The nature of the data I have is not arranged; however, I'm trying to extract the appropriate sentences for each query as a sample for ground truth. Also, the most critical problem is that I use the BERT model, which searches for the top 500 sentences in which the similarity value is greater than the 0.5 threshold. So the result is disorganized. For example, I got for a query
5 for the relevant, 495 for the irrelevant, and 6 for the Total relevant
How can I avoid the irrelevant number of sentences in the result?