Capsule Network on Social Media Text: An Application to Automatic Detection of Clickbaits

The advent of groundbreaking deep learning techniques like Capsule Network has changed the way of approaching a problem in data science research. Initially, Capsule Networks were built and tested on image data and found to be of great use. Their usage on textual data is still very limited. In this paper, we try to investigate whether Capsule Network can be used to address a research problem where the classification heavily depends on the textual data. In various classification task involving social networks and online sources, words and sentences across classes do not vary that much. But, the context and representation of those words play a significant role. One such problem is to correctly identify clickbaits. State of the art solutions either take into account various handcrafted features from the data or use efficient text classification techniques like LSTM. Our work is a stepping stone towards examining whether the need of network properties and feature engineering can be omitted while using Capsule Network. It relaxes the effort of manual feature construction from the data and looks beyond the sequence to sequence modelling of an LSTM based approach. Our proposed approach of clickbait detection using a Capsule Network outperforms various existing methods in terms of multiple performance metric.