Skip to content

Latest commit

 

History

History
125 lines (96 loc) · 8.65 KB

regressions-dl22-passage.d2q-t5.md

File metadata and controls

125 lines (96 loc) · 8.65 KB
Raw

Anserini Regressions: TREC 2022 Deep Learning Track (Passage)

Models: BM25 with doc2query-T5 expansions on original passages

This page describes document expansion experiments (with doc2query-T5), integrated into Anserini's regression testing framework, on the TREC 2022 Deep Learning Track passage ranking task using the MS MARCO V2 passage corpus. For additional instructions on working with the MS MARCO V2 passage corpus, refer to this page.

Note that the NIST relevance judgments provide far more relevant passages per topic, unlike the "sparse" judgments provided by Microsoft (these are sometimes called "dense" judgments to emphasize this contrast).

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 dl22-passage.d2q-t5

Indexing

Typical indexing command:

bin/run.sh io.anserini.index.IndexCollection \
  -threads 24 \
  -collection MsMarcoV2PassageCollection \
  -input /path/to/msmarco-v2-passage-d2q-t5 \
  -generator DefaultLuceneDocumentGenerator \
  -index indexes/lucene-inverted.msmarco-v2-passage.d2q-t5/ \
  -storeRaw \
  >& logs/log.msmarco-v2-passage-d2q-t5 &

The value of -input should be a directory containing the compressed jsonl files that comprise the corpus. See this page for additional details.

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 76 topics for which NIST has provided judgments as part of the TREC 2022 Deep Learning Track.

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-v2-passage.d2q-t5/ \
  -topics tools/topics-and-qrels/topics.dl22.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt \
  -bm25 &

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-inverted.msmarco-v2-passage.d2q-t5/ \
  -topics tools/topics-and-qrels/topics.dl22.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt \
  -bm25 -rm3 -collection MsMarcoV2PassageCollection &

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-inverted.msmarco-v2-passage.d2q-t5/ \
  -topics tools/topics-and-qrels/topics.dl22.txt \
  -topicReader TsvInt \
  -output runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt \
  -bm25 -rocchio -collection MsMarcoV2PassageCollection &

Evaluation can be performed using trec_eval:

bin/trec_eval -c -M 100 -m map -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt
bin/trec_eval -c -M 100 -m recip_rank -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt
bin/trec_eval -c -m ndcg_cut.10 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt
bin/trec_eval -c -m recall.100 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt
bin/trec_eval -c -m recall.1000 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default.topics.dl22.txt

bin/trec_eval -c -M 100 -m map -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt
bin/trec_eval -c -M 100 -m recip_rank -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt
bin/trec_eval -c -m ndcg_cut.10 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt
bin/trec_eval -c -m recall.100 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt
bin/trec_eval -c -m recall.1000 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rm3.topics.dl22.txt

bin/trec_eval -c -M 100 -m map -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt
bin/trec_eval -c -M 100 -m recip_rank -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt
bin/trec_eval -c -m ndcg_cut.10 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt
bin/trec_eval -c -m recall.100 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt
bin/trec_eval -c -m recall.1000 -l 2 tools/topics-and-qrels/qrels.dl22-passage.txt runs/run.msmarco-v2-passage-d2q-t5.bm25-default+rocchio.topics.dl22.txt

Effectiveness

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

MAP@100 BM25 (default) +RM3 +Rocchio
DL22 (Passage) 0.0748 0.0872 0.0868
MRR@100 BM25 (default) +RM3 +Rocchio
DL22 (Passage) 0.4221 0.4319 0.4273
nDCG@10 BM25 (default) +RM3 +Rocchio
DL22 (Passage) 0.3599 0.3721 0.3712
R@100 BM25 (default) +RM3 +Rocchio
DL22 (Passage) 0.2259 0.2444 0.2434
R@1000 BM25 (default) +RM3 +Rocchio
DL22 (Passage) 0.4632 0.4964 0.5024

The "BM25 (default)" condition corresponds to the p_d2q_bm25 run submitted to the TREC 2022 Deep Learning Track as a "baseline". As of 91ec67, this correspondence was exact. That is, modulo the runtag and the number of hits, the output runfile should be identical. This can be confirmed as follows:

# Trim out the runtag:
cut -d ' ' -f 1-5 runs/p_d2q_bm25 > runs/p_d2q_bm25.submitted.cut

# Trim out the runtag and retain only top 100 hits per query:
python tools/scripts/trim_run_to_top_k.pl --k 100 --input runs/run.msmarco-v2-passage-d2q-t5.dl22.bm25-default --output runs/run.msmarco-v2-passage-d2q-t5.dl22.bm25-default.hits100
cut -d ' ' -f 1-5 runs/run.msmarco-v2-passage-d2q-t5.dl22.bm25-default.hits100 > runs/p_d2q_bm25.new.cut

# Verify the two runfiles are identical:
diff runs/p_d2q_bm25.submitted.cut runs/p_d2q_bm25.new.cut

The "BM25 + RM3" and "BM25 + Rocchio" conditions above correspond to the p_d2q_bm25rm3 run and the p_d2q_bm25rocchio run submitted to the TREC 2022 Deep Learning Track as "baselines". However, due to a60e84, the results are slightly different (because the underlying implementation changed).