Modelling crosslinguistic n‑gram correspondence in typologically different languages

Abstract

Abstract N‑gram analysis (popularized e.g. by Biber et al., 1999 ) has become a popular method for the identification of recurrent language patterns. Although the extraction of n‑grams from a corpus may seem straightforward, it proves to be very challenging when applied cross-linguistically (cf. e.g. Ebeling and Ebeling, 2013 ; Granger and Lefer, 2013 ; Cermakova and Chlumska, 2017 ). The major issue is that the quantities of n‑grams of a certain length in typologically different languages do not correspond. Consequently, n‑grams of a given length may function differently across languages, rendering a direct comparison inadequate. Our paper introduces a function capable of modelling the relation between the quantities of n‑grams in typologically distant languages, using the example of Czech and English (and some other language pairs). Based on our model, we can suggest what n‑gram lengths should be contrasted to better reflect the size of n‑gram inventories in each language. The correspondence may not be intuitive (e.g. a Czech 2-gram may best correspond to an English 2.5-gram), but it still provides researchers with a general guide as to what might be useful to include in their analysis (e.g. in this case 2-grams in Czech and 2- and 3-grams in English).