A method trains a recursive reasoning unit (RRU). The method receives a graph for a set of words and a matrix for a different set of words. The graph maps each word in the set of words to a node with node label and indicates a relation between adjacent nodes by an edge with edge label. The matrix indicates word co-occurrence frequency of the different set of words. The method discovers, by the RRU, reasoning paths from the graph for word pairs by mapping word pairs from the set of words into a source word and a destination word and finding the reasoning paths therebetween. The method predicts word co-occurrence frequency using the reasoning paths. The method updates, responsive to the word co-occurrence frequency, model parameters of the RRU until a difference between a predicted and true word occurrence are less than a threshold amount to provide a trained RRU.