Diachronic Topic Trends
Aggregate relative frequencies of thematic keyword sets across the full temporal span of the corpus, normalised per 10,000 running tokens per year. Switch between four classification methods using the tabs above the chart, and use the chips below to show or hide individual topics.
Linguistic Insights
How to read these curves
Three things frame any interpretation. First, annual token counts are uneven — trends before roughly 1970 should be read with caution; sudden single-year spikes may reflect just a few influential songs in a sparse year. Second, because the corpus pools diverse genres, a rising curve can signal changing genre composition rather than language change; the large HipHop archive in particular lifts certain topics in recent decades. Third, the four methods — see explanations below — are independent, so where they agree a trend is robust, while divergence shows the category behaves differently when defined lexically versus distributionally.
Where the topic categories come from
Each curve aggregates a thematic keyword set. The base list (Schneider et al. 2022) was built largely on newspaper language, so it fits song lyrics only partly — about three-quarters of its keywords occur in the corpus, and its category sizes reflect journalism rather than pop (Politik is large, Liebe small). To close this gap, the list was extended with corpus-derived keywords like verliebt or Sehnsucht identified via G²-keyness. Even so, the curves seem reliable for relative change within a category over time, but not for cross-category comparison (larger keyword sets inflate absolute rates). Read each line as the trajectory of one topic as operationalised here, not as a complete thematic corpus map.
Keyword tokens per 10,000 running tokens per year · dots = sparse years (<5,000 total tokens) · 3-year smoothing applied
Methods
Keyword counting
For each year the script counts how often words from a predefined thematic list appear in the lyrics and divides that count by the total number of running tokens, giving a rate per 10,000 words. The keyword list (Schneider et al. 2022, extended with corpus-derived additions) assigns each word to one or more categories such as Gewalt or Liebe. What you see is essentially a dictionary lookup: no model, no inference, just counting.
Seeded LDA
Latent Dirichlet Allocation is a probabilistic topic model that treats each song as a mixture of latent topics. We use the 30 most frequent keywords from each category as seeds that pull the model towards human-readable themes. The result for each song is a set of topic proportions — how much of this song is about Politik, how much about Liebe — and the annual curve averages those proportions across all songs from that year.
Biterm Topic Model
Standard LDA struggles with short texts because individual documents carry too little signal. The Biterm Topic Model works on word pairs (biterms) co-occurring within a song rather than on full document statistics, making it better suited to sparse data. The model is trained without seeds, so topics emerge purely from co-occurrence patterns. Each discovered topic is then labelled by finding which predefined category its top words overlap with most. Because there is no seeding the BTM sometimes surfaces genre or register clusters — HipHop, Schlager — rather than clean semantic themes.
Seeded LDA with corpus-derived seeds
This method combines the seeding idea of Seeded LDA with vocabulary drawn from the corpus itself rather than from the external keyword list. The seeds come from the BTM's top words per topic, so they are grounded in song language from the start. This also helps where the newspaper-oriented keyword list has gaps.
These analyses are based on corpus data as of July 29, 2026.