In this article we explain an automatic approach to detect affixes from entries in a dictionary using the longest common substring algorithm, in the context of chemical named entity recognition on the CHEMDNER corpus. We then show selection and sorting methods in order to better integrate them in a machine learning system.