Models: BM25 with doc2query-T5 expansions
This page documents regression experiments on the MS MARCO passage ranking task with BM25 on docTTTTTquery (also called doc2query-T5) expansions, as proposed in the following paper:
Rodrigo Nogueira and Jimmy Lin. From doc2query to docTTTTTquery. December 2019.
These experiments are integrated into Anserini's regression testing framework.
The exact configurations for these regressions are stored in this YAML file. Note that this page is automatically generated from this template as part of Anserini's regression pipeline, so do not modify this page directly; modify the template instead.
From one of our Waterloo servers (e.g., orca), the following command will perform the complete regression, end to end:
python src/main/python/run_regression.py --index --verify --search --regression msmarco-v1-passage.docTTTTTquery
Typical indexing command:
bin/run.sh io.anserini.index.IndexCollection \
-threads 18 \
-collection JsonCollection \
-input /path/to/msmarco-passage-docTTTTTquery \
-generator DefaultLuceneDocumentGenerator \
-index indexes/lucene-inverted.msmarco-v1-passage.docTTTTTquery/ \
-storePositions -storeDocvectors -storeRaw \
>& logs/log.msmarco-passage-docTTTTTquery &
The directory /path/to/msmarco-passage-docTTTTTquery should be a directory containing jsonl files containing the expanded passage collection.
Instructions in the docTTTTTquery repo explain how to perform this data preparation.
For additional details, see explanation of common indexing options.
Topics and qrels are stored here, which is linked to the Anserini repo as a submodule. The regression experiments here evaluate on the 6980 dev set questions; see this page for more details.
After indexing has completed, you should be able to perform retrieval as follows:
bin/run.sh io.anserini.search.SearchCollection \
-index indexes/lucene-inverted.msmarco-v1-passage.docTTTTTquery/ \
-topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-topicReader TsvInt \
-output runs/run.msmarco-passage-docTTTTTquery.bm25-default.topics.msmarco-passage.dev-subset.txt \
-bm25 &
bin/run.sh io.anserini.search.SearchCollection \
-index indexes/lucene-inverted.msmarco-v1-passage.docTTTTTquery/ \
-topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-topicReader TsvInt \
-output runs/run.msmarco-passage-docTTTTTquery.bm25-tuned.topics.msmarco-passage.dev-subset.txt \
-bm25 -bm25.k1 0.82 -bm25.b 0.68 &
bin/run.sh io.anserini.search.SearchCollection \
-index indexes/lucene-inverted.msmarco-v1-passage.docTTTTTquery/ \
-topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-topicReader TsvInt \
-output runs/run.msmarco-passage-docTTTTTquery.bm25-tuned2.topics.msmarco-passage.dev-subset.txt \
-bm25 -bm25.k1 2.18 -bm25.b 0.86 &
Evaluation can be performed using trec_eval:
bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-default.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -M 10 -m recip_rank tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-default.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-default.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-default.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -M 10 -m recip_rank tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned2.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -M 10 -m recip_rank tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned2.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned2.topics.msmarco-passage.dev-subset.txt
bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.bm25-tuned2.topics.msmarco-passage.dev-subset.txt
With the above commands, you should be able to reproduce the following results:
| AP@1000 | BM25 (default) | BM25 (tuned) | BM25 (tuned2) |
|---|---|---|---|
| MS MARCO Passage: Dev | 0.2805 | 0.2850 | 0.2893 |
| RR@10 | BM25 (default) | BM25 (tuned) | BM25 (tuned2) |
| MS MARCO Passage: Dev | 0.2723 | 0.2768 | 0.2816 |
| R@100 | BM25 (default) | BM25 (tuned) | BM25 (tuned2) |
| MS MARCO Passage: Dev | 0.8192 | 0.8190 | 0.8277 |
| R@1000 | BM25 (default) | BM25 (tuned) | BM25 (tuned2) |
| MS MARCO Passage: Dev | 0.9470 | 0.9471 | 0.9506 |
Explanation of settings:
- The setting "default" refers the default BM25 settings of
k1=0.9,b=0.4. - The setting "tuned" refers to
k1=0.82,b=0.68, tuned on on the original passages, as described in this page. - The setting "tuned2" refers to
k1=2.18,b=0.86, tuned to optimize for recall@1000 directly on the expanded passages (in 2020/12); this is the configuration reported in the Lin et al. (SIGIR 2021) Pyserini paper.
