Using a Hierarchical Softmax Based on the Huffman Coding Tree for Authenticating Arabic Tweets
Bashar Talafha, Mohammad AL-Smadi, Mahmoud Al‐Ayyoub, Yaser Jararweh, Patrick Juola · 2019
Attributing a piece of text to its true author is called Authorship Authentication (AA). This work addresses the AA problem of Arabic tweets. Arabic language is both challenging and understudied. Existing approaches on authenticating Arabic tweets used bag of words features or Stylometric Features coupled with classifiers like SVM. However, the reported accuracy for these approaches is rather low and did not even reach 69%. In this work, we address this problem using two approaches. (a) A baseline approach that uses SVM along with BoW features, and (b) a character-level linear classifier (char-LC) with a rank constraint and a fast loss approximation along with word embeddings based on fasttext. Both approaches give significantly higher accuracies than the results reported in literature with 78.28% for the SVM along with BoW approach and 79.4% for the char-LC.