Robust part-of-speech tagging of social media text

Part-of-speech (PoS) taggers are an important processing component in many Natural Language Processing (NLP) applications, which led to a variety of taggers for tackling this task. Recent work in this field showed that tagging accuracy on informal text domains is poor in comparison to formal text domains. In particular, social media text, which is inherently different from formal standard text, leads to a drastically increased error rate. These arising challenges originate in a lack of robustness of taggers towards domain transfers. This increased error rate has an impact on NLP applications that depend on PoS information. The main contribution of this thesis is the exploration of the concept of robustness under the following three aspects: (i) domain robustness, (ii) language robustness and (iii) long tail robustness. Regarding (i), we start with an analysis of the phenomena found in informal text that make tagging this kind of text challenging. Furthermore, we conduct a comprehensive robustness comparison of many commonly used taggers for English and German by evaluating them on the text of several text domains. We find that the tagging of informal text is poorly supported by available taggers. A review and analysis of currently used methods to adapt taggers to informal text showed that these methods improve tagging accuracy but offer no satisfactory solution. We propose an alternative tagging approach that reaches an increased multi-domain tagging robustness. This approach is based on tagging in two steps. The first step tags on a coarse-grained level and the second step refines the tags to the fine-grained tags. Regarding (ii), we investigate whether each language requires a language-tailored PoS tagger or if the construction of a competitive language independent tagger is feasible. We explore the technical details that contribute to a tagger's language robustness by comparing taggers based on different algorithms to learn models of 21 languages. We find that language robustness is a less severe issue and that the impact of the tagger choice depends more on the granularity of the tagset that shall be learned than on the language. Regarding (iii), we investigate methods to improve tagging of infrequent phenomena of which no sufficient amount of annotated training data is available, which is a common challenge in the social media domain. We propose a new method to overcome this lack of data that offers an inexpensive way of producing more training data. In a field study, we show that the quality of the produced data suffices to train tagger models that can recognize these under-represented phenomena. Furthermore, we present two software tools, FlexTag and DeepTC, which we developed in the course of this thesis. These tools provide the necessary flexibility for conducting all the experiments in this thesis and ensure their reproducibility.