While food webs are pivotal tools to understand the structure, dynamics and functioning of ecosystems, their reconstruction is not trivial since feeding relationships are not always known. To this end, soil ecologists often simplify the problem by either grouping morphologically similar organisms into trophic groups with known interactions or by assuming that feeding relationships are predictable from consumer diets (e.g. frugivore or bacterivore). Interestingly, the scientific community has collected a considerable amount of information on trophic interactions and feeding habits, some of it being available in structured databases, or disseminated in the scientific and grey literature. However, the large-scale exploitation of these data for food web reconstruction is hampered by their dissemination in a multitude of heterogeneous datasets. The goal of our work is to propose a semantic data integration pipeline whose role will be to extract, aggregate, normalize, and integrate information about trophic interactions and diets into a trophic knowledge graph that will support the automatic reconstruction of soil food webs.