Attention-based Feature Construction and Authorship Identification

This paper uses the attention mechanism to integrate lexicon and syntactic features and identify authors. Firstly, a representation method of node embedding of syntax tree is proposed. The node of syntax tree is represented as the sum of embeddings of all its dependency arcs. The syntactic information and the association information among words are introduced into the deep learning model. Then the syntactic attention network is constructed, and the syntax-aware vector is obtained through the network. The vector incorporates information of dependencies, part-of-speech and words. Then, the representation of the sentence is obtained through the sentence attention network. Finally, the classifier is used for classification. The proposed model achieves high performance in authorship identification experiments on 3 datasets.