Deep Neural Networks for Disambiguation of Words and Named Entities

In natural-language understanding, addressing word and name ambiguity is a key issue nowadays. Some words and name mentions could refer to more than hundreds of concepts or entities. The ambiguity is resolved using the context words surrounding the target word or name mention. Deep learning based approaches take advantage of large annotated corpora like Wikipedia in order to build representations that encode the semantic meaning of the context. In the other hand, graph-based approaches exploit the relations among concepts and entities in the Knowledge Bases such as WordNet or BabelNet. Graph-based systems look for the coherence among concepts and entity candidates given a context.
The thesis proposed here will explore the optimal method for combining representation of context given by deep neural networks and graph-based approaches. Giving as a result a model that should disambiguate word and name mentions all together.