1. Introduction

Popular music and culture have long been intertwined—more so in recent memory with social media acting as a vehicle for proliferation. Platforms like TikTok and Instagram have transformed how music circulates, with viral moments, dance challenges, and fifteen-second clips launching songs to the top of the charts overnight. A single trending sound can eclipse years of traditional promotion. The algorithm has become the new radio DJ, and shareability often matters as much as songwriting.

Has pop music, then, also adapted? Has it turned more popular than music—more about sound bites than artistic substance? One must ask whether contemporary songwriting has optimized itself for virality at the expense of depth, trading narrative arcs for the hook that fits in a story, the chorus engineered to loop endlessly in short-form content. Does pop music still codify culture in the way that we have come to expect it to?

While the role of popular music as a repository for cultural memory is well established in popular music scholarship, the quantifiable ways this memory is encoded in both lyrical content and sonic structure remain underexplored. Bringing digital humanities methods to the study of popular music offers new ways to analyze cultural trends at scale. This paper contributes to this field by proposing a multidimensional approach that triangulates lyrical sentiment, lexical change, and shifts in sonic texture. It presents a novel computational framework for analyzing longitudinal trends in popular music using a corpus of peer-presented award nominations, combining lyrical analysis with a quantitative study of musicological features.

2. Literature review

Songs—and implicitly, songwriting—have been part of the cultural milieu for millennia and have therefore come to be vessels of memory themselves, as discussed by José van Dijck in her book Remembering Songs Through Telling Stories: Pop Music as a Resource for Memory. Van Dijck references the ethnomusicologist Thomas Turino who, in turn, employs Charles Sanders Peirce’s framework of indexicality to assert that “music is not about feelings but rather involves signs of feeling and experience” (Van Dijck 2009, 108). Turino’s argument relies on the understanding that music functions differently from, say, an essay: it does not define or keep track of feelings and experiences, but rather has referential sonic or lyrical fingerprints which evoke them. Van Dijck adds to this by showing how music has intragenerational implications, using comments from a Top 2000 website to show that shared music platforms and listening experiences facilitate the “transfer of personal and collective heritage” (Van Dijck 2009, 111) and strengthen a shared social consciousness and cultural memory. Music is, therefore, more than just a vessel for artistic expression: it is an index of cultural significance.

Van Dijck explores how digital materials like the CD and the Walkman have contributed to music as a social activity by making it easier and more accessible, inevitably leading to individual memories “creat[ing] a context for reminiscence and becom[ing] vehicles for collective identity construction” (Van Dijck 2009, 114). This has only been compounded with the rise of the streaming era and social media; listening to music is more of a communal experience than ever before. Digital methods have not only advanced the way we disseminate and listen to music but also the way we deconstruct and examine it. This shift towards computational methods is evident in the realm of digital humanities, which combines traditional humanistic methodologies with digital tools to excavate new layers of meaning from long-explored topics. One of the biggest merits of the digital humanities method is that it allows researchers to make arguments of scale using self-collected or pre-existing corpora, as well as highlight and examine specific case studies. A survey of algorithmic methods involving songs spanning 50 years of research in the digital humanities finds that many of these corpora leverage the complex rhyme detection schemes of rap music to train algorithms (Brown et al. 2024, 32), which have developed from rule-based programs to more complex ones involving large-language models and natural language processing techniques used today (Brown et al. 2024, 22–29). Existing research also borrows techniques used in the assessment of poetry, such as those by Hernández-Lorenzo et al. (Hernández-Lorenzo et al. 2022), who examine Spanish lyrics using a technique called topic modelling to uncover popular themes, as well as more structural analysis such as that of song length, syllabification, and rhyme schemes. Similarly, Tang et al. (Tang et al. 2021) have applied quantitative methods on Chinese Huar folk songs, focusing on rhythm and sentence analysis alongside hierarchical clustering to objectively distinguish stylistic differences, demonstrating the power of digital humanities in objectively verifying drifts and patterns in style and theme.

This paper’s computational approach requires a representative corpus. One option is to use a chart-based dataset like the Billboard Hot 100, which offers a mathematical measure of commercial success. However, this paper uses Grammy Song of the Year (SotY) nominees, a peer-curated alternative. Unlike charts, which can be influenced by marketing strategies, streaming algorithms, and social media virality (Steiner 2025), the Grammy SotY nominees are selected through peer evaluation, offering a more curated measure of artistic recognition. This provides a unique longitudinal record of what experts in the music industry have deemed the pinnacle of songwriting craft. This approach is supported by existing research, including longitudinal analyses of Grammy categories by Negro et al. (Negro et al. 2022) and specific lyrical analyses to predict SotY winners by Musthyala et al. (Musthyala et al. 2024). A scholarly precedent therefore solidifies SotY nominations as an apt candidate.

