Estimation of Average Information Content: Comparison of Impact of Contexts

Abstract

In this paper, we compare Linear Mixed Effect Models (LMM) which utilise the predictors Average Information Content (IC) and frequency for the prediction of lengths of aspect-marked verbs. IC is the information which target elements convey to their context. Focusing on typologically diverse languages, we took as contexts dependency frames and n-grams, and found that IC estimated from n-grams outperforms IC estimated from dependency frames: the models which utilise IC from n-grams achieve high correlations between predicted and actual verbs’ lengths, while models which utilise IC form dependency frames perform poorly. Only in few languages we found prediction effects of IC.