Temporal predictive regression models for linguistic style analysis

Abstract

This study focuses on modelling general and individual language change over several decades. A timeline prediction task was used to identify interesting temporal features. Our previous work achieved high accuracy in predicting publication year, using lexical features marked for syntactic context. In this study, we use four feature types (character, word stem, part-of-speech, and word n-grams) to predict publication year, and then use associated models to determine constant and changing features in individual and general language use. We do this for two corpora, one containing texts by two different authors, published over a fifty-year period, and a reference corpus containing a variety of text types, representing general language style over time, for the same temporal span as the two authors. Our linear regression models achieve good accuracy with the two-author data set, and very good results with the reference corpus, bringing to light interesting features of language change.