Domain Adaptation for Authorship Attribution: Improved Structural Correspondence Learning

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2226–2235, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Domain Adaptation for Authorship Attribution: Improved Structural Correspondence Learning Upendra Sapkota University of Alabama at Birmingham upendra@uab.edu Thamar Solorio University of Houston solorio@cs.uh.edu Manuel Montes-y-G´omez Instituto Nacional de Astrof´ısica, Optica y Electr´onica mmontesg@inaoep.mx Steven Bethard University of Alabama at Birmingham bethard@uab.edu Abstract We present the first domain adaptation model for authorship attribution to leverage unlabeled data. The model includes exten- sions to structural correspondence learning needed to make it appropriate for the task. For example, we propose a median-based classification instead of the standard binary classification used in previous work. Our results show that punctuation-based charac- ter n-grams form excellent pivot features. We also show how singular value decom- position plays a critical role in achieving domain adaptation, and that replacing (in- stead of concatenating) non-pivot features with correspondence features yields better performance. 1 Introduction Authorship Attribution (AA) can be used for his- torical purposes, such as disentangling the differ- ent authors contributing to a literary work. It can also help in understanding language evolution and change at the individual level, revealing a writer’s changes in linguistic patterns over time (Hirst and Feng, 2012). Authorship attribution can also help to settle disputes over the original creators of a given piece of text. Or it can help build a prose- cution case against an online abuser, an important application especially considering the rising trends in cyber-bullying and other electronic forms of teen violence1. The absorbing social media networks, together with the ever increasing use of electronic communications will require robust approaches to authorship attribution that can help to determine with certainty the author of a text, determine the provenance of a written sample, and in sum, help us determine the trustworthiness of electronic data. 1http://cyberbullying.org/ One of the scenarios that has received limited attention is cross-domain authorship attribution, when we need to identify the author of a text but all the text with known authors is from a differ- ent topic, genre, or modality. Here we propose to solve the problem of cross-domain authorship attri- bution by adapting the Structural Correspondence Learning (SCL) algorithm proposed by Blitzer et al. (2006). We make the following contributions: • We introduce the first domain adaptation model for authorship attribution that combines labeled data in a source domain with unla- beled data from a target domain to improve performance on the target domain. • We examine two sets of features that have previously been successful in cross-domain authorship attribution, explain how these can be used to select the “pivot” features required by SCL, and show that typed n-gram features (which differentiate between the the in their and the the in breathe) produce simpler mod- els that are just as accurate. • We propose a new approach for defining SCL’s pivot feature classification task so that it is able to handle count-based features, and show that this median-based approach outper- forms the …

Content truncated to 500 words