Topics
This is another approach in Neural IR, where there is direct interaction with query terms and document terms. When we say terms, we mean their tokenized representations which can be word or subword level.
Example
Each query term interacts with every document term, producing a matrix which can be fed to some MLP to get a score.
- Example:
monoBERT - Advantages:
- Can selectively focus on relevant information
- More effective than representation-based similarity
- Disadvantages:
- Computationally expensive during inference
- Hard to use in production