Add regressions for MS MARCO v2 doc segmented uniCOIL (#1850)

Encoding title/segment this time - calling it "v2" to distinguish from segment-only encoding.
castorini · Apr 21, 2022 · 5e0437b · 5e0437b
1 parent 115af94
commit 5e0437b
Show file tree

Hide file tree

Showing 21 changed files with 1,214 additions and 18 deletions.
diff --git a/README.md b/README.md
@@ -87,15 +87,15 @@ See individual pages for details!
 + Regressions for MS MARCO (V2) Document Ranking:
   + Unsupervised lexical, complete doc: [baselines](docs/regressions-msmarco-v2-doc.md), [doc2query-T5](docs/regressions-msmarco-v2-doc-d2q-t5.md)
   + Unsupervised lexical, segmented doc: [baselines](docs/regressions-msmarco-v2-doc-segmented.md), [doc2query-T5](docs/regressions-msmarco-v2-doc-segmented-d2q-t5.md)
-  + Learned sparse lexical: [uniCOIL noexp zero-shot](docs/regressions-msmarco-v2-doc-segmented-unicoil-noexp-0shot.md), [uniCOIL with d2q-T5 zero-shot](docs/regressions-msmarco-v2-doc-segmented-unicoil-0shot.md)
+  + Learned sparse lexical: [uniCOIL noexp zero-shot](docs/regressions-msmarco-v2-doc-segmented-unicoil-noexp-0shot-v2.md), [uniCOIL with d2q-T5 zero-shot](docs/regressions-msmarco-v2-doc-segmented-unicoil-0shot-v2.md)
 + Regressions for TREC 2021 Deep Learning Track, Passage Ranking:
   + Unsupervised lexical, original corpus: [baselines](docs/regressions-dl21-passage.md), [doc2query-T5](docs/regressions-dl21-passage-d2q-t5.md)
   + Unsupervised lexical, augmented corpus: [baselines](docs/regressions-dl21-passage-augmented.md), [doc2query-T5](docs/regressions-dl21-passage-augmented-d2q-t5.md)
   + Learned sparse lexical: [uniCOIL noexp zero-shot](docs/regressions-dl21-passage-unicoil-noexp-0shot.md), [uniCOIL with d2q-T5 zero-shot](docs/regressions-dl21-passage-unicoil-0shot.md)
 + Regressions for TREC 2021 Deep Learning Track, Document Ranking:
   + Unsupervised lexical, complete doc: [baselines](docs/regressions-dl21-doc.md), [doc2query-T5](docs/regressions-dl21-doc-d2q-t5.md)
   + Unsupervised lexical, segmented doc: [baselines](docs/regressions-dl21-doc-segmented.md), [doc2query-T5](docs/regressions-dl21-doc-segmented-d2q-t5.md)
-  + Learned sparse lexical: [uniCOIL noexp zero-shot](docs/regressions-dl21-doc-segmented-unicoil-noexp-0shot.md), [uniCOIL with d2q-T5 zero-shot](docs/regressions-dl21-doc-segmented-unicoil-0shot.md)
+  + Learned sparse lexical: [uniCOIL noexp zero-shot](docs/regressions-dl21-doc-segmented-unicoil-noexp-0shot-v2.md), [uniCOIL with d2q-T5 zero-shot](docs/regressions-dl21-doc-segmented-unicoil-0shot-v2.md)
 + Regressions for TREC News Tracks (Background Linking Task): [2018](docs/regressions-backgroundlinking18.md), [2019](docs/regressions-backgroundlinking19.md), [2020](docs/regressions-backgroundlinking20.md)
 + Regressions for [FEVER Fact Verification](docs/regressions-fever.md)
 + Regressions for [NTCIR-8 ACLIA (IR4QA subtask, Monolingual Chinese)](docs/regressions-ntcir8-zh.md)

