Skip to content

Latest commit

 

History

History
212 lines (160 loc) · 11.8 KB

regressions-msmarco-doc.md

File metadata and controls

212 lines (160 loc) · 11.8 KB
Raw

Anserini Regressions: MS MARCO Document Ranking

Models: various bag-of-words approaches on complete documents

This page documents regression experiments on the MS MARCO document ranking task, which is integrated into Anserini's regression testing framework. Note that there are four different bag-of-words regression conditions for this task, and this page describes the following:

  • Indexing Condition: each MS MARCO document is treated as a unit of indexing
  • Expansion Condition: none

All four conditions are described in detail here, in the context of doc2query-T5.

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.

Note that in November 2021 we discovered issues in our regression tests, documented here. As a result, we have had to rebuild all our regressions from the raw corpus. These new versions yield end-to-end scores that are slightly different, so if numbers reported in a paper do not exactly match the numbers here, this may be the reason.

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-doc

Indexing

Typical indexing command:

target/appassembler/bin/IndexCollection \
  -collection JsonCollection \
  -input /path/to/msmarco-doc \
  -generator DefaultLuceneDocumentGenerator \
  -index indexes/lucene-index.msmarco-doc/ \
  -threads 7 -storePositions -storeDocvectors -storeRaw \
  >& logs/log.msmarco-doc &

The directory /path/to/msmarco-doc/ should be a directory containing the document corpus in Anserini's jsonl format. See this page for how to prepare the corpus.

For additional details, see explanation of common indexing options.

Retrieval

Topics and qrels are stored here, which is linked to the Anserini repo as a submodule. The regression experiments here evaluate on the 5193 dev set questions.

After indexing has completed, you should be able to perform retrieval as follows:

target/appassembler/bin/SearchCollection \
  -index indexes/lucene-index.msmarco-doc/ \
  -topics tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt \
  -bm25 &

target/appassembler/bin/SearchCollection \
  -index indexes/lucene-index.msmarco-doc/ \
  -topics tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt \
  -bm25 -bm25.k1 3.44 -bm25.b 0.87 &

target/appassembler/bin/SearchCollection \
  -index indexes/lucene-index.msmarco-doc/ \
  -topics tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-doc.bm25-tuned2.topics.msmarco-doc.dev.txt \
  -bm25 -bm25.k1 4.46 -bm25.b 0.82 &

Evaluation can be performed using trec_eval:

target/appassembler/bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -M 100 -m recip_rank tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt

target/appassembler/bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -M 100 -m recip_rank tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt

target/appassembler/bin/trec_eval -c -m map tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned2.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -M 100 -m recip_rank tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned2.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned2.topics.msmarco-doc.dev.txt
target/appassembler/bin/trec_eval -c -m recall.1000 tools/topics-and-qrels/qrels.msmarco-doc.dev.txt runs/run.msmarco-doc.bm25-tuned2.topics.msmarco-doc.dev.txt

Effectiveness

With the above commands, you should be able to reproduce the following results:

AP@1000 BM25 (default) BM25 (tuned) BM25 (tuned2)
MS MARCO Doc: Dev 0.2305 0.2784 0.2773
RR@100 BM25 (default) BM25 (tuned) BM25 (tuned2)
MS MARCO Doc: Dev 0.2299 0.2778 0.2767
R@100 BM25 (default) BM25 (tuned) BM25 (tuned2)
MS MARCO Doc: Dev 0.7281 0.8069 0.8070
R@1000 BM25 (default) BM25 (tuned) BM25 (tuned2)
MS MARCO Doc: Dev 0.8856 0.9324 0.9357

Explanation of settings:

  • The setting "default" refers the default BM25 settings of k1=0.9, b=0.4.
  • The setting "tuned" refers to k1=3.44, b=0.87, tuned in 2019/06 using the MS MARCO document sparse judgments to optimize for MAP and used for TREC 2019 Deep Learning Track baseline runs.
  • The setting "tuned2" refers to k1=4.46, b=0.82, tuned in 2020/12 using the MS MARCO document sparse judgments to optimize for recall@100 (i.e., for first-stage retrieval); see this page additional details.

