As radio telescopes become more sensitive, the damaging effects of radio frequency interference (RFI) become more apparent. Near radio telescope arrays, RFI sources are often easily removed or replaced; the challenge lies in identifying them. Transient (impulsive) RFI is particularly difficult to identify. We propose a novel dictionary-based approach to transient RFI identification. RFI events are treated as sequences of sub-events, drawn from particular labelled classes. We demonstrate an automated method of extracting and labelling sub-events using a dataset of transient RFI. A dictionary of labels may be used in conjunction with hidden Markov models to identify the sources of RFI events reliably. We attain improved classification accuracy over traditional approaches such as SVMs or a na\"ive kNN classifier. Finally, we investigate why transient RFI is difficult to classify. We show that cluster separation in the principal components domain is influenced by the mains supply phase for certain sources.