diff --git a/docs/regressions-dl21-doc-segmented-unicoil-0shot-v2.md b/docs/regressions-dl21-doc-segmented-unicoil-0shot-v2.md
@@ -0,0 +1,127 @@
+# Anserini Regressions: TREC 2021 Deep Learning Track (Document)
+
+**Model**: uniCOIL (with doc2query-T5 expansions) zero-shot on segmented documents (title/segment encoding)
+
+This page describes experiments, integrated into Anserini's regression testing framework, on the [TREC 2021 Deep Learning Track document ranking task](https://trec.nist.gov/data/deep2021.html) using the MS MARCO V2 _segmented_ document collection.
+Here, we cover experiments with the uniCOIL model trained on the MS MARCO V1 passage ranking test collection, applied in a zero-shot manner, with doc2query-T5 expansions.
+
+The uniCOIL model is described in the following paper:
+
+> Jimmy Lin and Xueguang Ma. [A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques.](https://arxiv.org/abs/2106.14807) _arXiv:2106.14807_.
+
+**NOTE**: As an important detail, there is the question of what text we feed into the encoder to generate document representations.
+Initially, we fed only the segment text, but later we realized that prepending the title of the document improves effectiveness.
+This regression captures the latter title/segment encoding, which for clarity we call v2, distinguished from segment-only encoding, which is documented [here](regressions-dl21-doc-segmented-unicoil-0shot.md).
+The segment-only encoding results are deprecated and kept around primarily for archival purposes and ablation experiments.
+You probably don't want to use them.
+
+Note that the NIST relevance judgments provide far more relevant documents 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 V2 document collection, refer to [this page](experiments-msmarco-v2.md).
+
+The exact configurations for these regressions are stored in [this YAML file](../src/main/resources/regression/dl21-doc-segmented-unicoil-0shot-v2.yaml).
+Note that this page is automatically generated from [this template](../src/main/resources/docgen/templates/dl21-doc-segmented-unicoil-0shot-v2.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 dl21-doc-segmented-unicoil-0shot-v2
+```
+
+## Corpus
+
+Download, unpack, and prepare the corpus:
+
+```
+# Download
+wget https://rgw.cs.uwaterloo.ca/JIMMYLIN-bucket0/data/msmarco_v2_doc_segmented_unicoil_0shot_v2.tar -P collections/
+
+# Unpack
+tar -xvf collections/msmarco_v2_doc_segmented_unicoil_0shot_v2.tar -C collections/
+
+# Rename (indexer is expecting corpus under a slightly different name)
+mv collections/msmarco_v2_doc_segmented_unicoil_0shot_v2 collections/msmarco-v2-doc-segmented-unicoil-0shot-v2
+```
+
+To confirm, `msmarco_v2_doc_segmented_unicoil_0shot_v2.tar` is 72 GB and has an MD5 checksum of `c5639748c2cbad0152e10b0ebde3b804`.
+
+## Indexing
+
+Sample indexing command:
+
+```
+target/appassembler/bin/IndexCollection \
+  -collection JsonVectorCollection \
+  -input /path/to/msmarco-v2-doc-segmented-unicoil-0shot-v2 \
+  -index indexes/lucene-index.msmarco-v2-doc-segmented-unicoil-0shot-v2/ \
+  -generator DefaultLuceneDocumentGenerator \
+  -threads 18 -impact -pretokenized \
+  >& logs/log.msmarco-v2-doc-segmented-unicoil-0shot-v2 &
+```
+
+The path `/path/to/msmarco-v2-doc-segmented-unicoil-0shot/` 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 uniCOIL tokens.
+Upon completion, we should have an index with 124,131,414 documents.
+
+For additional details, see explanation of [common indexing options](common-indexing-options.md).
+
+## Retrieval
+
+Topics and qrels are stored in [`src/main/resources/topics-and-qrels/`](../src/main/resources/topics-and-qrels/).
+The regression experiments here evaluate on the 57 topics for which NIST has provided judgments as part of the TREC 2021 Deep Learning Track.
+The original data can be found [here](https://trec.nist.gov/data/deep2021.html).
+
+After indexing has completed, you should be able to perform retrieval as follows:
+
+```
+target/appassembler/bin/SearchCollection \
+  -index indexes/lucene-index.msmarco-v2-doc-segmented-unicoil-0shot-v2/ \
+  -topics src/main/resources/topics-and-qrels/topics.dl21.unicoil.0shot.tsv.gz \
+  -topicreader TsvInt \
+  -output runs/run.msmarco-v2-doc-segmented-unicoil-0shot-v2.unicoil-0shot.topics.dl21.unicoil.0shot.txt \
+  -hits 10000 -selectMaxPassage -selectMaxPassage.delimiter "#" -selectMaxPassage.hits 1000 -impact -pretokenized &
+```
+
+Evaluation can be performed using `trec_eval`:
+
+```
+tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m map src/main/resources/topics-and-qrels/qrels.dl21-doc.txt runs/run.msmarco-v2-doc-segmented-unicoil-0shot-v2.unicoil-0shot.topics.dl21.unicoil.0shot.txt
+tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.100 src/main/resources/topics-and-qrels/qrels.dl21-doc.txt runs/run.msmarco-v2-doc-segmented-unicoil-0shot-v2.unicoil-0shot.topics.dl21.unicoil.0shot.txt
+tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.1000 src/main/resources/topics-and-qrels/qrels.dl21-doc.txt runs/run.msmarco-v2-doc-segmented-unicoil-0shot-v2.unicoil-0shot.topics.dl21.unicoil.0shot.txt
+tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m recip_rank -c -m ndcg_cut.10 src/main/resources/topics-and-qrels/qrels.dl21-doc.txt runs/run.msmarco-v2-doc-segmented-unicoil-0shot-v2.unicoil-0shot.topics.dl21.unicoil.0shot.txt
+```
+
+## Effectiveness
+
+With the above commands, you should be able to reproduce the following results:
+
+| MAP@100                                                                                                      | uniCOIL (with doc2query-T5) zero-shot|
+|:-------------------------------------------------------------------------------------------------------------|-----------|
+| [DL21 (Doc)](https://microsoft.github.io/msmarco/TREC-Deep-Learning)                                         | 0.2718    |
+
+
+| MRR@100                                                                                                      | uniCOIL (with doc2query-T5) zero-shot|
+|:-------------------------------------------------------------------------------------------------------------|-----------|
+| [DL21 (Doc)](https://microsoft.github.io/msmarco/TREC-Deep-Learning)                                         | 0.9684    |
+
+
+| nDCG@10                                                                                                      | uniCOIL (with doc2query-T5) zero-shot|
+|:-------------------------------------------------------------------------------------------------------------|-----------|
+| [DL21 (Doc)](https://microsoft.github.io/msmarco/TREC-Deep-Learning)                                         | 0.6783    |
+
+
+| R@100                                                                                                        | uniCOIL (with doc2query-T5) zero-shot|
+|:-------------------------------------------------------------------------------------------------------------|-----------|
+| [DL21 (Doc)](https://microsoft.github.io/msmarco/TREC-Deep-Learning)                                         | 0.3700    |
+
+
+| R@1000                                                                                                       | uniCOIL (with doc2query-T5) zero-shot|
+|:-------------------------------------------------------------------------------------------------------------|-----------|
+| [DL21 (Doc)](https://microsoft.github.io/msmarco/TREC-Deep-Learning)                                         | 0.7069    |
+
+This run roughly corresponds to run `p_unicoil0` submitted to the TREC 2021 Deep Learning Track under the "baseline" group.
+The difference is that here we are using pre-encoded queries, whereas the official submission performed query encoding on the fly.
+
+## Reproduction Log[*](reproducibility.md)
+
+To add to this reproduction log, modify [this template](../src/main/resources/docgen/templates/dl21-doc-segmented-unicoil-0shot-v2.template) and run `bin/build.sh` to rebuild the documentation.
diff --git a/docs/regressions-dl21-doc-segmented-unicoil-0shot.md b/docs/regressions-dl21-doc-segmented-unicoil-0shot.md
@@ -1,6 +1,6 @@
 # Anserini Regressions: TREC 2021 Deep Learning Track (Document)
 
