Convolutional Neural Networks for Author Profiling in PAN 2017

Social media data allows researchers to establish relationships between everyday language and people’s sociodemographic variables, such as gender, age, language variety or personality. These variables configure social groups, where author profiling attempts to exploit the idea that they share a common language. This work describes our proposed method for the PAN 2017 Author Profiling shared task. We trained separate models for gender and language variety using a Convolutional Neural Network (CNN). We explored parameters such as the size of the input of the network, the size of the convolutional kernels, the number of kernels and the type of input. We found experimentally that sequences of words performed better than sequences of characters as input for the CNN. We obtained 0.66, 0.73, 0.81 and 0.57 of accuracy in the test partition for English, Spanish, Portuguese and Arabic respectively.