This paper employs sentiment analysis to explore the drifts and patterns in this corpus, a method which academic sources note offers scalability and objectivity for identifying large-scale trends. This technique is well established for this purpose; studies by DeWall et al. (DeWall et al. 2011) and Napier and Shamir (Napier and Shamir 2018) have used sentiment analysis on popular song lyrics to argue that music serves as a “window into cultural shifts.” However, this method has significant limitations when applied to a poetic form like song lyrics. Standard models struggle to detect irony and sarcasm (Tan et al. 2023; Hercig and Lenc 2017) and often misinterpret figurative language, slang, and contextual ambiguity. Most importantly, lyrics alone cannot capture the multimodal dissonance—a key concept explored by Schaab and Kruspe (Schaab and Kruspe 2024)—that occurs when sad lyrics are paired with upbeat music. As Simon Frith (Frith 1986) famously argued, by virtue of being written for the purpose of singing and not reading, lyrics are inherently “performative utterances”; their meaning is completed by the sound of the voice and the music. This paper acknowledges these limitations by employing sentiment analysis not as a definitive measure of emotion, but as a tool for tracking large, quantifiable trends in expressed linguistic sentiment, which is then triangulated with the musicological analysis explained in the forthcoming sections.

3. Methodology

As mentioned, this corpus consists of nominees for the SotY category for the Grammy Awards. A multistage data collection pipeline was used to compile songs across seven decades. First, a base list of nominees (song title, artist, year) was generated by scraping the official Grammy website using the BeautifulSoup library, which parsed JSON objects embedded in the site’s code. Next, song titles and artists were cleaned using regular expressions and passed to the LyricGenius library to source verified lyrics from the Genius website, with the lyrics.ovh API as a backup. Concurrently, the song title and artist were used to query the Spotify Web API to retrieve a suite of audio features, including danceability, energy, valence, mode, acousticness, and instrumentalness (other features collected include key, loudness, speechiness, liveness, tempo, duration [in ms], and time signature). All scraping and API calls were parallelized using the Dask library. Finally, the compiled lyrics were manually verified for accuracy using ChatGPT to flag potential anomalies, which were then individually reviewed. The final verified corpus consists of 353 nominees from 1960 to 2025 inclusive.

With the corpus assembled, this paper then employs a multidimensional analytical framework. Songs were first preprocessed and tokenized to remove stop words and clean them for analysis. To gauge overarching emotional trends, lyrics for each song were then processed using a pretrained DistilBERT model fine-tuned on the Stanford Sentiment Treebank (SST-2) dataset, which returns a binary “positive” or “negative” label for each song.

To analyze lyrical themes, TF-IDF scores were first computed for all songs—a method that moves beyond simple word counts, which can overrepresent extremely common words and introduce distortions. By combining term frequency (how often a word appears in a song) with inverse document frequency (how rare the word is across all songs), TF-IDF emphasizes words that are both significant and distinctive in a given song. These scores were then averaged across all songs within a decade, producing a measure of which words were consistently prominent in that period. This averaging reduces the influence of words that may appear frequently in a single song but are not characteristic of the decade as a whole. The result is a set of terms that captures the recurring lyrical motifs of each decade, highlighting words that are truly representative rather than just momentarily or coincidentally frequent.

This framework also quantifies structural shifts in the songwriting craft itself. Using the textstat and language_tool Python libraries, lyrics were analyzed for repetition, adherence to conventional grammatical style, and readability—specifically using the Flesch Reading Ease and Dale-Chall scores. Although primarily used in educational contexts, these readability scores serve as a robust quantitative proxy for measuring lexical complexity, allowing this study to track shifts in the linguistic style and accessibility of songwriting.

The final analytical dimension directly addresses the performative nature of song, moving beyond the text to analyze its sonic landscape and delivery. As discussed, the multimodal dissonance of a song—which notes that its lyrical mood can intentionally contradict its musical feel—demonstrates the insufficiency of a purely lyrical approach. This quantitative musicological data is therefore incorporated to provide this crucial sonic context. By triangulating these distinct dimensions of lyrical sentiment, lexical structure, and musicological features, a holistic, quantifiable understanding of how songwriting changes and adapts—or at least something close to it—can be developed.

