Cultural evolution of emotional expression in 50years of song lyrics

Charlotte O Brand,  Alberto Acerbi and Alex Mesoudi

Abstract
The cultural dynamics of music has recently become a popular avenue of research in the field of cultural evolution, reflecting a growing interest in art and popular culture more generally. Just as biologists seek to explain population-level trends in genetic evolution in terms of micro-evolutionary processes such as selection, drift and migration, cultural evolutionists have sought to explain population-level cultural phenomena in terms of underlying social, psychological and demographic factors. Primary amongst these factors are learning biases, describing how cultural items are socially transmitted from person to person. As big datasets become more openly available and workable, and statistical modelling techniques become more powerful, efficient and user-friendly, describing population-level dynamics in terms of simple, individual-level learning biases is becoming more feasible. Here we test for the presence of learning biases in two large datasets of popular song lyrics dating from 1965-2015 and 1965-2010, where we study the trends in emotional expression. We find an overall increase in emotionally negative lyrics and a decrease of positive ones. Our results provide some evidence of content bias, prestige bias and success bias in the proliferation of negative lyrics and decline of positive ones, and suggest that negative expression of emotions in music, and perhaps art generally, provides an avenue for people to not only process and express their own negative emotions, but also benefit from the knowledge that prestigious others experience similarly negative emotions as they do. 
Keywords: cultural evolution; emotions; learning biases; song lyrics

Introduction

Are the lyrics of contemporary pop songs happier or sadder than the lyrics of the popular songs of previous generations? Do current songs mention love more or less than they used to? Are the pop charts angrier now than in the past, or have they mellowed over time? Millions of people buy and listen to popular music every day (by ‘popular music’ or ‘pop music’, we mean music with wide appeal that is typically distributed to large audiences through the music industry, and not just the specific ‘pop’ genre), and their song choices offer a window into their emotional states and psychological preferences. Changing trends in pop music over time may therefore offer a means of measuring large-scale societal changes. In recent years, the availability of large datasets in electronic format has allowed long-term, population-level cultural dynamics to be identified in an increasingly precise, quantitative way (Michel et al. 2011). This, in turn, permits researchers to test hypotheses about cultural trends that previously could only be assessed informally by focusing on a small number of (potentially cherry-picked) cases.

A fruitful area of investigation concerns the analysis of emotions in human cultural expressions. Several tools have been developed to extract the emotional content of texts, known as ‘sentiment analysis’.  Some of these provide a classification of how words score on ‘positive’ and ‘negative’ content(Baccianella et al. 2010), and others provide additional scores for specific emotions (e.g. how ‘angry’ or ‘sad’ is a text; Pennebaker et al. 2007). In most cases, sentiment analysis has been applied on a short-term time scale, such as social media interactions (Lansdall-Welfare et al. 2012). However, some researchers have explored a longer time scale, analysing the expression of emotions in several decades of song lyrics (Dodds and Danforth 2010), newspaper articles (Iliev et al. 2016), in Grimm’s folktales (Mohammad 2013) or in centuries of literary works (Acerbi et al. 2013).

The quantitative description of trends is fundamental, but a further necessary step is to understand what drives these trends. Cultural evolution theory (Boyd and Richerson, 1985; Cavalli-Sforza and Feldman, 1981; Mesoudi 2011) provides a series of concepts that allows such an endeavour. Drawing a parallel with genetic evolution, this field focuses on how cultural variation is transmitted from person to person via social learning processes such as imitation, and the processes that change that transmitted variation over time. In particular, cultural evolutionists have focused on transmission or learning biases as key drivers of cultural evolutionary dynamics (R. L. Kendal et al. 2018; Rendell et al. 2010). Transmission biases are heuristics that individuals use to decide what, when and from whom to copy. They are rule-of-thumb principles such as ‘copy the majority’, ‘copy the successful’or ‘copy the prestigious’ that allow individuals to adaptively learn from others (Laland 2004). Importantly, different transmission biases give rise to different population-level cultural dynamics. A cultural trait introduced in a population in which, for example, individuals copy mostly from their parents will spread slower than the same cultural trait introduced in a population in which individuals copy mostly from a few successful or prestigious individuals (Cavalli-Sforza and Feldman, 1981).

‘Model-based’ biases describe from whom people learn: for example, a success bias describes a tendency to learn from successful others, and a prestige bias a tendency to learn from prestigious(high status, respected) others. ‘Content-based’ biases describe what kind of information people learn best, owing to its salience or memorability. For example, a bias to transmit emotionally salient content, or negative content, has been found in several laboratory experiments (Bebbington et al. 2017; Fessler et al. 2014; Stubbersfield et al. 2015, 2017). These biases can be compared with unbiased transmission, in which cultural variants are transmitted in equal proportion to their existing frequency in the population. While there are many theoretical models that examine the conditions under which different transmission biases are adaptive and their expected population-level consequences (J. Kendal et al. 2009; Rendell et al. 2010), and experiments which have tested the predictions of these models in controlled laboratory set-ups (Caldwell and Millen 2008; Mesoudi 2008; Mesoudi and O’Brien 2008; Morgan et al. 2012), less research has explored how cultural transmission biases may impact real-life cultural dynamics (although see Acerbi and Bentley 2014; Beheim et al. 2014; Miu et al. 2018).

