A theme is a list of words that tend to turn up in the same goods/services descriptions — video, computer, electronic, music, games, say, or store, retail, store services, featuring. Each trademark description is re-described as a mix of themes that sums to one: Amazon's 1997 filing, for example, is about 41% media and entertainment goods, 15% computer and information services, and 13% retail store services. Lead and atypicality compare a filing's theme mix with the average mix of its own Nice class in the five years before and the five years after its filing date.
Nobody wrote these lists. They come from latent Dirichlet allocation, a standard statistical method that starts from a random guess about which words belong together and adjusts the lists, pass after pass, until each description can be explained by a few themes and each theme uses a consistent set of words. The model was fitted once, on a sample of 10,000 descriptions from each of the 45 Nice classes (448,437 in all), using every word and adjacent word pair that appears in at least 50 sampled descriptions (62,168 terms), and then applied to every filing. The labels in the table below were added afterwards by reading each theme's leading words; the model itself only produces numbered word lists.
Fifty is coarse — roughly one theme per Nice class — and chosen on purpose. At fifty themes every theme's leading words genuinely occur together (average normalized pointwise mutual information 0.48, none below zero); at 200 and 500 themes, a tenth and a fifth of themes are word lists whose words rarely co-occur. The themes also come back when the model is refitted with a different random seed far more faithfully at fifty than at 200. The papers' findings do not depend on the choice: the five-year-proof lead penalty holds at 50, 200 and 500 themes and with fifty themes fitted separately inside each class. The cost of coarseness is visible below: a few themes merge unrelated vocabularies (marked mixed), and fast-growing vocabularies concentrated in the largest classes — artificial intelligence, blockchain, cloud computing — have no theme of their own, because the equal-per-class sample under-represents them.
Sorted by share of the whole corpus. "Heaviest in" lists the three classes where the theme carries the largest share of the class's language.
| # | Theme (label added by reading its words) | Corpus share | Class share | Heaviest in |
|---|
Shares are the average theme mix of up to 20,000 scored filings per class (filing years 1995–2019), weighted by class size for the corpus column. The full theme-by-class table is themes_T50.csv. A 500-theme model used as a robustness check has its own theme explorer.