features which attract human attention. Through the interview on impression for lyrics, we have found that “uniqueness of co-occurred terms” and “repetition” significantly influence human attraction. Therefore, we modeled these linguistic expressions as mathematical features. The proposed method detected attractive phrases using support vector machine with the modeled features. The results of the model evaluation experiments showed the 69% accuracy and 86% precision. Moreover, the results were compared and discussed with the key sentences detected by the existing summarization methods. As the result of comparison, it is confirmed that the proposed method detected attractive phrases more accurate than the existing summarization methods.