Comparing Annotation Types and n-Gram Sizes
Volker Gast · Cambridge University Press eBooks · 2022
The chapter addresses a problem of contrastive pragmatics: How can we study correspondences between pragmatic markers in two languages if one language has a class of elements that the other language lacks? Specifically, the contribution deals with modal particles of German (ja and doch) and their reflexes in English translations. As there is no predetermined set of potential English correspondences, traditional distributional analyses are not feasible, and methods from Natural Language Processing are explored instead. Using 32 types of n-grams, differing in length and type of annotation, three classification tasks are carried out, in order to identify cues in the English translations that reflect the presence (or absence) of a particle in the German original. The results show that lemma-unigrams and -bigrams are often most informative (i.e. most accurate), while trigrams and 1-skip-2-grams provide important information about concomitants of modal particles that unigrams and bigrams miss. The results show that linguistic observables (n-grams) as the basis of quantitative analyses need to be carefully selected and explored in terms of their contribution to linguistic analysis.