Lyrics and Emotions: A Sentiment Analysis of Popular Songs in Taiwan

(1)

LYRICS AND EMOTIONS: A SENTIMENT ANALYSIS

1

Lyrics and Emotions: A

Sentiment Analysis of Popular Songs in Taiwan

Master Thesis for the MA Digital Humanities

In partial fulfilment of the requirements for the degree of

Master of Arts

Ian-Rung, Huang

University of Groningen

June 2019

Supervisor. Dr. G. Bouma

Second reader. Dr. T. Caselli

(2)

2

Abstract

The studies on popular music rarely probe into the sentiment analysis and the

distribution of various emotions in Chinese song lyrics. The best songs of the year in the corpus of Golden Melody Awards, which aims to promote popular Chinese music in Taiwan, was used for helping to fulfill the research gap. The research questions of this study are “Do the properties of lyrics (length, sentiment, etc.) change over time? and “Do the lyrics of awarded songs differ from songs that were just nominated?” Mood classification analysis used involves two parts: One is applying API to decide the general impression of the binary mood and tone sentimentstrength; the other is combining fuzzy clustering of dimensions and categories from Russell’s (1980) model and Hu et al.s’ (2009) model as the study framework for analyzing the 142 song lyrics. The data was manually labelled with the 16 emotion classes which belong to the four dimensions +V+A, +V-A, -V+A, and -V-A,[(negative, positive) valence and (inactive, active) arousal]. The statistical tools SPSS was used for organizing, and visualizing the outcome of the data analysis. The major findings are summarized as follows. First, the songs of Golden Melody Awards in the period 2006-2018 had a similar SentimentStrength as those in the period 1990-1996. The accuracy rate of the judgement of the tone for the songs is 80.3%. Second, there was a significant

difference in the word count and in the category earnest respectively between the two periods, 1990-1996 and 2006-2018. Third, in 1990-1996 there was a tendency of more negative tone, significantly more words, a larger number in all the dimensions +V+A, +V-A, –V+A, and -V-A, and more number of the category sad than those in 2006-2018. Fourth, awarded songs tended to have a little stronger sentimentstrength, positive tone, significantly more words, a larger number in the dimensions +V-A and –V+A, and more number of several categories and earnest than the nominated songs. According to the findings in the current study, some implications for cross-cultural comparison of song appreciation, the writing of songs, various tools for tone accuracy judgement, and selection of different categories for the four dimensions were

addressed. There are some suggestions provided for the future studies, including selecting a larger quantity of data set, various types of song lyrics, more accessors for the manual mood annotation and concept-level sentiment analysis.

Keywords: Chinese popular songs in Taiwan, Golden Melody Awards, music,

(3)

LYRICS AND EMOTIONS: A SENTIMENT ANALYSIS 3

Abstract 2

Table of Contents 3

List of Tables 5

List of Figures 6

1 - Introduction 7

Background and Motivation 7

Research Purpose and Questions 8

Significance of the Study 9

2 - Literature Review

10

Definition of Emotion 10

Benefits of Music on Emotion 10

The Importance of Studying Lyrics 14

Studies on Music in Taiwan 16

Popular Music in Taiwan 16

Popular Music Studies in Taiwan Setting 17

Crisis of Taiwan Popular Music and Golden Melody Awards 19

Studies Related to Sentiment Analysis 21

Features and Functions of Sentiment Analysis 21

Studies of Sentiment on Various Texts 24

Studies Related to Sentiment Analysis of Music and Lyrics 27

Instrument for Studying Music Emotion Response 27

Studies of Sentiment Classification of Music and Lyrics 28

3 - Methodology

36

Source-data 36

Data Collection Procedure 37

A. The Document Level with API 38

A-1: Example of Implementing the API 38

A-2: The Result of Using API for a Song in the Study 39

B. The Words of Emotion 40

(4)

4

B-2 Chinese Lyric for "This Life and the Afterlife" and Distribution of the

Emotion Words 41

B-3: The Analytical Process of the Emotion Words 42

C. Further Analysis of the Data 43

4 - Results

44

The Distribution of Tone Judgement 44

The Word Count in the Whole Period and the Two Types of Songs 47

Distribution of the Four Dimensions 48

Distribution of the Categories 50

5 - Discussion

56

The Accuracy of the Automatic Tool 56

The Distribution of Tone 57

Word Count in the Two Periods and the Two Kinds of Songs 57

Distribution of the Four Dimensions 58

Distribution of the Categories 59

6 - Conclusion

63

Summary of Major Findings 63

Implications for the Music Work 63

Suggestions for Further Studies 65

(5)

5

List of Tables

1 Mood Categories and Number of Positive Examples 40

2 Differences in Group Statistics between 1990-1996 and 2006-2018 and between Two Kinds of Songs 47

3 Three Tones in the Two Kinds of Songs 47

4 Four Dimensions of the Two Kinds of Songs 49

5 t-test of Word Count of the Songs 52

6 16 Categories in Two Kinds of Songs and Two Periods of Time 53

7 Four Dimensions of the Songs in the Two Periods of Time 54

8 Dimension +V+A of the Songs in the Two Periods of Time 54

9 Dimension +V-A of the Songs in the Two Periods of Time 54

10 Dimension -V+A of the Songs in the Two Periods of Time 55

11 Dimension -V-A of the Songs in the Two Periods of Time 55

(6)

6

List of Figures

1 Coding for Generating the Outcome of Each Tone and SentimentStrength of

the Chinese 38

2 The Result of Each Tone and SentimentStrength from the Chinese Song “I May Be Ugly but I Am Tender” 39

3 Coding for Generating the Outcome of Each Tone and SentimentStrength of the Chinese Song “This Life and the Afterlife” and its result 39

4 Russell’s Model of Mood 41

5 Ian Rung’s Created Model of Mood 41

6 Mean SentimentStrength in the Two Periods of Time 45

7 Mean SentimentStrength in the Two Kinds of Songs in the Two Periods 45

8 Three Tones in the Two Periods of Time 46

9 Three Tones in the Two Kinds of Songs in the Two Periods of Time 46

10 The Mean of the 4 Dimensions in the Two Periods of Time 49

11 The Mean of the 4 Dimensions in the Two Kinds of Songs in the Two Periods of Time 50

12 The Categories in the Two Periods of Time 51

13 The Categories in the Two Kinds of Songs in the Two Periods of Time 52

(7)

7

1- Introduction

In this chapter, several parts will be presented. At the beginning, the background and motivation are introduced. Second, the research purpose and question are

discussed. Finally, contribution of the study is provided.