Note that prior to December 2021, runs generated with SearchCollection in the TREC format and then converted into the MS MARCO format give slightly different results from runs generated by SearchMsmarco directly in the MS MARCO format, due to tie-breaking effects.
This was fixed with #1458, which also introduced (intra-configuration) multi-threading.
As a result, SearchMsmarco has been deprecated and replaced by SearchCollection; both have been verified to generate identical output.
The commands below have been retained for historical reasons only, since they correspond to previously published results.
The following command generates with SearchMsmarco the run denoted "BM25 (tuned)" above (k1=0.82, b=0.68), which corresponds to the entry "docTTTTTquery" dated 2019/11/27 on the MS MARCO Passage Ranking Leaderboard:
$ sh target/appassembler/bin/SearchMsmarco -hits 1000 -threads 8 \
-index indexes/lucene-index.msmarco-passage-docTTTTTquery \
-queries tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-k1 0.82 -b 0.68 \
-output runs/run.msmarco-passage-docTTTTTquery.1
$ python tools/scripts/msmarco/msmarco_passage_eval.py \
tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.1
#####################
MRR @10: 0.27680089370991834
QueriesRanked: 6980
#####################This corresponds to the scores reported in the following paper:
Rodrigo Nogueira and Jimmy Lin. From doc2query to docTTTTTquery. December 2019.
And are identical to the scores reported in the docTTTTTquery repo.
The following command generates with SearchMsmarco the run denoted "BM25 (tuned2)" above (k1=2.18, b=0.86).
This does not correspond to an official leaderboard submission.
$ sh target/appassembler/bin/SearchMsmarco -hits 1000 -threads 8 \
-index indexes/lucene-index.msmarco-passage-docTTTTTquery \
-queries tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-k1 2.18 -b 0.86 \
-output runs/run.msmarco-passage-docTTTTTquery.2
$ python tools/scripts/msmarco/msmarco_passage_eval.py \
tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.2
#####################
MRR @10: 0.281560751807885
QueriesRanked: 6980
#####################This corresponds to the scores reported in the Lin et al. (SIGIR 2021) Pyserini paper.
As of February 2022, following resolution of #1730, BM25 runs for the MS MARCO leaderboard can be generated with the commands below.
For parameters k1=0.82, b=0.68:
$ sh target/appassembler/bin/SearchCollection \
-index indexes/lucene-index.msmarco-passage-docTTTTTquery/ \
-topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-topicreader TsvInt \
-output runs/run.msmarco-passage-docTTTTTquery.1 \
-format msmarco \
-bm25 -bm25.k1 0.82 -bm25.b 0.68
$ python tools/scripts/msmarco/msmarco_passage_eval.py \
tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.1
#####################
MRR @10: 0.27680089370991834
QueriesRanked: 6980
#####################
For parameters k1=2.18, b=0.86:
$ sh target/appassembler/bin/SearchCollection \
-index indexes/lucene-index.msmarco-passage-docTTTTTquery/ \
-topics tools/topics-and-qrels/topics.msmarco-passage.dev-subset.txt \
-topicreader TsvInt \
-output runs/run.msmarco-passage-docTTTTTquery.2 \
-format msmarco \
-bm25 -bm25.k1 2.18 -bm25.b 0.86
$ python tools/scripts/msmarco/msmarco_passage_eval.py \
tools/topics-and-qrels/qrels.msmarco-passage.dev-subset.txt runs/run.msmarco-passage-docTTTTTquery.2
#####################
MRR @10: 0.281560751807885
QueriesRanked: 6980
#####################
Note that the resolution of #1730 did not change the results.