In this paper, we explore the integration of LLMs with symbolic processing for achieving high granularity event extraction. We will show that the weakness of LLMs in producing structured information, often pointed out in the literature, can be overcomed by designing a domain tailored mapping function (hybridization). In order to support this claim, we compare the results of an in-context learning method with our hybrid methodology and we show that we can achieve superior results (+6.3 %) on a new dataset of subject-predicate-object triples in the medical domain (681 triples for 200 sentences). This result is achieved by leaving the LLM (Llama-3) free to generate the predicate types it is more familiar with, and then applying a mapping function. Besides improving explainability and controllability of the output, the intervention of such a function (which was implemented in five days), causes about a half reduction of GHG emissions produced when processing the corpus.