Knowledge-based Supervision for Domain-adaptive Semantic Role Labeling

Semantic role labeling (SRL) is a method for the semantic analysis of texts that adds a level of semantic abstraction on top of syntactic analysis, for instance adding semantic role labels like Agent on top of syntactic functions like Subject . SRL has been shown to benefit various natural language processing applications such as question answering, information extraction, and summarization. Automatic SRL systems are typically based on a predefined model of semantic predicate argument structure incorporated in lexical knowledge bases like PropBank or FrameNet. They are trained using supervised or semi-supervisedmachine learningmethods using training data labeled with predicate (word sense) and role labels. Even state-of-the-art systems based on deep learning still rely on a labeled training set. However, despite the success in an experimental setting, the real-world application of SRL methods is still prohibited by severe coverage problems (lexicon coverage problem) and lack of domain-relevant training data for training supervised systems (domain adaptation problem). These issues apply to English, but are even more severe for other languages, for which only small resources exist. The goal of this thesis is to develop knowledge-based methods to improve lexicon coverage and training data coverage for SRL. We use linked lexical knowledge bases to extend the lexicon coverage and as a basis for automatic training data generation across languages and domains. Links between lexical resources have already been previously used to address this problem, but the linkings have not been explored and applied at a large scale and the resulting generated training data only contained predicate (word sense) labels, but no role labels. To create predicate and role labels, corpus-based methods have been used. These rely on the existence of labeled training data as sources for label transfer to unlabeled corpora. For certain languages, like German or Spanish, several lexical knowledge bases, but only small amounts of labeled training data exist. For such languages, knowledge-based methods promise greater improvements. In our experiments, we target FrameNet, a lexical-semantic resource with a strong focus on semantic abstraction and generalization, but the methods developed in this thesis can be extended to other models of predicate argument structure, like VerbNet and PropBank. This