Skip to content

Latest commit

 

History

History
99 lines (66 loc) · 6.86 KB

regressions-msmarco-doc.md

File metadata and controls

99 lines (66 loc) · 6.86 KB

Anserini: Regressions for MS MARCO (Document)

This page documents regression experiments for the MS MARCO Document Ranking Task, which is integrated into Anserini's regression testing framework. For more complete instructions on how to run end-to-end experiments, 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.

Indexing

Typical indexing command:

nohup sh target/appassembler/bin/IndexCollection -collection TrecCollection -input /path/to/msmarco-doc \
 -index lucene-index.msmarco-doc.pos+docvectors+rawdocs -generator LuceneDocumentGenerator -threads 1 \
 -storePositions -storeDocvectors -storeRawDocs >& log.msmarco-doc.pos+docvectors+rawdocs &

The directory /path/to/msmarco-doc/ should be a directory containing the official document collection (a single file), in TREC format.

For additional details, see explanation of common indexing options.

Retrieval

Topics and qrels are stored in src/main/resources/topics-and-qrels/. The regression experiments here evaluate on the 5193 dev set questions; see this page for more details.

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

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -output run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -rm3 -output run.msmarco-doc.bm25-default+rm3.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -axiom -axiom.deterministic -rerankCutoff 20 -output run.msmarco-doc.bm25-default+ax.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -bm25prf -output run.msmarco-doc.bm25-default+prf.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -bm25.k1 3.44 -bm25.b 0.87 -output run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -bm25.k1 3.44 -bm25.b 0.87 -rm3 -output run.msmarco-doc.bm25-tuned+rm3.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -bm25.k1 3.44 -bm25.b 0.87 -axiom -axiom.deterministic -rerankCutoff 20 -output run.msmarco-doc.bm25-tuned+ax.topics.msmarco-doc.dev.txt &

nohup target/appassembler/bin/SearchCollection -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
 -bm25 -bm25.k1 3.44 -bm25.b 0.87 -bm25prf -output run.msmarco-doc.bm25-tuned+prf.topics.msmarco-doc.dev.txt &

Evaluation can be performed using trec_eval:

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-default.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-default+rm3.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-default+ax.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-default+prf.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-tuned.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-tuned+rm3.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-tuned+ax.topics.msmarco-doc.dev.txt

eval/trec_eval.9.0.4/trec_eval -m map -c -m recall.1000 -c src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc.bm25-tuned+prf.topics.msmarco-doc.dev.txt

Effectiveness

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

MAP BM25 (Default) +RM3 +Ax +PRF BM25 (Tuned) +RM3 +Ax +PRF
MS MARCO Document Ranking: Dev Queries 0.2310 0.1632 0.1147 0.1357 0.2788 0.2289 0.1895 0.1559
R@1000 BM25 (Default) +RM3 +Ax +PRF BM25 (Tuned) +RM3 +Ax +PRF
MS MARCO Document Ranking: Dev Queries 0.8856 0.8785 0.8369 0.8471 0.9326 0.9320 0.9264 0.8758

The setting "default" refers the default BM25 settings of k1=0.9, b=0.4, while "tuned" refers to the tuned setting of k1=3.44, b=0.87. See this page for more details. Note that here we are using trec_eval to evaluate the top 1000 hits for each query; beware, the runs provided by MS MARCO organizers for reranking have only 100 hits per query.