基於Word2Vec詞向量的網路情緒文和流行音樂媒合方法之研究

Matching Internet Mood Essays with Pop-Music Using Word2Vec

我們觀察到很多人在網路上發表與心情有關的文章會附帶一首與文章內容有高度相關性的流行音樂。藉由流行音樂的歌詞內涵及音訊來傳達想要文章中要表達的理念及情緒。本研究藉由機器的運算能力，幫助在撰寫文章的人根據所撰寫文字的內容意義推薦出以代表該篇文章的流行音樂，讓閱讀文章的人可以根據文章中附帶的流行音樂來輔助了解到撰寫文章的人想要表達的內涵。我們使用了類神經網路語言模型訓練工具Word2Vec來實作上述的研究目標，並與基礎方法Boolean representation、TF-IDF以及Okapi BM25比較效能，我們使用亞洲最大音樂串流服務商KKBOX從2005年至2015年的每月的TOP-100華語音樂排行榜作為音樂資料集，結果顯示使用Word2Vec的推薦效果mAP 5值可達到0.3185。實驗數據顯示，顯示若將此系統建置成一般使用者在現實情況下使用的網站或是工具，高達81%的使用者可以在被推薦五首歌之內得到適合文章內容的流行音樂，顯示本研究若被應用在現實生活中，將會有不錯的表現。

Many people share their feeling or story by writing emotional article on the Internet. They also attach a pop music in their text usually. This pop music has high relation with the meaning of the story. As the passed research show that people share their feeling through music all the time. This research use powerful computation power of computers to help people choose music when they are writing emotional article. We use neural network language model tool word2vec to build our recommender system. We also compare the performance with three baseline method including Boolean representation, TF-IDF, Okapi BM25. We use Chinese TOP-100 popular music monthly rank since 2005 to 2015 from Asia's largest music streaming provider KKBOX as our music dataset. The experiment result scored 0.3185 with mAP@5. According to our experiment result. 81% of users can get the correct music they want before five music recommended. It will be a usable system if we build a website or application.