Topics

Comprehensive interaction between all elements of query and document

In this kind of Neural IR, the model takes all query and document terms, and gives a score. This is more accurate than bi-encoders. The BERT model can be trained for such modelling, by feeding in query-document pairs. Such models are also called as cross-encoders.

  • Advantages:
    • Captures complex relationships between query and document
  • Disadvantages:
    • High computational cost
    • Difficult to use in production