Spotify Top 200 · week of 2026-03-05 · 8 regions · lyrics translated & emotion-scored

The Same Ache, Told Differently

Do song lyrics in different regions emphasize different emotional and thematic patterns? A full pipeline from ingestion to NLP classification, region by region.

We pulled the Spotify Top 200 for eight regional charts — Argentina, Colombia, Japan, Singapore, Spain, Taiwan, the USA, and the Global 200 — for the week of March 5, 2026, fetched their lyrics, translated everything to English, and ran each song through an emotion classifier tuned for lyrics: love, longing, joy, heartbreak, grief, despair, hope, lonely, sensual, anger.

The first result looked like a bug.

ARGENTINA → longing COLOMBIA → longing SPAIN → longing JAPAN → longing SINGAPORE → longing TAIWAN → longing USA → longing GLOBAL → longing

01 — The Illusion

Longing wins everywhere. That's the tell, not the finding.

532 of 1,471 songs — 36% — have longing as their single dominant emotion, more than the next two labels (sensual, love) combined. It's the top raw score in every one of the eight regions. If you stopped here, the headline writes itself: "global pop is wistful."

But look at how close together the regions actually sit. This is each region's average longing score, plotted on the classifier's full 0–1 scale:

0.0 0.5 1.0

Eight regions, eight different cultures, eight different languages of origin — and the entire spread is 0.092 points wide, all bunched at the top of the scale. A zero-shot classifier forced to score every label independently will lean on whichever label it's least stingy with, and it turns out that's longing, everywhere, regardless of what the song is actually about. Raw scores here don't compare regions. They compare a shared classifier bias against itself, eight times.

02 — Flattening the Mix

Subtract the global average, and a signal appears.

Think of it like a mixing board. Every region's lyrics run hot on the same few frequencies — sensual, longing — because those are just loud in pop songwriting generally, everywhere, always. That's the shared mix. It drowns out what makes any one region distinct.

So we flatten it: take each region's average score per emotion and subtract the eight-region global average for that same emotion. What's left isn't "how sensual is this region" — it's "how much more sensual than everyone else." Positive means a region leans into that emotion harder than the global pool. Negative means it leans away.

above the 8-region average below the 8-region average

That's the chart in every section below. Same ten emotions, same scale, every time — so a longer amber bar in Japan really is a bigger deviation than a shorter one in Colombia.

03 — Side A, Track 1

The Hopefuls

Japan, Singapore, and Taiwan all break the same direction — toward hope, love, and joy, and hardest away from anger of anywhere on the chart.

04 — Side A, Track 2

The Ache

The USA pulls hardest toward despair and heartbreak, away from love. The Global chart leans the same direction — toward longing, heartbreak, and grief — just far more mildly, which tracks: it's an aggregate of everywhere else.

05 — Side A, Track 3

The Heat

Spain, Colombia, and Argentina all spike hardest on anger, well past anywhere else on the chart — Spain and Colombia's charts share a Latin-market reggaetón overlap (literally the same songs some weeks), and Argentina's chart leans the same way even without that overlap.

06 — The Full Mix

All eight regions, one grid.

The same deviation-from-average measure, every region and every emotion at once. Read down a column to see which regions push an emotion hardest; read across a row to see a region's whole emotional signature in one line.

above average below average

07 — Two Ears, Two Verdicts

We ran two different classifiers. They barely agree.

Everything above comes from a zero-shot NLI model (facebook/bart-large-mnli) judging each song against ten hand-picked, lyric-specific labels. As a sanity check, the same 1,105 songs were also run through GoEmotions, a classifier trained on 28 labels from Reddit comments — a completely different taxonomy, model, and scoring mechanism.

Restricted to the four labels both share — love, joy, grief, anger — here's how much they actually agree:

90.3% / 90.3%
coverage, zero-shot / GoEmotions
0.30–0.48
score correlation (r), shared labels
29.8%
dominant-emotion agreement
r = 0.15
regional map agreement (the test that matters)

Coverage is identical now — both classifiers are held to the same "no scoreable lyrics" rule for unclassified, so that's no longer a place for a difference to hide. Dominant-emotion disagreement is mostly a taxonomy mismatch, not a real disagreement: zero-shot's sensual, longing, heartbreak, and lonely — the vocabulary that actually fits a love song — don't exist in GoEmotions at all, so a song zero-shot calls "heartbreak" collapses into GoEmotions' nearest bucket, usually just love or sadness.

The one number that matters more than the rest: z-score each classifier's regional averages on the four shared labels and ask whether the two draw the same map of which region leans which way. They don't — r = 0.15 overall, and on joy and grief specifically the two classifiers are telling close to opposite stories (r = −0.03 and r = −0.40). love is the one label where they mostly agree (r = 0.67). Every regional claim on this page comes from the zero-shot classifier — read it as that classifier's read on the charts, not as a classifier-agnostic fact.

04.1 · bart-large-mnli

Zero-shot NLI

  • +Custom vocabulary built for lyrics — longing, heartbreak, despair exist as labels at all
  • +Every label scored independently — same scoring contract as GoEmotions, magnitudes comparable in kind if not in scale
  • +Well calibrated at the shared 0.30 confidence bar — only 5 songs (0.5%) flagged low-confidence
  • Slower — one NLI pass per label, per lyric chunk
  • Labels are hand-picked, not benchmarked against any ground truth

04.2 · roberta-go_emotions

GoEmotions

  • +Fast — one supervised forward pass returns all 28 labels at once
  • +Trained on real labeled data, not hand-tuned hypotheses
  • Reddit-comment taxonomy: admiration, curiosity, approval don't map onto lyrical themes
  • No romantic/melancholic vocabulary — heartbreak and longing have nowhere to go
  • Conservative calibration at the shared 0.30 bar — 94 songs (8.5%) flagged low-confidence, vs. 5 for zero-shot

Neither is "more correct" — there's no labeled ground truth for song emotion to check either against. Zero-shot trades speed for a vocabulary that actually fits the medium; GoEmotions trades vocabulary for a faster, validated model. Every regional pattern on this page is a product of that first choice.

08 — Liner Notes

How this was made.

01
IngestWeekly Spotify regional charts → Genius API for lyrics, deduplicated across regions, checkpointed to survive rate limits. 42 of ~694 unique tracks had no retrievable lyrics.
02
Detect languageUnicode script detection first (catches mixed-script songs like Japanese lyrics with English hooks), FastText lid.176 for the rest.
03
TranslateGoogle Translate via deep-translator, everything to English. QA'd by re-running FastText on the output to confirm it actually reads as English.
04
Classify emotionTwo parallel approaches over the same lyrics — zero-shot NLI with a custom 10-label vocabulary, and GoEmotions' supervised 28-label model. Compared, not merged.
05
Aggregate & comparePer-song scores joined back to chart rank and region; region averages measured against the 8-region global mean to surface what's actually distinctive.

09 — Outro

The average isn't the story. The deviation is.

Raw emotion scores told us pop music is wistful, everywhere, always — technically true and completely useless. It took one subtraction to get somewhere real: Japan, Singapore, and Taiwan's charts lean hopeful, the USA's leans toward despair and heartbreak, Spain, Colombia, and Argentina run hot with anger. None of that shows up until you stop asking "how much" and start asking "how much more than everyone else" — and even then, it's one classifier's read: a second classifier trained on a different taxonomy draws a visibly different regional map on two of the four labels it shares with this one (§07).