Engineering Tree Kernels for Semantic Role Labelling Systems

Semantic Role Labelling (SRL) is a complex Natural Language Processing (NLP) task that has received a lot of attention in the latest years. An accurate shallow semantic parser, that recognized predicate-argument structures in a sentence and assigned each argument a semantic (or thematic) role, could be a key factor of larger NLP architectures, human-machine interaction (e. g. high level, semantic-oriented browsing and search systems), information extraction and dialogue based systems, just to name a few. The recognition of semantic structures within a sentence relies on lexical and syntactic information provided by earlier stages of an NLP process, such as lexical analysis, POS (part of speech) tagging and syntactic parsing. The complexity of the SRL task mostly lies in that: (a) this information is generally noisy, i. e. in a real-world scenario the accuracy and reliability of NLP subsystems are generally not very high; (b) the lack of a sound and complete linguistic or cognitive theory about the links between syntax and semantics doesn’t allow an informed, deductive approach to the problem. Still, the large amount of lexical and syntactic information available allows for an inductive approach to the SRL task, which indeed is generally