The dark side of text classification: addressing the negative aspects plaguing text classifiers

Text classification (automatically classifying text into a category) is a well explored area, and can achieve strong results on many tasks. However, there are areas which are still problematic and need to be further explored and improved. Text classifiers are vulnerable to attacks which modify text to trick the classifier. Thus, strong defenses against the attacks are needed. Furthermore, text classifiers can be used by bad actors to harm privacy and freedoms, thus algorithms which modify text to hide from censors are also needed. Finally, text classifiers have been found to unknowingly show bias against minority communities, thus strong bias detection and mitigation algorithms are need as well. In this thesis, we describe and motivate these problems and present new methods for their solution. Specifically, we present a new defense against attacks which is simple but effective in mitigating attacks. We present ways to compare and combine this defense with other text classifier defenses as well as address the flaws of attacks exposed by the defense. Next we present a new method to hide text from censorship systems by modifying it and leverages humans’ craftiness. Third, we present a community focused methodology to identify bias in text classifiers and datasets. Together these three contributions start to address the noted problems with text classification and open the door for future investigations and improvements of these areas.