Token Trigrams

The 500 strongest trigrams by each of three complementary association measures: logDice, MI³, and Lexical Gravity. The combined set is deduplicated and sortable by all columns. Use the search box to filter by token.

Linguistic Insights

Three complementary association measures

LogDice (Rychlý 2008) measures the strength of association between tokens; being scale-invariant and symmetric, it is robust across corpora of different sizes. MI³ raises the observed co-occurrence frequency to the third power before comparing it to the expected frequency. Lexical Gravity (Daudaravičius & Marcinkevičienė 2004; extended to longer n-grams by Gries & Mukherjee 2010) takes a different approach: rather than counting how often two tokens co-occur, it considers how many distinct word types compete for the neighbouring positions. A combination scores highly when each element strongly constrains what can appear next to it, identifying sequences that behave as fixed, low-variation units. Because these measures are defined for token pairs, applying them to trigrams requires decomposing each three-token sequence into its component associations.

Trigrams as formulaic sequences

Three-token sequences capture refrains, idioms, and stylized phrases that bigrams cannot. Repeated sequences such as na na na or la la la reach near-maximal logDice scores while carrying no referential content; they are purely performative and belong to the same paralinguistic layer (interjections and onomatopoeia) seen in the part-of-speech distribution. Longer fixed phrases such as tut mir leid or auf der Straße reflect the genre's colloquial register. Trigram association scores are generally lower than those of bigrams because the probability space is larger; high-scoring trigrams are therefore especially diagnostic of genre-specific formulaic language.

Token Trigrams – Full Corpus (all archives)

Trigram ↕ Freq ↕ Songs ↕ logDice ↕ MI3 ↕ Lex. Gravity ↕

Filter: f≥20 in ≥3 songs (full corpus & thematic archives) · f≥5 in ≥2 songs (artist archives) · Default sort: logDice ↓ · Click any column header to re-sort

These analyses are based on corpus data as of July 29, 2026.