See this page for more details on tuning.

In these runs, we are retrieving the top 1000 hits for each query and using trec_eval to evaluate all 1000 hits. This lets us measure R@100 and R@1000; the latter is particularly important when these runs are used as first-stage retrieval. Beware, an official MS MARCO document ranking task leaderboard submission comprises only 100 hits per query. See this page for details on Anserini baseline runs that were submitted to the official leaderboard.

Additional Implementation Details

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 in some cases they correspond to official MS MARCO leaderboard submissions.

To generate an MS MARCO submission with the BM25 default parameters, corresponding to "BM25 (default)" above:

$ sh target/appassembler/bin/SearchMsmarco -hits 100 -k1 0.9 -b 0.4 -threads 9 \
    -index indexes/lucene-index.msmarco-doc/ \
    -queries tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
    -output runs/run.msmarco-doc.bm25-default.txt

$ python tools/scripts/msmarco/msmarco_doc_eval.py \
    --judgments tools/topics-and-qrels/qrels.msmarco-doc.dev.txt \
    --run runs/run.msmarco-doc.bm25-default.txt

#####################
MRR @100: 0.23005723505603573
QueriesRanked: 5193
#####################

This run corresponds to the MS MARCO document ranking leaderboard entry "Anserini's BM25, default parameters (k1=0.9, b=0.4)" dated 2020/08/16, and is reported in the Lin et al. (SIGIR 2021) Pyserini paper.

To generate an MS MARCO submission with the BM25 tuned parameters, corresponding to "BM25 (tuned)" above:

$ sh target/appassembler/bin/SearchMsmarco -hits 100 -k1 4.46 -b 0.82 -threads 9 \
   -index indexes/lucene-index.msmarco-doc/ \
   -queries tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
   -output runs/run.msmarco-doc.bm25-tuned.txt

$ python tools/scripts/msmarco/msmarco_doc_eval.py \
   --judgments tools/topics-and-qrels/qrels.msmarco-doc.dev.txt \
   --run runs/run.msmarco-doc.bm25-tuned.txt

#####################
MRR @100: 0.2770296928568702
QueriesRanked: 5193
#####################

This run was not submitted to the MS MARCO document ranking leaderboard, but is reported in the Lin et al. (SIGIR 2021) Pyserini paper.

As of February 2022, following resolution of #1721, BM25 runs for the MS MARCO leaderboard can be generated with the commands below. For default parameters (k1=0.9, b=0.4):

$ sh target/appassembler/bin/SearchCollection \
    -index indexes/lucene-index.msmarco-doc/ \
    -topics tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
    -topicreader TsvInt \
    -output runs/run.msmarco-doc.bm25-default.txt \
    -format msmarco \
    -bm25 -hits 100

$ python tools/scripts/msmarco/msmarco_doc_eval.py \
    --judgments tools/topics-and-qrels/qrels.msmarco-doc.dev.txt \
    --run runs/run.msmarco-doc.bm25-default.txt

#####################
MRR @100: 0.22994387925437856
QueriesRanked: 5193
#####################

For tuned parameters (k1=4.46, b=0.82):

$ sh target/appassembler/bin/SearchCollection \
    -index indexes/lucene-index.msmarco-doc/ \
    -topics tools/topics-and-qrels/topics.msmarco-doc.dev.txt \
    -topicreader TsvInt \
    -output runs/run.msmarco-doc.bm25-tuned.txt \
    -format msmarco \
    -bm25 -bm25.k1 4.46 -bm25.b 0.82 -hits 100

$ python tools/scripts/msmarco/msmarco_doc_eval.py \
    --judgments tools/topics-and-qrels/qrels.msmarco-doc.dev.txt \
    --run runs/run.msmarco-doc.bm25-tuned.txt

#####################
MRR @100: 0.2766351807440808
QueriesRanked: 5193
#####################

Note that the resolution of #1721 did slightly change the results, since we corrected underlying issues with data preparation.