Co-authorship networks are undirected graphs of collaboration between scientists.  Analysis of co-authorship networks can help us to quantify and understand the structure and evolution of collaborations in academic societies. Additionally, patterns mined from co-authorship networks have a few important applications such as academic author ranking, reviewer recommendation and prediction of future research collaborations. The nodes in a co-authorship network are identified by researcher names contained in the bibliographical records used as the data source in network extraction. For manually created bibliographical databases there might be spelling errors and other inconsistencies in bibliographical records. In this paper, we present a semi-automatic technique for the construction of co-authorship networks that involves analysis of author names. Namely, different string similarity measures are employed to detect authors whose names appear differently in database records. The technique is experimentally evaluated on bibliographical records of the electronic library of the Mathematical Institute of the Serbian Academy of Science and Arts (eLib). In the paper we discuss the differences in both macroscopic and microscopic properties of the eLib co-authorship networks obtained with and without name analysis.