Background and Motivation

Music provides people with several functions. For example, people listen to music to regulate mood, to achieve self-awareness, and to express social relatedness (Schäfer, Sedlmeier, Städtler & Huron, 2013). Modern technology has simplified the process for us to retrieve, broadcast, and view the songs stored in a digital repository or an application (Hu, 2009). In-depth discussion of the important component of the vocal songs: lyrics, began in an impactful music survey: Approximately 17% of male adolescents and 25% of female adolescents expressed that they liked their favorite songs specifically because the lyrics were a reflection of their feelings(Fuld,

Mulligan, Altmann, Brown, Christakis, Clarke-Pearson,., ... Steiner, 2009). Knobloch-Westerwick, Musto and Shaw (2006) have stated that although young listeners might not understand all the details in lyrics, they can recognize them easily to obtain a general idea of the lyrical message.

Music or songs were perceived in various studies with sentiment analysis. For example, popular artists of western music were analyzed to check the tone of lyrics. Singh (2019) found a majority of song contents (more than 80%) are neutral. Only a few artists tended to have negative lyrics. Besides, API was used to extract different data field from Taylor Swift’s songs with exploratory analysis and Preetish (2018) found the length of content in her songs increase gradually. As the music are mostly Western-style or a single artist, I was intrigued to explore whether it is a universal tendency.

Moreover, a mood category specialized in the tone can be tremendously broad. For instance, there are various positive mood classification of excitement, pleasure, sleepiness, relaxation, and tranquilization. However, under my observation, most automatic machine analyzers generally archived them all in the tone “positive”. It is an important criterion to use music mood to seek or organize music, but most music string places do not support access to music by mood (Hu, 2009). Hu (2009) pointed

(8)

8

out that it is challenging for music digital libraries to provide users with access to music naturally and diversely. Even though music related text, such as lyrics can act as an independent resource and to complement audio content, not many music recommender systems perform satisfactorily and most only provide audio content, they especially lack in the Chinese song lyrics with clear annotation. According to Hu, we need to develop a proper set of mood categories to reflect the reality of music listening, and can be well applied in the music digital library (MDL) and music information retrieval (MIR) community.

Most of the sentiment classification research dealt with text of opinions.

Recently a few studies which investigated music of mood in general explored English music or songs, comparatively less research explored Chinese music or songs. Among the few examples, they basically focused on the particular topic-love (Hsu, 2016; Yang, 2014) or discussed the time in several clustered periods, for example, 1949 to 2000, 1989 to 1998, 80’s and 90’s, 1990 to 2008, 2009 to 2012 (Chen, 2010, Su, 1999, Wu, 2005) or the special focus on discussing lyrics with synaesthesia or metaphor (Chang, 2008, Chen, 2010) and concluded with various types of representation.

The studies on popular music rarely look into the sentiment analysis and the distribution of various emotions in the song lyrics simultaneously. This is one motivation of my conducting the current study. In addition, no research focuses on recent popular music from 2013 to 2018. The corpus for this study is the songs of Golden Melody Awards for promoting excellent Chinese music in Taiwan, which can help fulfill this gap. This is another reason why this study focused on this particular corpus through its whole history. Fell and Sporleder (2014) found that lyrics

properties seem to change over longer periods. They believed that classification of song lyrics can benefit both music retrieval and recommendation and basic

musicology research by showing researchers interesting trends of mining lyrics corpora.

Research Purpose and Questions

The purpose of this study is look into the difference of emotion words by using English tags to decide the Chinese words of emotion in order to verify the general emotion tone generated by API. Specifically, we would examine whether the song

(9)

9

lyrics of the two periods of time, 1990-1996 and 2006-2018, showed any different distribution. Besides, we were intrigued to find whether the lyrics in two kinds of songs, awarded and nominated, revealed different construction patterns. Therefore, the following are the research questions of this study. First, do the properties of lyrics (length, sentiment, etc.) change over time? Second, do the lyrics of awarded songs differ from songs that were just nominated?

Significance of the Study

In this study, we attempted to discover the sentiment strength of the song lyrics as a whole and extract the semantic features that were used for classification of the songs into mood categories. We focused on a corpus of Chinese popular songs in Golden Melody Awards, for which lyrics were assigned with manual mood

annotation. Owing to time and labor, we only focused on the best song of the year in the corpus. Consequently, there was only 142 song lyrics gathered for analysis. We showed a simple way to extract a fuzzy clustering model for each mood category, a list of words with similar mood and found the most representative mood in the four parts. We have examined the special meaning related to the mood words in context and also compared the tendency in the lyrics with other literature results. The results would show various distribution of the song lyrics in two periods of time and different construction of two kinds of songs. Therefore the findings may be suitable for serving as Chinese teaching materials put in the Internet or materials for music retrieval systems.

(10)

10

2-Literature Review

In this chapter, literature related to this study will be reviewed and presented in three sections. The first section deals with relation of emotion and music; the second focuses on the importance of studying lyrics; the third explores studies on music in Taiwan; the fourth section probes into studies related to sentiment analysis. The last section is related to sentiment classification of music and lyrics

Relation of Emotion and Music

Definition of emotion. Emotion is defined in different ways. For example, emotion is considered as an internal discomfort to create strong mental feeling and physical change (Wang, 1995). Emotion is thought as subjective mental feeling with different individuals’ cognition and physical change; the response of emotion should include positive and negative feeling (He, 2007). Chang (2006) defined emotion as a dynamic state of stimulus for influencing individuals, including complex and

uncontrollable affective response and physical change. However, to Ellis (1979), emotion is cognitive-sensory states, which refers to the person’s ideas to the event, not the event or stimulus itself. In a review of the main meaning of typical emotion, Lin (1999) emphasized intentionality as the main root of human emotion. That is, perception of the subjective value sense is integrated into the intentionality, which forms the deep structure or the pre-model for perception and then affects personal interpretation of the situation or forms concurrent emotion. Therefore, emotion involves complex process and people need to learn to grow with emotion and express appropriate emotion in life (Hsu, 2012).

From the definition of emotion, both physical response and feedback can reflect personal cognition, which is complex. People may decide to modify emotion through their cognitive process. The perception of song lyrics can be treated as a kind of cognitive evaluation to stimulus and interpretation of the song lyrics is the emotion response. The study of this whole process is interesting and challenging because our study composes the interrelation process of cognition and perception.

