Modeling the Readability of German Targeting Adults and Children: An empirically broad analysis and its cross-corpus validation

Proceedings of the 27th International Conference on Computational Linguistics, pages 303–317 Santa Fe, New Mexico, USA, August 20-26, 2018. 303 Modeling the Readability of German Targeting Adults and Children: An empirically broad analysis and its cross-corpus validation Zarah Weiß Detmar Meurers ICALL-Research.com Department of Linguistics University of T¨ubingen {zweiss,dm}@sfs.uni-tuebingen.de Abstract We analyze two novel data sets of German educational media texts targeting adults and children. The analysis is based on 400 automatically extracted measures of linguistic complexity from a wide range of linguistic domains. We show that both data sets exhibit broad linguistic adaptation to the target audience, which generalizes across both data sets. Our most successful binary classification model for German readability robustly shows high accuracy between 89.4%–98.9% for both data sets. To our knowledge, this comprehensive German readability model is the first for which robust cross-corpus performance has been shown. The research also contributes resources for German readability assessment that are externally validated as successful for different target audiences: we compiled a new corpus of German news broadcast subtitles, the Tagesschau/Logo corpus, and crawled a GEO/GEOlino corpus substantially enlarging the data compiled by Hancke et al. (2012). Zusammenfassung Wir untersuchen zwei neue Datens¨atze deutscher Bildungs- und Mediensprache f¨ur Kinder und Erwachsene. Die Analyse basiert auf 400 automatisch extrahierten Maßen sprachlicher Kom- plexit¨at, die verschiedene linguistische Dom¨anen abdecken. Unsere Ergebnisse zeigen, dass in beiden Datens¨atzen die sprachliche Gestaltung der Texte in ¨ahnlicher Weise breitfl¨achig an ihr jeweiliges Zielpublikum angepasst wird. Unser erfolgreichstes bin¨ares Klassifikationsmodell erzielt Genauigkeitswerte von 89,4% und 98,9% ¨uber beide Datens¨atze hinweg. Unseres Wissens handelt es sich bei diesem umfassend durch verschiedene linguistische Bereiche informtiertem Modell deutscher Text-Lesbarkeit um das erste, f¨ur das robuste Ergebnisse in einer korpus- ¨ubergreifenden Evaluation dokumentiert sind. Dar¨uber hinaus tragen wir mit unserer Arbeit zwei neue Datens¨atze zur Erforschung deutscher Text-Lesbarkeit bei, die auf Texten basieren, deren Eignung f¨ur ihre respektiven Zielgruppen extern durch wiederholte Rezeption validiert wurde: Wir haben aus Untertiteln deutscher Nachrichtenbeitr¨age das Tagesschau/Logo Korpus erstellt. Weiterhin haben wir das GEO/GEOlino Korpus betr¨achtlich erweitert, das urspr¨unglich von Hancke et al. (2012) erstellt wurde. 1 Introduction Readability assessment refers to the task of (automatically) linking a text to the appropriate target audi- ence based on its complexity. A diverse spectrum of potential application domains has been identified for this task in the literature, ranging from the design and evaluation of education materials, to information retrieval, and text simplification. Given the increasing need for learning material adapted to different audiences and the barrier-free access to information required for political and social participation, auto- matic readability assessment is of immediate social relevance. Accordingly, it has attracted considerable This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ 304 research interest over the last decades, particularly for the assessment of English (Crossley et al., 2011; Chen and Meurers, 2017; Feng et al., 2010). For German readability assessment, however, little progress has been made in recent years, despite a series of promising results published around the turn of the …

Content truncated to 500 words