BERT has been used lately as another step in ranking the retrieved documents from an IR engine. The model is usually trained on a set of relevant and irrelevant documents but what we could do instead is retrieve the top 100 documents and rerank based on cosine similarity using SciBERT.
nsorros
commented
4 years ago
BERT has been used lately as another step in ranking the retrieved documents from an IR engine. The model is usually trained on a set of relevant and irrelevant documents but what we could do instead is retrieve the top 100 documents and rerank based on cosine similarity using SciBERT.