-**Model**: uniCOIL (with doc2query-T5 expansions) zero-shot on segmented documents
+**Model**: uniCOIL (with doc2query-T5 expansions) zero-shot on segmented documents (segment-only encoding) - _Deprecated_, see below
 
 This page describes experiments, integrated into Anserini's regression testing framework, on the [TREC 2021 Deep Learning Track document ranking task](https://trec.nist.gov/data/deep2021.html) using the MS MARCO V2 _segmented_ document collection.
 Here, we cover experiments with the uniCOIL model trained on the MS MARCO V1 passage ranking test collection, applied in a zero-shot manner, with doc2query-T5 expansions.
@@ -9,6 +9,11 @@ The uniCOIL model is described in the following paper:
 
 > Jimmy Lin and Xueguang Ma. [A Few Brief Notes on DeepImpact, COIL, and a Conceptual Framework for Information Retrieval Techniques.](https://arxiv.org/abs/2106.14807) _arXiv:2106.14807_.
 
+**NOTE**: As an important detail, there is the question of what text we feed into the encoder to generate document representations.
+Initially, we fed only the segment text, but later we realized that prepending the title of the document improves effectiveness.
+This regression captures segment-only encoding and is kept around primarily for archival purposes; you probably don't want to use this one unless you're running ablation experiments.
+The version that uses title/segment encoding can be found [here](regressions-dl21-doc-segmented-unicoil-0shot-v2.md).
+
 Note that the NIST relevance judgments provide far more relevant documents 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 V2 document collection, refer to [this page](experiments-msmarco-v2.md).