Fig. 1

Corpus compilation pipeline. This pipeline, from bottom to top, shows the steps that were taken to compile the corpus (171 sentence pairs) that was later evaluated by annotators regarding the semantic similarity between the sentence pairs. First, two subsets, the anaerobiosis-FNR and the more general one, were compiled using different strategies. Then, a basic STS process was applied to both subsets in order to have a preliminary semantic similarity evaluation. This preliminary evaluation was used to select candidate sentences, creating a corpus that ended up with 40% of sentences from the anaerobiosis subset and 60% from the general subset