par2hier: towards vector representations for hierarchical content

Word embeddings have received a lot of attention in the natural language processing area for their capabilities of capturing inner words semantics (e.g. word2vec, GloVe). The need of catching semantics at a higher and more abstract level led to creation of models like paragraph vectors for sentences and documents, seq2vec for biological sequences. In this paper we illustrate an approach for creating vector representations for hierarchical content where each node in the hierarchy is represented as a (recursive) function of its paragraph vector and the hierarchical vectors of its child nodes, computed via matrix factorization. We evaluate the effectiveness of our solution against flat paragraph vectors on a text categorization task obtaining significant µF1 improvements.