Enhancing glass-box methods for part-of-speech tagging and text classification

Licentiate thesis, 2025

This thesis explores interpretable methods in two natural language processing (NLP) tasks, namely part-of-speech (POS) tagging and text classification. Currently, the NLP field is centered on the development and deployment of deep neural networks (DNNs), which have been established as the state-of-the-art and exhibit high performance in a variety of benchmarks. These models are to all intents and purposes black-boxes that lack interpretability, which is a major disadvantage when considering their use in high-stakes situations where transparency is essential.

The focus of this thesis is therefore on enhancing inherently transparent (glass-box) methods, thus establishing a foundation for their use in high-stakes scenarios. First, the task of POS tagging is considered. While often considered a solved problem, the findings here show that several challenges in POS tagging still remain. Based on these findings, a rule-based approach is explored for correcting the output of POS taggers. Second, interpretable text classification is studied with an enhanced linear classification method. The results demonstrate that a fully interpretable classifier can achieve a high performance when using the proposed enhancements, approaching that of pretrained DNN-based methods.