Lithuanian author profiling with the deep learning

Abstract

We address the Lithuanian author profiling task in two dimensions (AGE and GENDER) using two deep learning methods (i.e., Long Short-Term Memory-LSTM) and Convolutional Neural Network-CNN) applied on the top of Lithuanian neural word embeddings. We also investigate an impact of the training dataset size on the author profiling accuracy. The best results are achieved with the largest datasets, containing 5,000 instances in each class. Besides, LSTM was more effective on the smaller datasets, and CNN-on the larger ones. We compare the deep learning methods with the traditional machine learning methods (in particular, Naive Bayes Multinomial and Support Vector Machine), and frequencies of elements as the feature representation). The comparison revealed that the deep learning is not the best solution for our author profiling task.