Benefits of music on emotion. The issue of sentiment response to music has attracted research attention. Music has meaning even though music itself is not a language; by combining with lyrics or words in a song, music can reveal its inherent

(11)

11

meaning (Bisceglio, 2014; Yeoh, 1999). Music replaces and extends language and becomes a tool to affect people’s mental health, psychology and emotions (Trappe, 2012). In general there are four purposes of music: relaxing, concentrating,

energizing, and transitional. In the teaching of multiple intelligences to students, Campbell, Campbell and Dickinson (1999) suggested teachers play different music types to smooth students’ mood. These different types of music can affect people’s emotion, for example, fast rhyme, harmonious tone and raising melody can relax people and make them happy, but the opposite music may make people sad (Lai, 2004). You (2006) suggested people use different music to relieve negative emotion. Wu (2001) claimed that teaching music appreciation, teachers need to include cognition and affection, but it is not easy to express emotion, especially to children. Everyone has their favorite music type and different appreciation response. Music can meaningfully affect people’s sensitivity to others and improve emotional empathy and ability to get along with other people (Yeoh, 2016). Enjoyable and rewarding

experience of learning with music skills could positively affect personal and social development of children and young people (Hallam, 2010). Roberts (2017) found that world music lessons could promote students’ situational interest in terms of the four themes: fresh learning experience, physical engagement, related previous experiences, knowledge and real performance for real music. Ratings of familiarity, pleasantness and arousal could predict emotions associated with music, and the emotional state affected the focus of attention and subsequent discrimination (Ritossa & Rickard, 2004). Selection of appropriate and quality music was important in music education, but the research about the effects have not been conducted extensively (Droe, 2006). Affective responses of the fifth and sixth graders listening to music selection in music textbook were investigated with semantic differential scale (Hsu, 2012). Hsu (2012) found the students’ emotional response was quite consistent with the works’ intrinsic emotion and the strength of emotional response was high. Students have different responses to different works, have more familiarity with positive tone. Students’ emotion response and familiarity were affected by their music learning experience. Hsu tried to explore students’ emotion to music and help them appreciate music. In this study, Hsu has recruited 1015 students to listen to 12 songs from textbooks, among which almost all were western classic music without lyrics and had only two Chinese music. Therefore, students only used their instinct for judging the

(12)

12

music emotion. With students of different age, and pop music songs with lyrics, the perspectives of emotion appreciation should be different.

The notion of musical emotions is controversial (Juslin & Vastfjall, 2008). Juslin and Vastfjall investigated emotional responses to music and found out six underlying mechanisms beside the cognitive appraisal. The first is automatic brain stem reflexes to musical events; second, evaluative conditioning—evoke the same emotion as the original event; third, emotional contagion—perceiving and experiencing the same emotion; fourth, visual imagery is associated; fifth, episodic memory is evoked and the last is musical expectancy, which is fulfilled or violated. They differ in terms of information focus, ontogenetic development, key brain regions, cultural impact, induction speed, degree of volitional influence, modularity and dependence on musical structure. The inconsistent or non-interpretable findings often resulted from failed control of underlying mechanism (Juslin & Vastfjall, 2008). Juslin and

Vastfjall’s new paradigms pertaining to the study lie in that mechanisms not unique to music can also evoke emotions, guide future study and solve disagreement.

Evans and Schubert (2008) explored relationships between expressed and felt emotions in music because of disagreeing with the Gabrielsson model which claimed that it is often positive between the two loci of emotion; when listening to music one feels the emotion that the music expresses. They found 61% of the positive

relationship and the music pieces were preferred more than those showing non-positive relationships, which were affected by the internal and external locus of the music. When reading the lyrics, the process often includes the expressed part and the felt part of emotions, so the analysis is a very interesting and learning process

enlightening. We are interested in the perceived part from the emotion words expressed in the lyrics and music factors.

To enhance liking or preference for popular music, familiarity with the music is an important factor (Yeoh, 1999). Yeoh (1999) found that Malaysian undergraduates and teenagers preferred popular music the most as it was most frequently aired over the media in Malaysia and other countries and it was most familiar, especially to non-musicians. The popularity of music is related to its connection to emotions

(Swaminathan & Schellenberg, 2015). Swaminathan and Schellenberg (2015) claimed the acoustic cues contribute to the musical expression of music. General cultural cues can make listeners recognize musical emotions from other cultures with

(13)

13

modest accuracy, which is correct for basic emotions, but not for complex emotions. Emotions are perceived more strongly than experienced and perceived and felt emotions may also have different quality,

It is uncertain and debatable that emotional reactions to music can occur without cognitive appraisals (Swaminathan & Schellenberg, 2015). Swaminathan and

Schellenberg (2015) found that responses to emotions were accessed with two continuous dimensions, arousal (low to high) and valence (negative to positive), the same as the circumplex model, which shows that all emotions could be shown as points in two –dimensional (arousal by valence) space (Russell, 1980; Russell & Carroll, 1999). Combined affective responses to music were found to show support for the evaluative space model, which means that positive and negative feelings were independent and could appear simultaneously (Cacioppo, Gardner, & Berntson, 1997). Confusion between perceived and experienced emotions might affect self-reports of mixed emotions even though listeners were shown to distinguish mixed perceptions from mixed feelings (Hunter et al, 2010). They claimed that whether self-reports of mixed emotions are reflected in physiological responses can be researched in the future.

There were some psychological research related to assessing music emotion, but the inquiry was not well suitable for the research purpose. We need to consider the different music sources, expressed by music or induced by music as the affective response to music could be subjective to the environment, personal motivation and contexts of listening. In the study we focused on the perception of the emotion expressed in the song lyrics (Kim et al., 2010).

These articles are interesting as musical works, especially intrinsic emotion and familiarity, could directly affect the students’ emotional response. Because the music styles are mostly Western and there is no lyrics in them, we are curious to find out whether the lyrics of Chinese songs directly related to the emotion would evoke some types of emotional perception in this study. Since we are non-musicians, who like popular music, which contains similar quality like western music and meaningful lyrics, I would like to use popular song lyrics of our own cultural music as the corpus to study emotions. Therefore, this study can benefit my mental health as I am

(14)

14

Importance of Studying Lyrics

There were ubiquity and rapid grow of music quantity and there were several advantages for using lyrics to obtain natural genre clusters (Logan, Kositsky & Moreno, 2004). For example, lyrics transcription were available online, easy to collect, non-subjective, providing rich description of the song. Various methods could be used to analyze text semantically. Logan, Kositsky and Moreno (2004) used Probabilistic Latent Semantic Analysis (PLSA) to handle polysemy and synonymy problems. They found most frequent words in the genre respectively and top five words for various topics, but failed to distinguish various songs or artist styles. Similarity of artist styles based on lyrics was better than random but worse than acoustic similarity owing to the bias in the ground truth towards acoustic similarity. Incorrect errors proved that using a combination of two methods might be a better technique.

Recently the distribution of music has changed from the physical to online with the digital code, high quality and portable content. Music lovers can collect large music data which need to be managed and accessed efficiently (Cunningham et al., 2005). When attracted by the songs, many people will want to appreciate lyrics because people may like or dislike the song because of the lyrics (Cunningham et al., 2005). It was worthy of deeply exploring properties of lyrics when analyzing and classifying music. It is easier to withdraw and process lyrics than audio data,

especially for non-musicians because they often derived lyrics in the music retrieval system. Lyrics could provide semantic content, melodic and rhythmic properties of the audio signal as well as textual elements such as chorus, verse and bridge (Fell & Sporleder, 2014). Cunningham, Downie and Bainbridge (2005) modelled music information behavior and songs people dislike and found the most consistent and dominant factor for disliking songs is the song lyrics. It is assumed that a song with a more complex story or message requires more complex lyrics and might be rated more favorably.

