Your interests/skills: Deep Learning, Natural Language Processing, Knowledge Bases

Step 1: Implement a BERT model with two heads: one for mention detection and one to generate mention embeddings. Find a proper way to embed entities from a knowledge base (e.g. Wikipedia2vec). Then perform candidate selection using similarity search between mention embeddings and entity embeddings.

Step 2: Look for ways to improve the performance and especially the time consumption of your system.

Step 3: Thoroughly evaluate your system. Compare it with at least two other approaches (one state of the art system and one baseline) on different benchmarks. Analyse your results in detail.