PAN 2017: Author Profiling - Gender and Language Variety Prediction

We present the results of gender and language variety identification performed on the tweet corpus prepared for the PAN 2017 Author profiling shared task. Our approach consists of tweet preprocessing, feature construction, feature weighting and classification model construction. We propose a Logistic regression classifier, where the main features are different types of character and word n-grams. Additional features include POS n-grams, emoji and document sentiment information, character flooding and language variety word lists. Our model achieved the best results on the Portuguese test set in both—gender and language variety—prediction tasks with the obtained accuracy of 0.8600 and 0.9838, respectively. The worst accuracy was achieved on the Arabic test set.