Social research networks such as Mendeley and CiteULike offer various services for collaboratively managing bibliographic metadata. Compared with traditional libraries, metadata quality is of crucial importance in order to create a crowdsourced bibliographic catalog for search and browsing. Artifacts, in particular PDFs which are managed by the users of the social research networks, become one important metadata source and the starting point for creating a homogeneous, high quality, bibliographic catalog. Natural Language Processing and Information Extraction techniques have been employed to extract structured information from unstructured sources. However, given highly heterogeneous artifacts that cover a range of publication styles, stemming from different publication sources, and imperfect PDF processing tools, how accurate are metadata extraction methods in such real-world settings? This paper focuses on answering that question by investigating the use of Conditional Random Fields and Support Vector Machines on real-world data gathered from Mendeley and Linked-Data repositories. We compare style and content features on existing state-of-the-art methods on two newly created real-world data sets for metadata extraction. Our analysis shows that two-stage SVMs provide reasonable performance in solving the challenge of metadata extraction for crowdsourcing bibliographic metadata management.