Contextualized Offline Relevance Weighting

This repository contains some resources of our paper:

• Contextualized Offline Relevance Weighting for Efficient and Effective Neural Retrieval. In SIGIR 2021.

Framework

1. Offline System

2. Online System

Resources

1. Selected neighbour documents^*:

   • TREC-19-DL-Passage: Download
   • TREC-20-DL-Passage: Download
   • TREC-19-DL-Document: Download
   • TREC-20-DL-Document: Download
   • Format: seed_doc_id \t neighbour_doc_id \t rank \t recall_frequency \t best_bm25_score \n

2. The final relevance scores of all recalled neighbour documents^*:

   • TREC-19-DL-Passage: Download
   • TREC-20-DL-Passage: Download
   • TREC-19-DL-Document: Download
   • TREC-20-DL-Document: Download
   • Format: query_id \t neighbour_doc_id \t rank \t rel_score \n

   ^* Note that the neighbour documents here belong to the top-100 (s=100) seed documents by BM25, and the neighbour document for document ranking task refers to the passages segmented from documents using the script convert_msmarco_passages_doc_to_anserini.py in docTTTTTquery repo.

3. Retrieval run files (top-1,000):

   • TREC-19-DL-Passage: BM25, Ours(s=30), Ours(s=50).
   • TREC-20-DL-Passage: BM25, Ours(s=30), Ours(s=50).
   • TREC-19-DL-Document: BM25, Ours(s=30), Ours(s=50).
   • TREC-20-DL-Document: BM25, Ours(s=30), Ours(s=50).
   • Format: query_id Q0 doc_id rank score run_name

For more details of our method and experiments, please refer to our paper.

Citation

If you find our paper/resources useful, please cite:

@inproceedings{Chen2021_sigir,
 author = {Xuanang Chen and
           Ben He and
           Kai Hui and
           Yiran Wang and
           Le Sun and
           Yingfei Sun},
 title = {Contextualized Offline Relevance Weighting for Efficient and Effective Neural Retrieval},
 booktitle = {Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval},
 pages = {1617-1621},
 year = {2021},
}