Unsupervised Induction of Relation within a Reconstruction Minimization Framework

Day - Time: 03 February 2015, h.11:30
Place: Area della Ricerca CNR di Pisa - Room: A-27

Andrea Esuli


Relation extraction is the task that aims to extract structured information between elements from unstructured sources such as natural language text. This comes in help when one wants to create a knowledge graph in which the nodes are entities (person, organization, event), and the edges that connects the nodes are the relations between entities (works_for, attend, CEO_of). In this work we approached the task of relation extraction using a novel, unsupervised model, to induce relations starting from a raw text. The method primarily consists of two components, (i) an encoding component which predicts the relation between two found entities given a rich set of syntactic and semantic features, and (ii) a reconstruction component expressed as a tensor factorization model which relies on the relations predicted by (i) to predict back the argument fillers (as word embeddings). The entire model can be seen as an autoencoder that instead of learning the model parameters from labeled data, learns the best label that allows the reconstruction component to reconstruct the input.