Order Embeddings for Supervised Hypernymy Detection

Resumen

In this work we present a supervised approach to partially order word embeddings, through a learned order embedding, and we apply it in supervised hypernymy detection. We use neural network as an order embedding to map general purpose word embeddings to a partially ordered vector set. The mapping is trained using positive and negative instances of the relationship. We consider two alternatives to deal with compound terms: a character based embedding of an underscored version of the terms, and a convolutional neural network that consumes the word embedding of each term. We show that this distributional approach presents interesting results in comparison to other distributional and path-based approaches. In addition, we observe still good behavior on different sized portions of the training data. This may suggest an interesting generalization capability.