Document network embedding aims at learning representations for a structured\ntext corpus i.e. when documents are linked to each other. Recent algorithms\nextend network embedding approaches by incorporating the text content\nassociated with the nodes in their formulations. In most cases, it is hard to\ninterpret the learned representations. Moreover, little importance is given to\nthe generalization to new documents that are not observed within the network.\nIn this paper, we propose an interpretable and inductive document network\nembedding method. We introduce a novel mechanism, the Topic-Word Attention\n(TWA), that generates document representations based on the interplay between\nword and topic representations. We train these word and topic vectors through\nour general model, Inductive Document Network Embedding (IDNE), by leveraging\nthe connections in the document network. Quantitative evaluations show that our\napproach achieves state-of-the-art performance on various networks and we\nqualitatively show that our model produces meaningful and interpretable\nrepresentations of the words, topics and documents.\n