Entities are real world objects such as persons, places, or events that appear in natural language text such as web pages, news, and journals. Entity Linking, a nascent field in Natural Language Processing, is the task of linking entities in text to their referent entries in a Knowledge Base (KB), which is a repository of information such as Wikipedia. There’s a huge application of entity linking in automatic knowledge base population, prevention of identity crimes, etc. It can also provide background information about unfamiliar concepts during document reading, rendering a smooth and joyful reading experience without frequent “context switch”.
This thesis taps into the power of convolutional neural network, and proposes an architecture that makes use of deep learning layers, convolution, max pooling, and fully-connected neurons with dropout to approach the problem of entity linking. Based on a pre-trained word2vec word embedding and another ad-hoc trained layer of word representation, we were able to outperform previous state-of-art models, which handcrafted a large number of features, by a modest margin.
Visualization of the neural network is also provided in order to understand what happens under the hood. Our experiment showed that it clearly captured the desired features, indicating the efficacy of neural network in dealing with entity linking.