What does this research mean for the field?

For training sentence-BERT models in dense retrieval and reranking tasks, multiple negatives ranking loss outperforms softmax and triplet losses, and system fusions significantly improve retrieval performance. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To enhance performance in dense retrieval and reranking tasks utilizing sentence-BERT and various loss functions.

April 1, 2026Open Access

ditlab at the NTCIR-17 Transfer Task

Key Points

To enhance performance in dense retrieval and reranking tasks utilizing sentence-BERT and various loss functions.
Trained sentence-BERT on a Japanese version of the mMARCO dataset.
Executed both dense retrieval and dense reranking subtasks.
Compared model performance based on different loss functions: softmax, triplet, and multiple negatives ranking losses.
Multiple negatives ranking loss outperformed other loss functions in both subtasks.
System fusions led to significant performance improvements, particularly in the retrieval task.

Abstract

The ditlab team participated in the Transfer task composed of dense retrieval and dense reranking subtasks. We trained sentence-BERT by using a Japanese version of mMARCO dataset and commonly used for both subtasks. We compared three types of models that were trained according to three types of losses: softmax, triplet, multiple negatives ranking losses. The results show that the multiple negatives ranking loss was the best for both subtasks. In addition, system fusions significantly improved the performance especially for the retrieval task.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper