Genre Analysis Via Constituent Tree Structure

Abstract

Among the many tasks within the field of natural language processing, genre analysis is one of the most difficult as there is no objective standard of what the features of a genre are. Past works have attempted to apply a combination of syntactic and lexical machine learning and deep learning models to categorize texts by genre effectively. Syntactic features have additionally been found to be important features in authorship analysis. This paper applies previous findings related to the use of syntactic features to the area of genre analysis, specifically testing whether constituency based parse trees derived from the Penn Treebank, and other related lexical features, are valuable to different supervised machine learning models, such as the Naive Bayes and Maximum Entropy classifiers in determining genre. The accuracies of these models as compared to the baseline show that these syntactic features are indeed important and result in a significant increase in accuracy.