以潛在語意為基礎運用模糊分群之歌詞情緒偵測

Abstract

音樂是一種傳遞情感的語言。作詞作曲家透過音樂傳遞某種特定的意念，而聆聽者也能藉由音樂來抒發自己的心情。因此在近幾年來，音樂情緒偵測逐漸成為音樂資訊檢索領域中所積極探討的議題之一。音樂情緒偵測主要可分為兩個方向，一是針對歌曲內容，如節奏、音高、曲調等來辨識歌曲意境；二則是利用歌詞進行文字分析，萃取出其特徵詞以偵測歌詞所隱含的情感。然而，歌詞在用字上因彈性較大，容易出現同詞異義或異詞同義的現象；加上情緒本身具有主觀及重複等特性，使得歌詞在音樂情緒偵測上的效果仍受到一定程度的限制。

本研究以語意詞彙進行特徵詞篩選，再利用潛在語意分析找出詞彙間的關連以顯現歌詞的真實意義，最後以Fuzzy C-Means將歌詞進行分群，來偵測歌詞所隱含的情緒。本研究先將ANEW與WordNet兩種語意詞彙結合後，從歌詞中萃取出足以代表整首歌曲情緒的特徵詞集，再以潛在語意中的奇異值分解找出詞彙與詞彙、歌曲與詞彙、以及歌曲與歌曲間的關聯，並透過低維近似找出每一首歌曲對應到低維向量空間時的座標，以此低維座標對歌曲進行模糊分群。根據實驗結果顯示，利用語意詞彙篩選特徵詞確實能提升情緒偵測的準確性；且將潛在語意應用在歌詞分析上，也較傳統向量空間模型來得可信許多。而相較於硬式分群，模糊分群法在歌詞情緒的偵測上也有較傑出的表現。最後本研究也發現，直覺的情緒類型如狂歡、激烈等較容易被偵測出來，反映出鮮明的情緒在選詞用字上通常也較為明確。

Music is often characterized as the language of emotions. Composers and lyricists convey their specific thought by music, and listener can express their feeling by music. For this reason, music emotion recognition becomes one of the most important issues of music information retrieval gradually. There are two main aspects about music emotion recognition. The one is the content of music, such as tempo, pitch, and tune. The other one is analysis of terms in lyrics, which can represent the mood implied by songs. However, the elasticity of selection of words through the lyrics is so high that it is hard to find the real meaning of words because of polysemy and synonymy. In addition, the characteristics of emotion, subjectivity and repetition, are the original limitation of music emotion recognition.

In this work, we propose an approach to lyrics emotion detection using fuzzy clustering based on latent semantic. First, we extract the features by the combination of two lexical dictionary, ANEW (Affective Norms for English Words) and WordNet, and then implement latent semantic analysis to discover the relation between terms and terms, terms and songs, and songs and songs. Finally, we put the matrix into fuzzy clustering to build model, which tries to detect music mood. According to the experiment result, the features picked by lexical dictionary can really improve the accuracy of music emotion detection, and the application of latent semantic on lyric analysis conforms to the reality more than the application of traditional vector space model. Moreover, we find fuzzy clustering has better representation than hard clustering in music emotion detection, and the intuitional emotion like party and fiery can be detected more easily because of the selection of words in these emotion are more precisely.