Recently access to music contents has prompted a lot of studies. For example, researchers pointed out that structure detection could be done well by text-based methods and music often correlates with lyric structure. An experiment on using standard natural language processing tools was conducted to analyze song lyrics. Mahedero, Martinez, Cano, Koppenberger and Gouyon (2005) found several profits

(15)

15

of this NLP, such as identifying language, extracting structure, categorizing themes, and searching similarity. They claimed that song semantics could be decoded from lyrics and the analysis of lyrics combining with audio and cultural data was the basis for developing the whole music information retrieval systems.

In addition, music mood has become an important criterion in recent studies as music information behavior to seek and organize music (Hu, 2009). Hu pointed out the situation that it is necessary to develop a suitable set of mood categories to

represent the real music listening and adopted by the music digital library (MDL) and music information retrieval (MIR) community. In the modern time people can access and share information conveniently from the Internet and with computer technologies, they need automatic tools for classifying music and recommendation. Fell and

Sporleder (2014) believed that classification of song lyrics could benefit both music retrieval and recommendation and basic musicology research by showing researchers interesting trends of mining lyrics corpora. However, not many systems perform well and most only provide audio content. Music related text, such as lyrics can act as an independent resource and to complement audio content (Hu, 2009).

According to TechNews (2017), there are many kinds of music service platforms; it is easy to listen to music in daily life. Spotify, the biggest world music service company, has discussed the prospect of five music trends. Among them, the third one is ‘the music ranking is decided by oneself.” The listening behavior directly affects

the popularity of the music. Spotify believed that the audience really understand their own taste and their preference can be satisfied. Spotify predicted the future music market will be the time of streamed music activated by the music fans. Such analysis has provided an evolutionary vision about the future music and exploring the detailed lyrics can be an important work in order to really appreciate the musical essence. So we like to compare the past music with the popular songs lyrics in depth to find any differences.

In our study, we are doing a kind of music participation by analysis of emotion with the accessor’s emotion projected into the process. The reward and nominated popular songs in Golden Melody Awards would be great choices to understand the beauty of the linguistic aspects in the music as the popular songs have emotion words, which can be read and appreciated, not like music, which can only be heard and imagined. In the future, the collected content-based song lyrics may be deeply

(16)

16

mapped with extra musical features, such as audio, videos, radio play history, chart positions, and so on to the sentimental comments of song commenters, which may contribute to the improvement of MIR recommender systems and user-defined MIR filtering algorithms (Cunningham, Downie & Bainbridge, 2005). Similar to what Cunningham et al. claimed, our study aimed to have such contributions in the future.

Studies on Music in Taiwan

Popular music in Taiwan. Popular music is widely attractive music, typically distributed to large audience through the music industry. The forms and styles can be enjoyed and performed by people with little or no musical training. Through the mixture of musical genres, new popular music forms are created to reflect the ideals of a global culture… Scholars have classified music as "popular" based on various factors, including whether a song or piece becomes known to listeners mainly from hearing the music; its attraction to different listeners, its treatment as a

marketplace product in a capitalist context, and other factors (Wikipedia). Mandopop (2019) is a blended term of Mandarin popular music in order to describe Mandarin-language popular songs of that time. It is now used as a general term to describe popular songs performed in Mandarin.Long time ago, Chinese popular music use folk songs and ancient poetry to be inspiration to write lyrics; nowadays, the lyrics are direct, combining the features in the Western popular music and creation with various features (Wikipedia).

Chiuo (2015) pointed out four tendencies for popular music. The first is the added value of mobile video. The second is the surprising visual experience created by new high technology. The third is the fan browsing economy invigorated by social community application. The fourth one is the role relation among producers,

benefiters, and consumers at the new media time. Taiwan popular music is considered as the only energy of Chinese music platform (Dato, 2017). According to Dato, the history of Taiwan popular music was described as 60s’ 70s’, 80s, 90s’ and the 21st century multiple styles, the last of which included non-mainstream and independent music, which provided nice elements of rich Taiwan culture. I wonder whether the lyrics will represent a special emotion style during recent time. In this study the corpus of songs were divided into two periods: 1990-1996 and 2006-2018 for a comparison.

(17)

17

Different headings were used by Tingyi (2015) for the developmental history of Taiwan’s popular music. The first is Japanese period and after the war, when there were hybrid music of combining the Japanese melody with Taiwanese lyrics, even the Western Jazz melody, creating the unique style of Taiwan folk songs. Second, the national government period promoted Chinese and forbid Taiwanese music. After 1965 American soldiers stayed in Taiwan for Vietnam War, Western popular music were prevalent among knowledgeable people until later 70s’ when many kinds of songs, including patriotic songs were popular. Third, in Taiwanization, some singers promoted singing our own local songs and formed the popularity of campus folk songs. The period of abolishing martial law relieved the oppressed society and artists created alternative music and the new Taiwanese song activity, using new elements in music to express humanity meaning, raising the Taiwanese folk style. Fourth, later on the commercialization of popular music created the main stream of 80s’ and lead up the Chinese popular music. A lot of bands appeared at this time and independent music were developed into a new period with clear character and excellent non-mainstream music style, opening up a new page of Taiwan music creation. Piekeeper (2011) claimed that Taiwanese like singers with strong personal character and

unending creation. Now it is opened with the emphasis of local culture.

MyMusic (2018), the second big streamed music service platform company, announced a big survey of Taiwan popular music tendency in 2017, revealing most Taiwanese people choose music from their own data bank (65.9%), followed by ranking (61.9%), their own search (49%) and theme song list (45.7%). People show multiple choices of music and personalization of music listening style. One thing worth noticing in the survey is that the number of time of listening to Chinese music is double that to Western music and repeated for six times, which means that people appreciate Chinese music for much longer time and then they are assumed to enjoy or learn things from the songs. The exploration of the lyrics emotion is meaningful in terms of this aspect of music content.

Popular music studies in Taiwan setting. Researchers explored popular songs in Taiwan in various perspectives. For example, Lai (2004) examined 423 high school students listening to 10 clips of music of four types and chose their perception from 7 emotions. He found different music types for different emotions, that building musical circumstance and taste of music appreciation were effective on mending students’

(18)

18

disordered emotions. The main influence of music appreciation acceptance was the familiarity with music repertoire. Lai considered various types of music forms affecting relief of emotional stress, especially the current popular music easily bring free, easy and joyful feelings to students.

