Braunschweig, KatrinThiele, MaikEberius, JulianLehner, WolfgangSaake, GunterHenrich, AndreasLehner, WolfgangNeumann, ThomasKöppen, Veit2018-10-242018-10-242013978-3-88579-610-7https://dl.gi.de/handle/20.500.12116/17432Named entity extraction is an established research area in the field of information extraction. When tailored to a specific domain and with sufficient pre-labeled training data, state-of-the-art extraction algorithms have achieved near human performance. However, when presented with semi-structured data, informal text or unknown domains where training data is not available, extraction results can deteriorate significantly. Recent research has focused on crowdsourcing as an alternative to automatic named entity extraction or as a tool to generate the required training data. While humans easily adapt to semi-structured data and informal style, a crowd-based approach also introduces new issues due to monetary costs or spamming. We address these issues by combining automatic named entity extraction algorithms with crowdsourcing into a hybrid approach. We have conducted a wide range of experiments on real world data to identify a set of subtasks or operators, that can be performed either by the crowd or automatically. Results show that a meaningful combination of these operators into complex processing pipelines can significantly enhance the quality of named entity extraction in challenging scenarios, while at the same time reducing the monetary costs of crowdsourcing and the risk of misuse.enText/Conference Paper1617-5468