Authorship Attribution of Micro-Messages

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1880–1891, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Authorship Attribution of Micro-Messages Roy Schwartz Oren Tsur Ari Rappoport Institute of Computer Science Hebrew University of Jerusalem {roys02|oren|arir}@cs.huji.ac.il Moshe Koppel Department of Computer Science Bar Ilan University koppel@macs.biu.ac.il Abstract Work on authorship attribution has tradition- ally focused on long texts. In this work, we tackle the question of whether the author of a very short text can be successfully iden- tified. We use Twitter as an experimental testbed. We introduce the concept of an au- thor’s unique “signature”, and show that such signatures are typical of many authors when writing very short texts. We also present a new authorship attribution feature (“flexible pat- terns”) and demonstrate a significant improve- ment over our baselines. Our results show that the author of a single tweet can be identified with good accuracy in an array of flavors of the authorship attribution task. 1 Introduction Research in authorship attribution has developed substantially over the last decade (Stamatatos, 2009). The vast majority of such research has been dedicated towards finding the author of long texts, ranging from single passages to book chapters. In recent years, the growing popularity of social me- dia has created special interest, both theoretical and computational, in short texts. This has led to many recent authorship attribution projects that experi- mented with web data such as emails (Abbasi and Chen, 2008), web forum messages (Solorio et al., 2011) and blogs (Koppel et al., 2011b). This paper addresses the question to what extent the authors of very short texts can be identified. To answer this question, we experiment with Twitter tweets. Twitter messages (tweets) are limited to 140 char- acters. This restriction imposes major difficulties on authorship attribution systems, since authorship at- tribution methods that work well on long texts are often not as useful when applied to short texts (Bur- rows, 2002; Sanderson and Guenter, 2006). Nonetheless, tweets are relatively self-contained and have smaller sentence length variance com- pared to excerpts from longer texts (see Section 3). These characteristics make Twitter data appealing as a testbed when focusing on short texts. Moreover, an authorship attribution system of tweets may have various applications. Specifically, a range of cyber- crimes can be addressed using such a system, includ- ing identity fraud and phishing. In this paper, we introduce the concept of k- signatures. We denote the k-signatures of an author a as the features that appear in at least k% of a’s training samples, while not appearing in the training set of any other author. When k is large, such signa- tures capture a unique style used by a. An analysis of our training set reveals that unique k-signatures are typical of many authors. Moreover, a substantial portion of the tweets in our training set contain at least one such signature. These findings suggest that a single tweet, although short and sparse, often con- tains sufficient information …

Content truncated to 500 words