Lin (2005) explored elementary school upper-grade level classical music appreciation with implemented popular music elements and obtained the following findings. The contents and features of the popular music induced music curriculum matched the students’ preference and promoted their interests, participation and motivation. This curriculum offered students more adaptive, meaningful and interesting music learning experiences, and made students interested in learning, changed ideas about classical music, and increased learning in the cognitive and affective domain, but not psychomotor domain as it was just based on music

appreciation. Lin suggested teachers incorporate popular music into the curriculum, design various, interesting ordered musical activities, choose appropriate popular music to match teaching goals, gender and preference of the students. Lin also

suggested researchers investigate how lyrics of popular music can influence students. Students’ emotional intelligence could be improved by applying music

exploration activities. Kuo (2005) used EQ Scale and assignments in EG and CG groups and found that the EG students got higher grades than the CG students.

Students also improved various abilities, such as emotion recognition, expressing and balancing ability, social skill ability, and self-motivation and self-recognition. Both pop and classical music can improve students’ ability of identifying music element and transferring ability (Chen, 1998). The students prefer pop music to classic music but increase preference of classic music. Chen (1998) suggested teachers incorporate pop music into the junior high music curriculum, find suitable repertoire from the bulk of pop music and study the pop music in Taiwan.

Popular music was applied to improve the ability (competence) of sixth graders’ music aesthetic judgment, in terms of feelings, styles understanding, and value judgment of learning music (Sun, 2006). Sun (2006) found that the EG students have higher grades than the CG students. The studies suggested that integrating popular music into the classical music teaching might effectively improve students’ judgment of aesthetic musical merits. However, interaction and discussion between the teacher and students were important in the course.

(19)

19

Hsu (2008) also investigated the relation of 232 elementary school students’ music preference, emotion features and music perception by using measurement scales (Hsu, 2008) with positive and negative emotion features and eight emotion types to test students. She found there was a significant correlation between music preference, emotion features and music perception. Students had positive music preference, especially the third graders have high music preference; emotion traits were related to music emotions.

Eighth graders’ music preference of the Chinese popular songs were surveyed. Hsu (2010) found that rock and roll was the most favorite type of songs, melody was the most decisive reason for preferences, idols were their favorite singers, talent and singing techniques of singers decided their preference of singers’ features. Students’ preference were affected by their backgrounds. The investigation questions in Hsu’s study were limited and there were no questions about the construction of the lyrics themselves, or what emotion was contained in the songs or lyrics, which also were important factors to decide the audience’s preference of music.

Most of the music types in these studies are western classic music with no lyrics. From the review, we see popular music is interesting to attract students and affect their emotion. The studies on popular music seldom look into the sentiment analysis and the distribution of various emotions in the song lyrics simultaneously. This is another motivation toward exploring the current study.

Crisis of Taiwan popular music and Golden Melody Awards. The crisis of the pop music record industry has started since 1998 and Kuo (2003) explored the solution in terms of six areas: technology, law, market, industry structure,

organization structure and employers and looked at the four periods: 1970, 1980, 1990 and after 1998. Several factors might affect such changes. Kuo claimed that the music people could not only rely on passing singing songs to other people, they needed to be true to themselves, or the popular music would be stagnant. Lin considered that the music situation in Taiwan is disappointing. Lin (2016) claimed to respect the individual creator’s independent thoughts as the center of independent music.

Chong (2018) noted in the past, freedom, creativity and special political atmosphere have created the golden time of Chinese popular music in Taiwan, but currently the music industry in Taiwan is rooted out. Low ebb of popular music in Taiwan have appeared recently even though Golden Melody Awards, the so-called

(20)

20

Chinese Grammy has given awards for the highest performance of Chinese popular music for almost three decades. He thought that to make the institute to have the right to speak and set the norm for Chinese popular music was probably the last thing for Golden Melody Awards to do in Taiwan.

The international authoritative music magazine Billboard complimented Golden Melody Awards as an eye-catching event in Chinese music field and the only awards with creditability and Grammy for Chinese music (Lieng, 2017). Piekeeper (2011) claimed that Golden Melody Awards maintains high level of credibility and

evaluation, as it pays attention to the music essence and multiple development phrases and progress of music properties. Besides, the judges of Golden Melody Awards might come from the independent music circle and often encouraged the excellent music quality. Piekeeper (2011) considered sometimes the annual awards criteria might be different and the judges’ musical background might affect the results. Moreover, the media and the public pressure might affect the award results of the consequent year if the preceding year had unexpected awards. Golden Melody Awards is not only a rewarding ceremony, but also the target performance exchange platform and trial balloon for Chinese popular music (Chiuo, 2015).

Hsiong (2018) commented on 30 years of Golden Melody Awards and listed many merits it has embraced. First, it has created many target popular music singers at the time, whose songs could hypnotize the people’s life affection or stability, impulse or fashion, monitor our soul and sensitize the fashionable culture with visual and audio organs. There are some catching divine tunes in each year and behind the successful songs they contain romantic artistic character to summon the listeners’ longing and broken heart; the songs walk through our heart and affect the public. The elegance and aesthetic perspectives of Golden Melody Awards are the highest target for creating various types of music. Therefore, it is hoped to last forever and not become a legend. Music Finance (2017) claimed that Golden Melody Awards has become not only an award ceremony, but the biggest exchange platform of Chinese music. Every year new performers can get the awards as they have created new music styles. Nobody can predict the winners as no one knows the judges.

Another important thing pointed out by MyMusic (2018) is the gold plating of the Golden Melody Awards. After the announcement of the nominated song list, the listening rate increased to 50% and the awarded album grew1.6 times, and some

(21)

21

famous singers even reached 4 times of listening rate. It is obvious that the Golden Melody Awards is really influential, which contributes to one more attraction for doing the study with Golden Melody Awards. The vice manager of MyMusic, Huang H. L., claimed that she was happy to see more Chinese creation from new generation could match the modern language expression and return to the pure music essence. We can appreciate popular music in Taiwan in different perspectives, which is the important contribution of the event.

From the review of related literature in various areas, such as emotion, music, Taiwan popular music, and sentiment classification, I considered that conducting a study related to exploring the emotion in song lyrics in the corpus Golden Melody Awards is rewarding and fulfilling. I would apply the research methodsautomatic analysis and manual annotation in the studies to assist the exploring process and make the analysis more efficiently. I think we all need to contribute ourselves to revise this negative situation and attract more people’s attention to Taiwan popular music. Certainly I would want to make the song lyrics shine their light during the whole history of the existence of Golden Melody Awards. This is another important motivation of conducting this study.

Studies Related to Sentiment Analysis

Features and functions of Sentiment Analysis. Since 1950s there has been

natural language processing research in English language and about 20 years ago English sentiment analysis started to develop. Chinese sentiment classification has started only for ten years and the researchers need to use English research for reference (Peng, Cambria & Hussain, 2017).