4. Sentiment analysis

The initial computational analysis involves running the DistilBERT sentiment model on the lyrics of all 353 nominees, grouped by decade. The results are visualized in Figure 1.

Figure 1: Bar plot of the sentiment analysis of songs in the corpus by decade from the 1960s to 2020s. The bars represent the count of positive (green) and negative (red) sentiments. The bars are annotated by the difference between positive and negative counts. The smaller the difference, the more “balanced” that decade is.

As the bar plot shows, the counts of positive (green) and negative (red) sentiments for each decade reveal a distinct and almost symmetrical pattern. From the 1960s, the count of positive sentiments increases, peaking in the 1980s, and then steadily decreases into the 2020s. The count of negative sentiments follows an inverse pattern, starting high, bottoming out in the 1980s, and then rising significantly in the 2010s and 2020s.

Two periods stand out as significant:

The 1980s: This decade shows a remarkably positive lean, with 21 more songs classified as positive than negative, the largest positive gap in the entire dataset.

The 2010s–2020s: These decades are the converse, showing a clear and consistent lean towards negative sentiment, with 12–14 songs being more negative.

4.2 Theme

Another way to frame this sentiment analysis is to use the prevalence of individual words as a proxy for theme to conduct a detailed thematic comparison of songs from the 1980s and the 2010s–2020s and understand how these sentiments are contextualized. To illustrate the themes, word clouds using the lyrics cleaned of stop words from each decade are presented in Figure 2.

Figure 2: Word clouds for cleaned lyrics by decade from the 1960s to 2020s. Emphasized themes are coloured red. The size of the word is representative of its TF-IDF score.

Of the words depicted in the figure, two are of special note: overt references to love and time showed up repeatedly in every decade, although increasingly less often. Love is perhaps not so surprising, but time is unexpected. The quantitative anomaly here necessitates a close reading to accompany the distant; as it turns out, songs in the 1980s and 2010s–2020s have very different attitudes towards time. When time is brought up in song lyrics in the 1980s, they are usually borrowing its property of perpetual recurrence. Take Dionne Warwick’s “That’s What Friends Are For” (Warwick 1985) as an example: “For good times and bad times / I’ll be on your side forevermore / That’s what friends are for.” Time looping back around is a comfort: whether it is good or bad will change, but the companionship will remain. The song incorporates time to act as a reminder to treasure life regardless of circumstances.

In the 2010s and 2020s, however, time appears to be fleeting. The relief that is ascribed to it in the 1980s is replaced by the dread and finality of the curtain coming down. Time is no longer cyclical; it is a dead end. This is visible in Lady Gaga and Bradley Cooper’s collaboration “Shallow” (Lady Gaga and Cooper 2018), whose lyrics are representative of a search for meaning: “In all the good times, I find myself longin’ / For change / And, in the bad times, I fear myself.” These lyrics are especially pertinent because they refer to the same good and bad times that Warwick does but shift the perspective from joyous to wistful and fearful. In a similar vein, while Harry Styles’s “As It Was” (Styles 2022) does not explicitly mention time, it speaks to the same despair when reminiscing of times past and too late to change: “In this world, it’s just us / You know it’s not the same as it was.” Styles’s song is lamenting, lonely, and evasive; he “[doesn’t] wanna talk about the way that it was.” In combination, these songs promote the overarching perspective that time has transformed from a restorative cycle of renewal into an irreversible linear progression characterized by loss, anxiety, and alienation.

What is evident from these analyses is that while the themes prevalent in songs across decades have not changed, there is a perspectival shift in how they are written into songs which correlates, as Batt pointed out, to the mood zeitgeist. Moreover, recent songs which incorporate these themes are less explicitly referential, which explains why the relative frequency of the words love and time diminishes in the word clouds in Figure 2. This analysis demonstrates a key methodological finding which, in turn, illustrates the advantage of a digital humanities approach: while the quantitative analysis successfully identified time as a stable keyword, it was blind to its semantic inversion. It is only by conflating the quantitative data with qualitative close reading that the full trend is revealed.

