Authorship Verification on Short Text Samples Using Stylometric Embeddings

摘要

Given the increasing amounts of textual data published online and our inability to reliably identify a person by their writing style, impersonation in the context of social media applications becomes a real-world problem. This work explores how deep learning and metric learning techniques can be applied to the challenge of authorship verification-given a collection of text samples by one author and another document of unknown origin, determine if the new document is written by the same author or not. Using fastText word embeddings, deep LSTMs, and triplet loss, we propose a system that is able to learn stylometric embeddings of different documents and measure their stylistic distance. Unlike most approaches that work on entire documents, our system is able to work on very short text samples of 1-3 sentences, which resembles the length of typical social media posts. We successfully evaluated our approach on the PAN 2014 challenge on authorship verification for English text. The presented system outperforms competing approaches in the PAN 2014 challenge when using 10 short text samples or more.