In this study, we test the extent to which transmission biases can explain trends in the emotional content of two datasets of English language song lyrics. The first dataset ('billboard') contains the lyrics of the songs included in the annual Billboard Hot 100 from 1965 to 2015, a widely known US chart (n = 4913 songs). The second dataset ('mxm') contains the lyrics of the English language songs present in the musixmatch.com website, the world’s largest lyrics platform where users can search and share lyrics, from 1965 to 2010 (n = 159,015 songs).

Figure 1. Proportion of the term ‘love’ (left panel) and ‘hate’ (right panel) in all song lyrics by year for the dataset billboard which contains the lyrics of the songs included in the annual US Billboard Hot 100 (n = 4913 songs). The proportions here are small as we are reporting the proportion of the word out of the total number of words in 100 songs each year (on average 30,000 words, i.e. 300 words/song) and on different scales (the frequency of positive emotion words is usually higher than the frequency of negative emotion words). To have an intuitive idea of the change, from 1965 to 1990, in the top-100 billboard songs, the word ‘hate’ was used each year around four or five times overall (30,000*0.00015), whereas now the average is around 24 (30,000 × 0.0008).

Preliminary analyses found a substantial decrease in the use of positive emotion-related words coupled with an increase in the use of negative emotion-related words in both datasets. Specific emotion-related words show considerable change in use during the time frame considered. For example, use of the term 'love' more than halved in both datasets, whereas the term ‘hate’ increased in frequency substantially (see Figure 1 for the ‘billboard’ dataset. The trends are qualitatively similar in the mxm dataset). These results are broadly consistent with previous analyses of song lyrics (DeWall et al. 2011; Dodds and Danforth 2010) and literary fiction (Morin and Acerbi 2017), suggesting a general cultural or artistic trend for increasingly negatively valenced emotional expressions.

The main goal of this study is to test hypotheses about possible drivers of these two trends. Before testing for the aforementioned transmission biases, we first checked whether linguistic effects could explain the patterns. We considered (a) a possible increase in slang words, (b) an asymmetric semantic change (for which words denoting negative emotions had acquired positive or neutral connotation, e.g. ‘wicked’, but not vice versa) and (c) a general increase in lyric complexity (although note the latter would predict that both negative- and positive-emotion words would decrease in frequency). After finding that the trend persisted after controlling for these linguistic effects (see Supplementary Material), we then examined whether cultural transmission biases might explain the patterns.

We considered, in a fully preregistered study, three hypotheses derived from cultural evolution theory as outlined above:

(H1) Success bias: the emotional trends result from artists copying the best-selling songs from the preceding years.

(H2) Prestige bias: the emotional trends result from artists copying the songs of ‘prestigious’ artists (independently of the success of the songs) from the preceding years.

(H3) Content bias: there is a general psychological preference for lyrics that reflect negative emotions in songs, thus songs with more negative content rank higher in the charts.

Note that we applied H1 and H2 to both positive and negative content, while H3 was applied only to negative content. This was because, although there is ample evidence for a content bias towards negative emotion, as referenced above, we are not aware of any evidence or theory that would predict a content bias for positive emotions.

After running our preregistered analyses to test these three preregistered hypotheses and depositing a preprint, we received the useful suggestion to also include an additional variable to control for unbiased transmission. Including this additional variable affected the interpretation of our results. Hence below we report both the preregistered analyses without unbiased transmission and the revised analyses including unbiased transmission. The use of unbiased transmission here assumes that the emotional trends of lyrics result from artists randomly copying the lyrics of any of the available songs in the preceding years, without taking into account chart rank or artist prestige.

Our results are mixed. We found a small effect of success bias on the likelihood of a word being positive in the billboard dataset, as well as a small effect of both prestige and content bias on the likelihood of a word being negative in the billboard dataset. However, these effects vastly reduced or became non-existent in the mxm dataset. Moreover, when controlling for unbiased transmission, the effects of success and prestige largely disappeared in both datasets. We therefore conclude that content bias may play a role in the likelihood of using negative emotion words in song lyrics, but that success and prestige biases (as we have implemented them) are not strong enough to explain the trends compared with unbiased transmission. We discuss possible explanations for the apparent content bias effect, as well as our interpretation and implementation of success, prestige and unbiased transmission in relation to the content of song lyrics.