Age and gender classification of tweets using convolutional neural networks

Abstract

Determining age and gender from a series of texts is useful for areas such as business intelligence and digital forensics. We explore the use of convolutional neural networks together with word2vec word embeddings for this task in comparison to handcrafted features. The network constructed consists of five layers and is trained using adadelta. It starts with an embedding layer where a word is represented by a vector, followed by a convolutional layer composed of three filters, each with 100 feature maps. It is followed by a max-over-time pooling layer which is done on each map and the resulting features are concatenated before a dropout layer and a softmax layer. The network was trained to classify age and gender for English and Spanish tweets. The predictions per tweet were aggregated using the majority prediction as the final prediction for the user who gave the tweets. The results outperform previous experiments. The highest English age and gender classification accuracy obtained are 49.6\% and 72.1\% respectively. The highest Spanish age and gender classification accuracy obtained on the other hand are 56.0\% and 69.3\% respectively.