Skip to content

Latest commit

 

History

History
145 lines (105 loc) · 10.3 KB

regressions-dl20-passage.splade-pp-ed.cached.md

File metadata and controls

145 lines (105 loc) · 10.3 KB
Raw

Anserini Regressions: TREC 2020 Deep Learning Track (Passage)

Model: SPLADE++ CoCondenser-EnsembleDistil (using cached queries)

This page describes regression experiments, integrated into Anserini's regression testing framework, using the SPLADE++ CoCondenser-EnsembleDistil model on the TREC 2020 Deep Learning Track passage ranking task, as described in the following paper:

Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stéphane Clinchant. From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective. Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2353–2359.

In these experiments, we are using cached queries (i.e., cached results of query encoding).

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). For additional instructions on working with MS MARCO passage collection, refer to this page.

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 and then run bin/build.sh to rebuild the documentation.

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 dl20-passage.splade-pp-ed.cached

We make available a version of the MS MARCO Passage Corpus that has already been encoded with SPLADE++ CoCondenser-EnsembleDistil.

From any machine, the following command will download the corpus and perform the complete regression, end to end:

python src/main/python/run_regression.py --download --index --verify --search --regression dl20-passage.splade-pp-ed.cached

The run_regression.py script automates the following steps, but if you want to perform each step manually, simply copy/paste from the commands below and you'll obtain the same regression results.

Corpus Download

Download the corpus and unpack into collections/:

wget https://rgw.cs.uwaterloo.ca/pyserini/data/msmarco-passage-splade-pp-ed.tar -P collections/
tar xvf collections/msmarco-passage-splade-pp-ed.tar -C collections/

To confirm, msmarco-passage-splade-pp-ed.tar is 4.2 GB and has MD5 checksum e489133bdc54ee1e7c62a32aa582bc77. With the corpus downloaded, the following command will perform the remaining steps below:

python src/main/python/run_regression.py --index --verify --search --regression dl20-passage.splade-pp-ed.cached \
  --corpus-path collections/msmarco-passage-splade-pp-ed

Indexing

Sample indexing command:

bin/run.sh io.anserini.index.IndexCollection \
  -threads 16 \
  -collection JsonVectorCollection \
  -input /path/to/msmarco-passage-splade-pp-ed \
  -generator DefaultLuceneDocumentGenerator \
  -index indexes/lucene-inverted.msmarco-v1-passage.splade-pp-ed/ \
  -impact -pretokenized -storeDocvectors \
  >& logs/log.msmarco-passage-splade-pp-ed &

The path /path/to/msmarco-passage-splade-pp-ed/ should point to the corpus downloaded above.

The important indexing options to note here are -impact -pretokenized: the first tells Anserini not to encode BM25 doclengths into Lucene's norms (which is the default) and the second option says not to apply any additional tokenization on the SPLADE-distil CoCodenser Medium tokens. Upon completion, we should have an index with 8,841,823 documents.

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 54 topics for which NIST has provided judgments as part of the TREC 2020 Deep Learning Track. The original data can be found here.

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.splade-pp-ed/ \
  -topics tools/topics-and-qrels/topics.dl20.splade-pp-ed.tsv.gz \
  -topicReader TsvInt \
  -output runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached.topics.dl20.splade-pp-ed.txt \
  -impact -pretokenized &

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-inverted.msmarco-v1-passage.splade-pp-ed/ \
  -topics tools/topics-and-qrels/topics.dl20.splade-pp-ed.tsv.gz \
  -topicReader TsvInt \
  -output runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rm3.topics.dl20.splade-pp-ed.txt \
  -impact -pretokenized -rm3 &

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-inverted.msmarco-v1-passage.splade-pp-ed/ \
  -topics tools/topics-and-qrels/topics.dl20.splade-pp-ed.tsv.gz \
  -topicReader TsvInt \
  -output runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rocchio.topics.dl20.splade-pp-ed.txt \
  -impact -pretokenized -rocchio &

Evaluation can be performed using trec_eval:

bin/trec_eval -m map -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m ndcg_cut.10 -c tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.100 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.1000 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached.topics.dl20.splade-pp-ed.txt

bin/trec_eval -m map -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rm3.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m ndcg_cut.10 -c tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rm3.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.100 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rm3.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.1000 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rm3.topics.dl20.splade-pp-ed.txt

bin/trec_eval -m map -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rocchio.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m ndcg_cut.10 -c tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rocchio.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.100 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rocchio.topics.dl20.splade-pp-ed.txt
bin/trec_eval -m recall.1000 -c -l 2 tools/topics-and-qrels/qrels.dl20-passage.txt runs/run.msmarco-passage-splade-pp-ed.splade-pp-ed-cached+rocchio.topics.dl20.splade-pp-ed.txt

Effectiveness

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

AP@1000 SPLADE++ CoCondenser-EnsembleDistil +RM3 +Rocchio
DL20 (Passage) 0.5001 0.5096 0.5084
nDCG@10 SPLADE++ CoCondenser-EnsembleDistil +RM3 +Rocchio
DL20 (Passage) 0.7198 0.7131 0.7280
R@100 SPLADE++ CoCondenser-EnsembleDistil +RM3 +Rocchio
DL20 (Passage) 0.7653 0.7553 0.7704
R@1000 SPLADE++ CoCondenser-EnsembleDistil +RM3 +Rocchio
DL20 (Passage) 0.8995 0.9046 0.9069

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking). For computing nDCG, remember that we keep qrels of all relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the -l 2 option in trec_eval). The experimental results reported here are directly comparable to the results reported in the track overview paper.

Reproduction Log*

To add to this reproduction log, modify this template and run bin/build.sh to rebuild the documentation.