Lately, the focus has shifted towards hybrid methods, either by exploiting the best aspects of both the above-mentioned types of techniques, eg, by using rules to bootstrap the ML classification process,31,32 or by aggregating several ML techniques into cascaded classifiers.3 The latter has showed promising results in Bio-NER contexts. Consequently, we have followed this direction and designed our recognition process as a sentence-based classification via an ensemble of four classifiers. The finer-grained annotation level required to capture the content and conceptual structure of a scientific article16 has motivated our choice of sentence-based classification. This article aims to bring the following contributions, envisioned to support other researchers working on the topic, as well as to enable the development of automated mechanisms for building argumentative discourse networks or for tracking the evolution of scientific artifacts: (i) we propose, develop and evaluate a hybrid Machine Learning ensemble, as opposed to the existing research that makes use of a single classification technique; and (ii) we use classification features built strictly from a local, publication perspective, as opposed to corpus-wide statistics used within all the other approaches. This last aspect can make the difference between a model biased towards the domain/corpus used for training and one that makes use of more generic elements and hence displays an increased versatility. Our experimental results show that such a model achieves accuracy comparable to the state of the art, even without relying on corpus-based features.