• Use Multiprocessing for entity disambiguation
• Prevent LLM from residing in memory for the whole pipeline
• Set up Dockerfile, include model files
• Creating a test set for evaluation
• Set up Google docs document to explain our approaches (for presentation etc.)

Notes Maik & Floris

• Found a working LLM -> Zephyr 7B (outperforms LLaMa 70B model)
• Experiments with Entity Linking: words that are closer to the mention score more points
• Classifiying questions as 'yes/no' or 'entity'
  • First approach: look at 'where', 'what', 'who' etc. at the beginning of the question
  • Now working on: classifier trained on on 150K+ questions and type datasets
• Answer extraction
  • 'yes/no' question: pre-trained classifier from boolq dataset (question, context, answer), seems to work pretty well
  • 'entity' question: for now, return the first found entity in the answer that is not in the question
• Start creating a test set