Improving LSTMs' under-performance in authorship attribution for short texts

We present a novel approach for conducting authorship attribution over tweets using Long-Short Term Memory networks (LSTMs). Vanilla LSTMs use the last hidden state for prediction. Our strategy introduces a mechanism based on Max Pooling to process all the hidden states simultaneously, which helps the model to better detect authors’ stylometry. We obtain a 4% accuracy improvement with respect to vanilla LSTMs.