Neural induction of a lexicon for fast and interpretable stance classification.

Large-scale social media classification faces the following two challenges: algorithms can be hard to adapt to Web-scale data, and the predictions that they provide are difficult for humans to understand. Those two challenges are solved at the cost of some accuracy by lexicon-based classifiers, which offer a white-box approach to text mining by using a trivially interpretable additive model. However current techniques for lexicon-based classification limit themselves to using hand-crafted lexicons, which suffer from human bias and are difficult to extend, or automatically generated lexicons, which are induced using point-estimates of some predefined probabilistic measure on a corpus of interest. In this work we propose a new approach to learn robust lexicons, using the backpropagation algorithm to ensure generalization power without sacrificing model readability. We evaluate our approach on a stance detection task, on two different datasets, and find that our lexicon outperforms standard lexicon approaches