Entity Similarity through word embedding and Named Entity Recognition using word vectors

Details

Supervisors

Saerens, Marco ; Françoisse, Kevin

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques

Abstract

The recently introduced Word2vec and GloVe models are efficient methods to build quality word embeddings. In this thesis we investigated and implemented both models to compute Entity Similarity. We assessed two different approaches. The first one considers entities as normal words to include them in the final word embedding. The second approach uses the already built word embedding to project the entities inside it. A complete data enrichment pipeline was also designed to increase the data quality and improve the final results. Currently, Named Entity Recognition state-of-the-art uses Conditional Random Fields. We built a word vector based Multilayer Bidirectional Long-Short-Term-Memory Recurrent Neural Network using the deep learning framework Tensorflow. Providing only few feature and archi- tecture engineering the model achieved near to state-of-the-art results. Considering a few more optimizations explained in the thesis, Recurrent Neural Network using word vector could become the next state-of-the-art method.