Lexicon Induction for Interpretable Text Classification

The automated classification of text documents is an active research challenge in document-oriented information systems, helping users browse massive amounts of data, detecting likely authors of unsigned work, or analyzing large corpora along predefined dimensions of interest such as sentiment or emotion. Existing approaches to text classification tend toward building black-box algorithms, offering accurate classification at the price of not understanding the rationale behind each algorithmic prediction. Lexicon-based classifiers offer an alternative to black-box classifiers by modeling the classification problem with a trivially interpretable classifier. However, current techniques for lexicon-based document classification limit themselves to using either handcrafted 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 in the corpus of interest. This paper proposes LexicNet, an alternative way of generating high accuracy classification lexicons offering an optimal generalization power without sacrificing model interpretability. We evaluate our approach on two tasks: stance detection and sentiment classification. We find that our lexicon outperforms baseline lexicon induction approaches as well as several standard text classifiers.