- Use incoming link count as a feature
- Classify context + entity into categories like person/organization/location
- If x is a redirect to y, add a context link/surface from x to y
- Try different neighbour files: (Full neighbourhood, only reciprocal neighbourhood, triangle neighbourhood)
- Extend language models with the contexts of links linking to the entity
- Use categories to draw links between entities, e.g.: both Bayern Munich and Borussia Dortmund are German football clubs (--> draw link between them)
- Small performance idea: Only parse language models on demand in
readLanguageModels - Use surface link occurrence (how often does a surface occur at all), not only surface link probability to filter out surfaces
- Implement random walks on windows to reduce neighbourhood size
- Full test matrix:
- Use/don't use overlapping trie hits for training
- Use different training data filter strategies (remove candidate, remove entire group)