search

nyu-dl / dl4marco-bert

BSD 3-Clause "New" or "Revised" License
476 stars 87 forks source link

What is the most popular loss for doc ranking and What is the most popular loss for text similarity? #28

Closed guotong1988 closed 5 years ago

guotong1988 commented 5 years ago

Based on my understanding, both the doc ranking task and text similarity task take sentence pairs as model input. We use different loss to get better result for each of them. Thank you very much.

rodrigonogueira4 commented 5 years ago

I think it depends on the problem. Margin loss and cross-entropy loss are popular choices for document ranking. This article provides a nice overview: https://www.nowpublishers.com/article/Details/INR-016