Scientific texts represent a rich source of unstructured knowledge. Extracting this knowledge in a supervised manner can become highly expensive in time and human resources. Moreover supervised models are domain- and language-dependent which make them hard to maintain and extend. Hence unsupervised methods have received a lot of attention from researchers in the fields of information extraction and data mining. In this paper, we present our experiments with self-organizing maps (SOMs) for the task of open relation extraction. We combine contextual features of different level (lemmas and parts-of-speech) to help the algorithm to automatically discover lexical and morphological patterns in the corpus. The evaluation results show that our model yields a better performance than the widely used K-means clustering algorithm with the same feature set.