Improve section-related behaviour

[kermitt2]

When we recognized explicitly a "dataset section" (like "materials and methods" section, with the preliminary CRF section classifier), we should further restrict and control the second level DL classification on sentences in other filtered sections, e.g. like introduction, to avoid too many annotated data sentences for the same dataset.