Sentiment Analysis (SA) or Opinion Mining (OM) is the computational study of people’s opinions, attitudes and emotions toward an entity. The entity can represent individuals, events or topics. These topics are most likely to be covered by reviews (Medhat, Hassan, & Korashy, 2014). Sentiment Analysis can be considered as a classification process in the following sequence: Products Reviews, Sentiment Identification, Feature Selection, Sentiment Classification, and Sentiment Polarity. Sentiment analysis is more challenging than traditional fact-based analysis as it is used to treat opinions, sentiment and subjectivity in text, especially in dealing with opinions (Pang & Lee, 2008).

(22)

22

The methodology of sentiment classification can be implemented by either technical tools or manual annotation, which is a type of text research aka mining. It applies a mix of statistics, natural language processing (NLP), and machine learning to identify and extract subjective information from text files, for instance, a reviewer’s feelings, thoughts, judgments, or assessments about a particular topic, event, or a company and its activities as mentioned above (Altexsoft, 2018). Sentimental prediction identifies the text containing sentiment or opinion by deciding sentiment polarity as positive, negative or neutral and finally summarization module aggregates the results obtained from previous two steps (Zubair Asghar, Khan, Ahmad, Masud Kundi, 2014).

Archived by Montserrat-Howlett (2013), the shortlist of 10 practical tools were made for people to track user sentiment in English. Among them, Hootsuite,

Tweetstats, Facebook Insights, Pagelever and Social Mention aimed at managing and precisely measuring the social media networks. Besides, Meltwater, Google Alerts, People Browser, Google Analytics and Marketing Grader were designed for assessing the tone of commentary and track influencers, trends and competitors and are

powerful to monitor one’s efforts generated from marketing or business related works. The tools are convenient for researchers or company managers to adopt for

investigation.

Sentiment analysis is used to help users organize a number of comments and obtain better information content (Yang, 2011). It can help to classify articles in terms of positive and negative emotions. Comments often contain both subjective and objective facts and can cause mistakes in classification, especially movie reviews. Yang considered it important to decide the objective or subjective tone of the sentences. Yang focused on the subjective parts in the study and excluded the

narrative parts. Using the subjective analysis and PMI-IR, Yang could obtain the best accuracy from the first 2000 features, showing how opposition phrases affected the classification of movie comments. We could also focus on the sentences containing the emotion words and tried to find out the emotion distribution of the song lyrics after a comparison of all the emotion words.

As a kind of text mining techniques, Lee (2014) used word-of-mouth analysis and showed that the classification system was good for product features and opinions extracting words. Sentiment analysis and product analysis helped to understand

(23)

23

Internet consumer preferences, advantages and disadvantages of the product features, and understand the brand competitive edge through brand analysis. Companies could adjust their marketing strategies and made improvement on the products through the immediate market trend. The automated system analysis was convenient and useful to grasp consumers’ opinions. Emotion is frequently used to express friendships, social support or online argument, so Thelwall et al. (2010) tried to detect sentiment strength in short informal text, claiming that the analysis should be related to user behavior. Their study provided some accuracy of positive and negative features by

SentiStrength.

According to Symeonidis (2018), sentiment analysis was used to examine the problem of studied texts and it has numerous applications, such as posts and reviews, uploaded by users on microblogging platforms, news, forums, ad campaign, and electronic businesses, regarding the opinions they have about a product, service, event, person or idea, for understanding how people are responding to it. The

application of sentiment analysis on social media has been taken the dominant status. With the fast developing of World Wide Web applications, sentiment classification would have a huge opportunity to help people with automatic analysis of customers’ opinions from the web information. Automatic opinion mining will benefit both consumers and sellers (Ye, Shi & Li, 2006). If we look at numbers only, it can give us a false sense of hope that our content is generating leads for our brand or business. With sentiment analysis, we dig deeper and look at “quality metrics”, which include opinions, feelings, satisfaction ratings, the quality of shares, comments, re-tweets, replies, ratings or conversations, as well as the quality of engagement over time (Delahaye Paine, 2011). Therefore, the detailed information provided more beneficial reference.

In 1986, Whissell published the first edition of “the Dictionary of Affect in Language”. This is one of many acceptable ways to measure emotions, using a dictionary of research and meta-measurement frameworks, including conceptual definitions and considerations for the state, - feature difference. The "dictionary" contains more than 4,000 words, and each word in the dictionary is accompanied by a score of the emotional dimension of the evaluation and activation. In use, dictionaries can be applied to any language material, from freely generated text to text lists and document passages. The significant effects related to dictionary-based scores are

(24)

24

summarized, the reliability and effectiveness of the instrument are discussed, and some potential limitations are proposed. In short, the role of the dictionary and its potential for future development are considered. Nowadays we can easily get access to the Whissell’s dictionary in electronic version on the web page and use it as a reference when conducting a sentiment analysis.

Studies of Sentiment Analysis on various texts. A special domain of mining movie reviews was provided by Ye, Shi, and Li (2006) because “word semantics in a particular review could contradict with overall semantic direction (good or bad) of that review”. Ye, Shi, & Li divided two kinds of approaches for sentiment

classification: machine learning as well as semantic orientation (SO) methods. They claimed that no studies on using sentiment classification for Chinese review were reported. They tried to summarize the process of using SO approach for sentiment classification, but failed to apply to Chinese movie reviews due to two problems. These steps can serve as interesting reference steps for distinguishing Chinese text with sentiment classification.

In a study (Quan & Ren, 2009), Quan and Ren (2009) used a fine-grained

scheme to annotate various text emotion expressions in blogs in eight emotion classes (expectant, joy, love, surprise, anxiety, sorrow, angry, and hate). Quan and Ren analyzed the text in three levels: document, paragraph and sentence. Blogs often contain complete utterances and easy to be divided in the three fine ways. Songs contain different parts and various length. There were fewer studies on the valuable source of Chinese content with sentiment analysis (Zhang, Zeng, Li, Wang & Zuo, 2009). Their study used two levels: sentence and document level with three machine learning based approaches for two sets of Chinese articles. The method was proved to be effective and beneficial, but it remained unclear for other genres, such as whether the middle-size range of Chinese popular song lyrics in Taiwan in almost three decades display a similar tendency. Therefore, we decided to explore the general emotion of the whole song and examined the sentences with emotion.

In psychology, Russell’s (1980) emotion models was very popular in MIR research and it was a dimensional model where there were 28 emotions existing in a continuous multidimensional space. Using the circumplex model of affect, Hu, Downie and Ehmann (2010) investigated the usefulness of text features in music

(25)

25

mood classification on 18 mood categories were derived from user tags. The Positive and Negative Affect Schedule (PANAS) is a tool for assessing self-reported affect by category and proved all individual emotions (and associated labels) existed as

incidences of positive or negative affect, similar to valence. Positive or negative categories were also used to measure the emotion in this study (Kim, et al., 2010).