5. Structure and lexical field

These sentimental and thematic drifts are easily reflected in the listening experience and what you feel when you listen to different songs, but what is less manifest is the lexical and semantic shifts that mirror developments in language—that is, what you hear. The fingerprints of this structural transformation are better disguised; these changes must first make their way into the mainstream lexicon and colloquial parlance and, as a result, sound natural to hear. Therefore, they are more difficult to gauge by simply listening but may be detected under the microscopic lens of quantitative scrutiny.

To quantify, this analysis measures the prevalence of lyrical repetition. Repetition is a key compositional tool in popular music, deliberately employed by songwriters to function as a “hook” and enhance memorability. To track the changing prominence of this device, one simple approximation is to split the song lyrics by line and calculate the average count of repeated phrases for every decade. More specifically, because multiple phrases may repeat in a song, this experimental setup selects the highest count for every song and then averages across a decade to normalize by the number of songs. The results are presented in Figure 3.

Figure 3: Line plot of the average highest count of repeated phrases in a song by decade from the 1960s to 2020s.

Two things are interesting in this plot: the 1980s are, once again, an outlier with the peak value, and the average repeated phrase count has slightly decreased and then levelled out afterwards. In his article titled “Simply Irresistible,” Don Traut’s (Traut 2005) research on recurring accent patterns offers some explanation: the 1980s were a time when accent patterns were formalized as hooks into mainstream music. The hook, as aptly defined by Traut, is “the most memorable line or part of a song, the part that first comes to mind when one thinks of [it]” (Traut 2005, 57). Like Traut, who claims that accent patterns often find themselves becoming hooks by virtue of the repeated delivery of the title, or a slightly longer phrase containing the title (Traut 2005, 66–67), this paper treats the repetition of a phrase as a proxy for the hook of the song. One might think of a fascinating contemporary example: “that’s that me espresso” (Carpenter 2024).

Linguist Adam Aleksic (Aleksic 2024) recently examined the influence of algorithmic playlisting on contemporary linguistic evolution. He uses “Espresso” as an example, citing how Spotify’s algorithm pushing the music into everyone’s recommendations has left a linguistic footprint by way of a rise in the use of the grammatical construction “that’s that me” from the hook of the song on social media (Aleksic 2024). In conjunction with the knowledge that there is an increasing rise in the count of the hook phrase from Figure 3, it is fair to say that the songwriting process is increasingly more focused on one particular lyric or set of lyrics. In other words, the linguistic architecture of the song is laid on the foundation of hook.

What does this observation signify? Are songs perhaps valuing a tongue-in-cheek, viral moment more than the storytelling? In particular, because “that’s that me” is a technically incorrect yet wildly popular neologism, what does that say about the utility (or lack thereof) of time-honoured structural and linguistic rules—like those of grammar—in song lyrics? Fortunately, the language tool library in Python can measure this quantitatively. Figure 4 shows the average number of grammatical errors in song lyrics from the 1960s to the 2020s.

Figure 4: Line plot of the average number of grammatical errors in song lyrics by decade from the 1960s to 2020s.

This plot clearly shows an increase in grammatical errors, which numerically hints at artists using poetic license to heighten the appeal of silly, memorable pop earworms over, perhaps, more narrative-driven expository songs. This trend is exemplified by neologisms like “me espresso” (Carpenter 2024) and Chappell Roan’s “Femininomenon.” The viral success of such terms suggests a potential feedback loop between lyrical invention and online shareability, propelling both the song and the new term into the mainstream lexicon. Figure 5 uses the Python textstat library to plot metrics such as the readability and complexity of song lyrics to further see how the language has changed. Specifically, this paper uses the Dale-Chall readability score and Flesch Reading Ease—which are metrics typically used to assess the complexity of a text for pedagogical purposes. Figure 5 shows an increasing Dale-Chall readability score, which signifies that song lyrics are increasingly using words which are less commonly used in the common lexicon. It also shows a decreasing Flesch Reading Ease, which connotes the increasing structural complexity of the lyrics—individual sentences may be longer, or more syllables may be packed into a sentence. Combined, these aforementioned results speak to how the structure and style have changed to continue to appeal to audiences in parallel with culturally and temporally driven shifts in language.

Figure 5: Line plots of the average Dale-Chall readability score and Flesch Reading Ease by decade. The Dale-Chall metric estimates reading grade level based on whether words appear in a list of the 3,000 most common English words, while Flesch Reading Ease assigns higher scores to more readable text (with no lower bound and a maximum of 121.22). Both measures were calculated using the textstat library.

