Authorship attribution is an important problem in many areas including information retrieval and computational linguistics, but also in applied areas such as law and journalism where knowing the author of a document (such as a ransom note) may be able to save lives. The most common framework for testing candidate algorithms is a text classification problem: given known sample documents from a small, finite set of candidate authors, which if any wrote a questioned document of unknown authorship? It has been commented, however, that this may be an unreasonably easy task. A more demanding problem is author verification where given a set of documents by a single author and a questioned document, the problem is to determine if the questioned document was written by that particular author or not. This may more accurately reflect real life in the experiences of professional forensic linguists, who are often called upon to answer this kind of question.
A note to forensic linguists: In order to bridge the gap between linguistics and computer science, we strongly encourage submissions from researchers from both fields. We understand that research groups with expertise in linguistics use manual or semi-automated methods and, therefore, they are not able to submit their software. To enable their participation, we will provide them with the opportunity to analyze the test corpus after the deadline of software submission (mid-April). Their results will be ranked in a separate list with respect to the performance of the software submissions and they will be entitled to describe their approach in a paper.
To develop your software, we provide you with a training corpus that comprises a set of known documents by a single person and exactly one questioned document. There are several such problem instances covering English, Greek, and Spanish and a varying number of known documents (1-10 per problem). All documents within a single problem instance will be in the same language and best efforts are applied to assure that within-problem documents are matched for genre, register, theme, and date of writing. The document lengths vary from a few hundred to a few thousand words.
Your software must take as input the absolute path to a set of problems. For each problem there is a separate sub-folder within that path including the set of known documents and the single unknown document of that problem. The software has to output a single text file
answers.txt with all the produced answers for the whole set of evaluation problems. Each line of this file corresponds to a problem instance, it starts with the ID of the problem followed by a binary (Y)ES/(N)O answer to the question "Is the unknown document written by the author of the known documents?". If you do not want to provide answers for some problems, you can either replace the answer character with "-" or just do not include a line for that problem to your answers. For example, an
answers.txt file may look like this:
EN01 Y EN02 N EN03 - EN04 Y EN07 Y ...
Optionally, you may also provide a score, a real number in the set [0,1] inclusive, where 0 corresponds to NO and 1 to YES. This score should be round with two decimal digits and will allow a more detailed evaluation of your approach. In this case, the scores have to be placed next to the binary answers. It is possible to provide scores even for problems you are not able to provide binary answers. For example, an
answers.txt file with scores may look like this:
EN01 Y 0.90 EN02 N 0.25 EN03 - 0.53 EN04 Y 0.86 EN07 Y 0.74 ...
Use a single whitespace to separate problem ID, binary answer, and score. The naming of the output file is up to you. We reccomend to use the name of the participant group-run.
Performance of the binary classification will be measured as follows:
Participants are be ranked by combining these measures via F1.
In addition, participants may also provide a score, a real number in the set [0,1] inclusive, where 0 corresponds to NO and 1 to YES. A separate ranking for those participants who also submit real scores [0,1] according to the ROC-AUC. For the calculation of ROC curves, any missing answers are assumed to be wrong answers.
Once you finished tuning your approach to achieve satisfying performance on the training corpus, you should run your software on the test corpus.
During the competition, the test corpus will not be released publicly. Instead, we ask you to submit your software for evaluation at our site as described below.
After the competition, the test corpus is available including ground truth data. This way, you have all the necessities to evaluate your approach on your own, yet being comparable to those who took part in the competition.
We ask you to prepare your software so that it can be executed via a command line call. You can choose freely among the available programming languages and among the operating systems Microsoft Windows 7 and Ubuntu 12.04. We will ask you to deploy your software onto a virtual machine that will be made accessible to you after registration. You will be able to reach the virtual machine via ssh and via remote desktop. Please test your software using one of the unit-test-scripts below. Download the script, fill in the required fields, and start it using the sh command. If the script runs without errors and if the correct output is produced, you can submit your software by sending your unit-test-script via e-mail to firstname.lastname@example.org. For more information see the PAN 2013 User Guide below.
Note: By submitting your software you retain full copyrights. You agree to grant us usage rights only for the purpose of the PAN competition. We agree not to share your software with a third party or use it for other purposes than the PAN competition.
The following table lists the F3 performances achieved by teams participating in predator identification:
|Authorship attribution performance|
Bar Ilan University, Israel
|0.718||Oren Halvani, Martin Steinebach, and Ralf Zimmermann|
Fraunhofer Institute for Secure Information Technology SIT, Germany
|0.671||Robert Layton, Paul Watters, and Richard Dazeley|
University of Ballarat, Australia
University of Tartu, Estonia
|0.659||Magdalena Jankowska, Vlado Kešelj, and Evangelos Milios|
Dalhousie University, Canada
|0.659||Darnes Vilariño, David Pinto, Helena Gómez, Saúl León, and Esteban Castillo|
Benemérita Universidad Autónoma de Puebla, Mexico
Technical University of Moldova, Moldova
|0.647||Vanessa Wei Feng and Graeme Hirst|
University of Toronto, Canada
|0.612||Paola Ledesma°, Gibran Fuentes*, Gabriela Jasso*, Angel Toledo*, and Ivan Meza*|
°Escuela Nacional de Antropología e Historia (ENAH) and *Universidad Nacional Autónoma de México (UNAM), Mexico
Amirkabir University of Technology, Iran
|0.600||Michiel van Dam|
Delft University of Technology, The Netherlands
|0.600||Erwan Moreau and Carl Vogel|
Trinity College Dublin, Ireland
|0.576||Arun Jayapal and Binayak Goswami|
Nuance Communications, India
|0.553||Cristian Grozea° and Marius Popescu*|
°Fraunhofer FIRST, Germany, and *University of Bucharest, Romania
|0.541||Anna Vartapetiance and Lee Gillam|
University of Surrey, UK
Know-Center GmbH, Austria
|0.417||Cor J. Veenman° and Zhenshi Li*|
°Netherlands Forensic Institute and *Delft University of Technology, The Netherlands
University Politehnica of Bucharest, Romania
A more detailed analysis of the detection performances with respect to precision, recall, and granularity can be found in the overview paper accompanying this task.
We refer you to: