Explainable NLP for Human-AI Collaboration

With more data and computing resources available these days, we have seen many novel Natural Language Processing (NLP) models breaking one performance record after another. Some of them even outperform human performance in some specific tasks. Meanwhile, many researchers have revealed weaknesses and irrationality of such models, e.g., having biases against some sub-populations, producing inconsistent predictions, and failing to work effectively in the wild due to overfitting. Therefore, in real applications, especially in high-stakes domains, humans cannot rely carelessly on predictions of NLP models, but they need to work closely with the models to ensure that every final decision made is accurate and benevolent. In this thesis, we devise and utilize explainable NLP techniques to support human-AI collaboration using text classification as a target task. Overall, our contributions can be divided into three main parts. First, we study how useful explanations are for humans according to three different purposes: revealing model behavior, justifying model predictions, and helping humans investigate uncertain predictions. Second, we propose a framework that enables humans to debug simple deep text classifiers informed by model explanations. Third, leveraging on computational argumentation, we develop a novel local explanation method for pattern-based logistic regression models that align better with human judgements and effectively assist humans to perform an unfamiliar task in real-time. Altogether, our contributions are paving the way towards the synergy of profound knowledge of human users and the tireless power of AI machines.Open Acces