Improving Cross-domain Authorship Attribution by Combining Lexical and Syntactic Features: Notebook for PAN at CLEF 2019
Abstrakt
Authorship attribution is a problem in information retrieval and computationallinguistics that involves attributing authorship of an unknown documentto an author within a set of candidate authors. Because of this, PAN-CLEF2019 organized a shared task that involves creating a computational model thatcan determine the author of a fanfiction story. The task is cross-domain becauseof the open set of fandoms to which the documents belong. Additionally, theset of candidate authors is also open since the actual author of a document maynot be among the candidate authors.We extracted character-level, word-level andsyntactic information from the documents in order to train a support vector machine.
Our approach yields an overall macro-averaged F1 score of 0.687 on thedevelopment data of the shared task. This is an improvement of 18.7% over thecharacter-level lexical baseline. On the test data, our model achieves an overallmacro F1 score of 0.644.We compare different feature types and find that charactern-grams are the most informative feature type though all tested feature typescontribute to the performance of the model.