Fell and Sporleder (2014) performed a human annotation experiment to decide whether lyrics alone could not classify music genre well. They asked 11 participants to do song classification for genres and found that it was difficult for them to detect subtle stylistic properties and suggested that some genres were inherently more difficult to detect than others. Fell and Sporleder found that song quality seemed to affect lyrics quality in part. Type-token ratio and the length were the most two important for pop/rock music. In general, the best songs feature a higher type-token ratio, and told more stories than those which contain sex and violence. Fell and Sporleder claimed that “the musical style will change over shorter time spans reflecting for example change in taste regarding instrumentation and recording techniques” (p. 628). They also found that lyrics properties seemed to change over longer periods and that newer songs texts tended to be longer, older songs contained more repetitive structures, more well-matched between text blocks and had a higher chance of containing a chorus.

Asghar, Khan, Anmad and Kundi (2014) indicated that a product or service often contained some favorable or unfavorable features and users’ positive or negative opinions about them were usually interesting to refer. Besides text analysis, feature extraction in sentiment analysis is a popular research area recently. Asghar, Khan, Anmad and Kundi pointed out with feature based sentiment analysis, researchers needed to conduct feature extraction, sentiment prediction, sentiment classification and optional summarization modules. The features in the text could be extracted in various forms, such as lexical-syntactic or stylistic, syntactic and discourse based. The feature extraction is interesting and we would focus on lexical part.

Current sentiment analysis methods deal with the emotion tendency analysis (Wu, Lu and Zhuo, 2015). Sentiment analysis aims to explore user reviews and emotion polarity from texts or extends to natural language processing fields. The last one relied on lexicons and corpora, but was limited for sentiment analysis in Chinese language. Wu, Lu and Zhuo (2015) stated that owing to the difference between

(26)

26

English and Chinese, machine translation could not effectively classify Chinese text. They claimed that different language use resulted from users’ different habits, therefore, sentiment analysis needed to understand sentence structures. Besides, they considered “different individuals’ interpretations of these words influence the

environment, and the environment in turn influences their subsequent interpretations” (p. 2). This article about Chinese is influential.

Sentiment analysis was considered as a growing field between linguistics and computer science to decide the sentiment of positive/negative opinion, contained in text (Taboada, 2016). Machine learning is the most common approach and needs a lot of data set for training and learning the aspects and sentiments associated. It is a tendency for models to focus on a simple global classification of reviews, instead of rating individual aspects of the reviewed product. Not many of the methods could provide high accuracy; consequently, sentiment classification still needs a lot room to make improvement (Collomb, Costea, Joyeux, Hasan & Brunie, 2014).

Sentiment analysis research in Chinese language was reviewed by Peng,

Cambria and Hussain (2017), using the monolingual and multilingual points of view. The former involves polarity detection and the latter involves machine translation techniques. They looked into various sentiment corpora and lexica construction in different framework. They proposed a cognitive representation of Chinese concepts and interrelation to solve problems of limited resources, solving various research problems of NLP, which tended to be related emotion situations and was a promising direction for future research. The final goal of sentiment analysis is to point out sentiment or opinion labels of the targeted text. The problems involve classifying sentiment and identifying emotion/subjectivity. Peng, Cambria and Hussain reviewed the procedure and pointed out the basis of sentiment analysis lied in the corpora and lexica.There are two paths of conducting the research: machine learning based and knowledge based methods. They used the first one to classify sentiment in a binary way (positive or negative) preliminarily and treated the sentiment classification as a topic-based categorization problem. The sentiment was the topic and was treated as either positive or negative. The other way was using the knowledge based method to study language rules and syntactical or semantic relations in the sentiment lexicon words with sentiment. Polarity were used to label the sentiment of words in a text. The text polarity was counted by concluding each word polarity within syntactic

(27)

27

rules. It is rare to have Chinese sentiment resources, therefore it is important to create the Chinese sentiment resources in our study. We also used the traditional machine learning to manually develop features to identify sentiment from various domain knowledge.

Emotion traits and music perception had a prediction for music preference. This study is interesting in using positive and negative features to test eight emotion types of music, which are basic types of categories of emotion terms used for reference in our study. Different from the studies on the students’ direct response from listening to some clips of music from a few parts of music, it is intriguing to do the investigation on various song lyrics which contain various types of emotion expressions. There is no response of scales in our study, but the direct exploration from the lyrics text itself with direct perception and cognition involved in the process. The eight basic emotions were used in this study. There was in-depth demonstration of emotion from the lyrics in terms of the emotion words which all conveyed emotion between the perception and interaction of the assessors with the song lyrics.

Studies Related to Sentiment Analysis of Music and Lyrics

Instrument for studying music emotion response. There are many types of instrument for studying music emotion response, listed in this section.

Adjective checklist. Many scholars used different tools to understand the

underlying emotion for music. A popular one is called Adjective checklists, which include suitable adjectives for the theme. Hevner’s (1935) adjective clusters were revised and were used by 90 to test 133 college students with music background. Schubert (2003) used 67 adjectives and 23 adjectives from the circumplex theory of emotion (addressed in the later section) and obtained 9 dimensions of music from 46 adjectives, such as happy, light, serene, sentimental, sad, sacred, yearning, angry and exciting. These adjectives are basic emotion words for music, but they are too

complex. Asmus (2009) found that complex music is an important factor of affecting emotion response. We will limit the number of adjectives to simplify the selection process of parsing a lot of data.

Questionnaire. A questionnaire was conducted by You (2009) to examine 966

elementary school students listening to 20 world music and found different responses from different graders and various responses for different clusters of adjectives. You

(28)

28

(2009) claimed that using music of different speed would produce different response; slow speech would show sad and calm emotion; fast speech will show happy,

passionate, and nervous emotion; there were significant differences between grades, gender, and music learning experience. However, the self-statement might not be objective.

Semantic differential method. The semantic differential technique was

developed by Charles E. Osgood (Osgood, Tannenbaum, & Suci, 1957), to measure the connotations of words or concepts. Using factor analysis of large sets of semantic differential data, Osgood found three recurring attitudes that people used in assessing words and phrases: evaluation (such as “good/bad”), potency (such as “strong/weak”), and activity (such as “passive/active”).

The circumplex model of affect. Some research showed that mood could be

assessed by a continuum of descriptors or simple metrics of multi-dimensions. The circumplex model of affect means all affective states arise from two fundamental neurophysiological systems, one related to valence (a pleasure–displeasure continuum) and the other to arousal, or alertness. Russell (1980) set up a basis for organizing low-dimensional models, such as Valence-Arousal (V-A) space, where emotion intensity (Arousal) could exist from high to low and music polarity (Valence) could range from positive to negative. The music representation of two dimensions was supported to be valid. Russell found that 8 main circular emotion of the first two domains were sufficient to explain the emotion. These clusters of emotion were useful as they showed similarity of the emotions and could be simplified into some shorter lists for exploring the audience’s response to emotion words, especially when we only intended to focus on the written lyrics. It is easier and more efficient to merely center on a certain main emotions instead of many overlapping ones.

Studies of sentiment classification of music and lyrics. Evans and Schubert (2008) explored the relation between music expression and emotion and self-perception of 45 students, asking them to choose two familiar words with 11 scales for 4 domains: valence, arousal, dominance, and emotional strength. One group listened for 10 seconds and imagined the detail of the music, and the other listened to the whole music. They found the complex interrelation of music with the

environment, self-experience and memory.

(29)

29

pressure with vector space model (VSM), but not effective. Chen, Hsu, Chang and Luo expanded Thayer’s (1989) model of energy and stress and divided it into

energetic, calm, happy and anxious and generated four emotions: exuberance, anxiety, contentment, and depression. They found that using lyrics to mark listening to music was not effective, but they could get average distinction of music and lyrics when using lyrics to mark the answers. They thought modifiers might be added to judge the emotion to raise the effect of distinction. As the perception of song lyric emotion was subjective, using lyrics is the direct way to explore the emotion of the song writers.

The VSM was revised by Xia, Wang, Wong, and Xu (2008), who included only sentiment related words and invented a sentiment-VSM (s-VSM) approach. This approach could obtain the effective modifying words to strengthen or weaken the major emotion of the lyrics and decrease dimension of the features. Meyers (2007) explored music mood by using Lyricator system to give an emotional score for the lyrical content. Xia et al. (2008) used a fuzzy clustering method to decide the major emotion from a lyric. Grammatical information was used to weigh the clusters for individual sentences in terms of the factors such as tense and inter-sentence relation. The clusters with the largest overall weight were treated as the major song emotion and were classified by its mean V-A values into one of the four quadrants of the V-A model. They showed that their system outperformed the Lyricator system in different categories. We will use the clustering methods to put similar emotion words together into the same quadrants and count their frequency.

It is challenging for music digital libraries to provide users with access to music naturally and diversely. Not many systems perform satisfactorily and most only provide audio content, they especially lack in the Chinese song lyrics with clear annotation (Hu, 2009). Music related text, such as lyrics, can act as an independent resource and to complement audio content (Hu, 2009).

Researchers used different mood categories in the experiments of automatically classifying music mood, but it is difficult to make a comparison among them (Hu, 2009). Hu (2009) listed two methods to classify the music mood. One is to use Russell’s two dimensional arousal-valence mood model, which is based on

psychology, but it ignored the social context of listening to music. The other is to use mood categories from professional assigned mood labels on popular websites (e.g., AMG). The real life listening context represents real-life music information service,

(30)

30

but the professional mood labels may not reflect users’ viewpoints. Hu proposed to combine social tags, linguistic resources and human expertise to design mood

categories from social tags, adding that as human are the ultimate judges on labels in information retrieval tasks. We would also invite three human assessors to do the music label manually and choose music based on human judgment consensus.

Features from psycholinguistic resource, which contained affective lexicon translated from English words (ANEW) was extended to analyze each lyric sentence (Hu, Chen & Yang, 2009). Hu, Chen and Yang (2009) found that emotion units contained an emotion word in the lexicon and considered the influence of modifiers and tense on the emotion units. The fuzzy clustering method could decrease the error effect in the analyzing process of sentence emotions. The weight of each sentence could show the clustered emotion. The highest weight of the clustered emotion was treated as the major emotion of the lyrics. If we want to understand the emotion of songs, we can detect the lyric emotion.

Hu, Chen and Yang (2009) criticized one dimensional model of studies on emotion analysis of text as it was insufficient to show only positive-negative for emotions. They adopted Russell’s mood model which contained two dimensions, valence and arousal (with four symbols for the mood distribution in four quartered parts: –V+A, +V+A, -V-A, +V-A for 16 moods. They used Cheng’s method to detect sentence tenses with three categories: past, current and future. (Increase when u is after adversative or progressive words; decrease when u is before them.) The range of [-1.0, 1.0] is used to show the effect of the tenses as a modifying factor. In a lyric, an emotion unit after an adversative word affects the lyric emotion more strongly than before it. They found lyric sentences often belonged to several groups which had similar emotions and could be unified to a major lyric emotion, so they could remove the isolated sentences acting as noises. Their findings showed the small distribution of +V,-A, which conformed to the reality. The largest is +V,+A, then –V,-A, and –V,+A. Complicated and unusual sentence structures may cause some errors. They concluded that combining features seemed to be more effective. The findings of Hu, Chen and Yang showed the small distribution of +V-A, conformed to the reality. The largest is +V+A, then –V-A, and –V+A. They claimed that lyrics did not express more arousal dimension than valence. Some lyric emotions were implicitly expressed and needed to use knowledge and imagination to recognize them. We would use the similar way to

(31)

31

present the emotion words and examine the distribution of emotion words in Chinese lyrics.

Hu and Downie (2010) tried to use multi-modal mood classification to analyze features which could improve classification of individual mood categories. The dataset contained 5296 songs divided into 18 mood categories which were consistent with Russell’s mood model. They found that lyrics features were better than audio features in seven categories, but audio features were better in only one category. They claimed that in both MIR and music psychology there was much debate on music mood categories. Many used two to six mood categories, but the most popular one was Russell’s model, which contained two dimensions: valence (negative-positive) and arousal (inactive-active). This model had 28 emotion-denoting adjectives positioned in a continuous multidimensional space. Hu and Downie pointed out that most research classified music mood did not analyze or compare usefulness of the specific feature values. Text sentiment classification often used feature analysis. It is likely that different classification models would uncover various features. The previous research prompted this study to use feature ranking methods.

There were 20 years of audio message study of music emotion, but it failed to obtain correct emotion audio features and had no satisfying effect (Xia, Yang, Zhong & Liu, 2010). Xia, Yang, Zhong, and Liu (2010) proposed to use song lyrics to do sentiment analysis, which was not popularly seen at the time. They claimed that song emotions were often expressed by music, singing and lyrics. Originally they were afraid that singers might express the song emotions in a different way from the song lyrics, but actually only less than 5% of Chinese popular songs belonged to this type of situation. Therefore, they proposed to use natural language processing techniques to explore emotion in song lyrics. They used Thayer (1989) model of energy and stress and divided it into contentment, depression, anxious/frantic and exuberance. They said that energy was correct in the audio message evaluation, but stress was not obvious on it. Therefore, they only focused on stress in the lyrics and divided it into light-hearted and heavy-hearted. Xia et al. (2010) expanded emotion features into 12 kinds, including complexity and connotation. The design principles of s-VSM were to count the emotion with emotion words, to clarify ambiguity of the emotion words in context, and to consider the inversion, enhancement and decrease of emotion with the negation and modifiers. In the four areas, s-VSM performed better than VSM. They