Learning a semantic database from unstructured text

In this paper, we aim to learn a semantic database given a text corpus. Specifically, we focus on predicting whether or not a pair of entities are related by the hypernym relation, also known as the 'is-a' or 'type-of' relation. We learn a neural network model for this task. The model is given as input a description of the words and the context from the text corpus in which a pair of nouns (entities) occur. In particular, among other things the description includes pre-trained embeddings of the words. We show that the model is able to predict hypernym noun pairs even though the dataset includes many incorrectly labeled noun pairs. Finally, we suggest ways to improve the dataset and the method.