A Syntax-Aware Encoder for Authorship Attribution

Authorship Attribution (AA) refers to the task of predicting the author of a given text by learning unique writing styles of different authors. It often involves the extraction of social language characteristics of texts such as writing styles to identify authors. Among the many facets of text styles, we postulate that authors have unique and signature syntactic information entangled in their texts. This paper investigates this hypothesis and proposes a Syntax-Aware Encoder (SAE), a novel graph convolutional network based model augmented with the dependency tree to learn the syntax representation for short texts. Subsequently, we leverage the learned latent representations to perform AA. Extensive experiments are conducted on two social medial short-text datasets. The results show that SAE outperforms the state-of-the-art baselines on the AA task.