Master's Thesis Robin Otto

Transfer Learning for Named Entity Linking with Deep Learning

The involved challenges in Named Entity Linking are to differentiate between words in general and afterwards to link the occurrences of the same entities.
For example, in “Paris is the capital of France. I enjoy visiting the city”, the challenge would be to associate the word “Paris” with the location, as well as to link the occurrences of “Paris” and “city” to the same entity.

Current supervised neural networks designed for Named Entity Linking use publicly available datasets for training and testing. They address the problem of automated information extraction in general and try to maximize the accuracy and speed of these networks. This thesis focuses especially on the aspect to apply Transfer Learning to networks trained for Named Entity Linking.
If this approach succeeds, one could, for example, apply these networks to enterprises and their internal datasets. An example could be the integration of such networks in order to extract the skill base of the employees from CVs, commit histories, etc. Another instance for this approach would be the application on legal documents, like terms of service.

In order to make automated information extraction possible for datasets not publicly available, it is necessary to compare the state-of-the-art supervised deep learning neural networks coping with such problems.

Afterwards, rather than using the private datasets for training, we want to integrate and apply the best state-of-the-art neural networks by using the concept of Transfer Learning. Furthermore, we aim at evaluating the different types of Transfer Learning to determine the most beneficial technique for this particular problem. The expected result is to gain insight about the accuracy of state-of-the-art algorithms trained on publicly available datasets. In addition to this, the goal is a neural network that is capable of extracting information from private datasets. However, we assume this network will provide less accuracy.