Author identification of short texts using dependency treebanks without vocabulary

Abstract

Abstract How to classify short texts effectively remains an important question in computational stylometry. This study presents the results of an experiment involving authorship attribution of ancient Greek texts. These texts were chosen to explore the effectiveness of digital methods as a supplement to the author’s work on text classification based on traditional stylometry. Here it is crucial to avoid confounding effects of shared topic, etc. Therefore, this study attempts to identify authorship using only morpho-syntactic data without regard to specific vocabulary items. The data are taken from the dependency annotations published in the Ancient Greek and Latin Dependency Treebank. The independent variables for classification are combinations generated from the dependency label and the morphology of each word in the corpus and its dependency parent. To avoid the effects of the combinatorial explosion, only the most frequent combinations are retained as input features. The authorship classification (with thirteen classes) is done with standard algorithms—logistic regression and support vector classification. During classification, the corpus is partitioned into increasingly smaller ‘texts’. To explore and control for the possible confounding effects of, e.g. different genre and annotator, three corpora were tested: a mixed corpus of several genres of both prose and verse, a corpus of prose including oratory, history, and essay, and a corpus restricted to narrative history. Results are surprisingly good as compared to those previously published. Accuracy for fifty-word inputs is 84.2–89.6%. Thus, this approach may prove an important addition to the prevailing methods for small text classification.