6. Sonic landscape

There remains one last dimension to explore—that of the sonic texture of songs. This additional context is critical, as it is “frequently the first aspect to attract (or repel) a listener” according to musicologist Allan F. Moore in Song Means: Analysing and Interpreting Recorded Popular Song (Moore 2012, 29). In popular music, this texture is not merely bells and whistles; things like timbre and loudness—what Moore refers to as “secondary domains”—are important features in their own right (Moore 2012, 29–30).

Figure 6 depicts systematic changes in the sonic texture of popular music. The distribution shapes—the widths, multimodality, and tails of the violins—capture signals beyond simple averages. Production-oriented and rhythmic features like loudness, energy, and danceability all move in tandem, showing not just increasing medians but significant mode consolidation at higher values in recent decades. This convergence suggests a large fraction of popular music has become more homogeneous, engineered for a high-energy, high-loudness, and rhythmically immediate soundscape. Concurrently, the collapse of acousticness distributions towards zero confirms a major shift in what Moore terms the harmonic filler layer (Moore 2012, 38), moving away from acoustic instrumentation and towards dominant electronic and electric production.

Figure 6: Violin plots of five key audio features (Danceability, Energy, Loudness, Speechiness, and Acousticness) for the dataset, grouped by decade from the 1960s to 2020s.

While some features converge, the violin shapes also reveal stylistic diversification. Speechiness, in particular, displays much fatter right-hand tails and broader distributions in later decades, indicating a growing subpopulation of tracks utilizing rap, spoken, or talk-sung content. This broadens the spectrum of popular vocal delivery and prioritizes the melodic layer over the others in the typical “four textural layers” of a song (Moore 2012, 37).

Figure 7 visualizes what Allan F. Moore refers to as the “‘feel’” of a recording (Moore 2012, 29). This chart plots energy (intensity and activity of a song) against valence (the musical “positiveness”) to map the emotional feel of songs by decade. We can divide this chart into four quadrants: top-right for happy, excited, anthemic songs; top-left for tense, angry, or dark songs; bottom-left for sad, sombre songs; and bottom-right for calm, serene songs.

Figure 7: Scatter plot of energy versus valence for the song dataset from the 1960s to the 2020s, colour-coded by decade. Dashed lines indicate the mean value for each axis.

Looking at the distribution, there is a clear migration over time. The 1960s (dark purple) have a significant number of songs in the two lower-energy quadrants. As the decades progress, the dots seem to float upwards: the 2010s and 2020s (light green and yellow) are heavily concentrated in the upper half of the chart. The takeaway here is clear: while the emotional mood (valence) of songs still varies from positive to negative, the energy with which that mood is delivered has overwhelmingly increased. Modern popular songs are far more likely to be high energy, whether they are happy or sad, delivering their message with what Moore would call a “thicker” and less “natural” sonic palette than in previous decades (Moore 2012, 46–49) and solidifying Schaab and Kruspe’s construct of multimodal dissonance.

7. Conclusion

The results of this trifold study thereby confirm the thesis of this paper: that popular songs function as poignant and evolving artifacts of cultural memory. While its main contribution is a methodological approach towards distilling songs and songwriting for markers of information that they are imbued with, the paper also attests the value of a digital humanities perspective. By applying a triangulated framework to seven decades of Grammy SotY nominees, it charts a clear affective shift in songwriting: the peak lyrical optimism of the 1980s inverts into a pronounced and consistent pessimism in the 2010s and 2020s. This emotional turn is complemented by quantifiable adaptations in craft and a structural trend towards greater lexical complexity and growing disregard for formal grammatical conventions.

Furthermore, these lyrical and structural changes are emboldened by a shifting sonic landscape. This paper identifies a clear migration towards high-energy, high-loudness, and electronically produced soundscapes. This convergence has normalized “multimodal dissonance,” where lyrically sombre or anxious themes are frequently packaged in high-energy, danceable productions—a technique that arguably captures the complex “mood zeitgeist” of the contemporary era.

This paper campaigns for a mixed-methods approach—future works may relax its reliance on an elementary, binary sentiment model by employing multi-aspect sentiment analysis and/or large language models (LLMs) to develop a more flexible and sophisticated architecture. Further scholarship may also conflate other musicological elements with techniques for lexical analysis to build a more granular understanding of the multimodal shifts this paper has introduced. Such work would build on this paper’s foundation, further demystifying the intricate and timeless craft of songwriting and its enduring power to capture and reflect the feelings of a generation.

Competing interests

The author has no competing interests to declare.

Contributions