Preview

PDF, English - main document Download (1MB) | Lizenz: Rights reserved - Free Access

Abstract

Event extraction is the task of automatically finding events in texts. It is an important step towards automatic text understanding because events not only describe what happens, but also assign roles to participating entities. Events are complex semantic structures. Finding events in an information extraction setting consists of finding a word which indicates the event on the lexical surface, called the trigger, and a set of arguments, entity mentions which play a role in the event, along with the roles they play.

Many event extractors published to date capture only intra-sentential contexts and rely on shallow features like the neighbor words and immediate dependency relations. This thesis is first concerned with expanding the information available to an event extractor. We propose a method to make the global (document-wide) context available to the decoding process of a local (intra-sentential) state-of-the-art event extractor. The resulting system shows the best evaluation results to date (summer 2018). Our system improves overall performance because it can improve the identification and classification of triggers. We could not devise successful features for a global event argument detection. This is the starting point for the second part of the thesis, which proposes a syntax-based event extractor which can use multimodal sentence representations (lexical and syntactic) to better perform event extraction. We evaluate the use of such a system in detail and show which syntax representation methods perform best.