Skip to content

Latest commit

 

History

History
172 lines (129 loc) · 6.59 KB

experiments-ance.md

File metadata and controls

172 lines (129 loc) · 6.59 KB
Raw

Pyserini: Reproducing ANCE Results

This guide provides instructions to reproduce the following dense retrieval work:

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, Arnold Overwijk. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval

Note that we often observe minor differences in scores between different computing environments (e.g., Linux vs. macOS). However, the differences usually appear in the fifth digit after the decimal point, and do not appear to be a cause for concern from a reproducibility perspective. Thus, while the scoring script provides results to much higher precision, we have intentionally rounded to four digits after the decimal point.

MS MARCO Passage

ANCE retrieval with brute-force index:

python -m pyserini.search.faiss \
  --index msmarco-v1-passage.ance \
  --topics msmarco-passage-dev-subset \
  --encoded-queries ance-msmarco-passage-dev-subset \
  --output runs/run.msmarco-passage.ance.bf.tsv \
  --output-format msmarco \
  --batch-size 36 --threads 12

The option --encoded-queries specifies the use of encoded queries (i.e., queries that have already been converted into dense vectors and cached). As an alternative, replace with --encoder castorini/ance-msmarco-passage to perform "on-the-fly" query encoding, i.e., convert text queries into dense vectors as part of the dense retrieval process.

To evaluate:

$ python -m pyserini.eval.msmarco_passage_eval msmarco-passage-dev-subset \
    runs/run.msmarco-passage.ance.bf.tsv

#####################
MRR @10: 0.3302
QueriesRanked: 6980
#####################

We can also use the official TREC evaluation tool trec_eval to compute other metrics than MRR@10. For that we first need to convert runs and qrels files to the TREC format:

$ python -m pyserini.eval.convert_msmarco_run_to_trec_run \
    --input runs/run.msmarco-passage.ance.bf.tsv \
    --output runs/run.msmarco-passage.ance.bf.trec

$ python -m pyserini.eval.trec_eval -c -mrecall.1000 -mmap msmarco-passage-dev-subset \
    runs/run.msmarco-passage.ance.bf.trec

map                   	all	0.3362
recall_1000           	all	0.9584

MS MARCO Document

ANCE retrieval with brute-force index:

python -m pyserini.search.faiss \
  --index msmarco-v1-doc.ance-maxp \
  --topics msmarco-doc-dev \
  --encoded-queries ance_maxp-msmarco-doc-dev \
  --output runs/run.msmarco-doc.passage.ance-maxp.txt \
  --output-format msmarco \
  --batch-size 36 --threads 12 \
  --hits 1000 --max-passage --max-passage-hits 100

Same as above, replace --encoded-queries with --encoder castorini/ance-msmarco-doc-maxp for on-the-fly query encoding.

To evaluate:

$ python -m pyserini.eval.msmarco_doc_eval \
    --judgments msmarco-doc-dev \
    --run runs/run.msmarco-doc.passage.ance-maxp.txt

#####################
MRR @100: 0.3796
QueriesRanked: 5193
#####################

We can also use the official TREC evaluation tool trec_eval to compute other metrics than MRR@100. For that we first need to convert runs and qrels files to the TREC format:

$ python -m pyserini.eval.convert_msmarco_run_to_trec_run \
    --input runs/run.msmarco-doc.passage.ance-maxp.txt \
    --output runs/run.msmarco-doc.passage.ance-maxp.trec

$ python -m pyserini.eval.trec_eval -c -mrecall.100 -mmap msmarco-doc-dev \
    runs/run.msmarco-doc.passage.ance-maxp.trec

map                   	all	0.3796
recall_100            	all	0.9033

Natural Questions (NQ)

ANCE retrieval with brute-force index:

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.ance-multi \
  --topics dpr-nq-test \
  --encoded-queries ance_multi-nq-test \
  --output runs/run.ance.nq-test.multi.bf.trec \
  --batch-size 36 --threads 12

Same as above, replace --encoded-queries with --encoder castorini/ance-dpr-question-multi for on-the-fly query encoding.

To evaluate, first convert the TREC output format to DPR's json format:

$ python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
    --topics dpr-nq-test \
    --index wikipedia-dpr-100w \
    --input runs/run.ance.nq-test.multi.bf.trec \
    --output runs/run.ance.nq-test.multi.bf.json

$ python -m pyserini.eval.evaluate_dpr_retrieval \
    --retrieval runs/run.ance.nq-test.multi.bf.json \
    --topk 20 100

Top20	accuracy: 0.8224
Top100	accuracy: 0.8787

Trivia QA

ANCE retrieval with brute-force index:

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.ance-multi \
  --topics dpr-trivia-test \
  --encoded-queries ance_multi-trivia-test \
  --output runs/run.ance.trivia-test.multi.bf.trec \
  --batch-size 36 --threads 12

Same as above, replace --encoded-queries with --encoder castorini/ance-dpr-question-multi for on-the-fly query encoding.

To evaluate, first convert the TREC output format to DPR's json format:

$ python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
    --topics dpr-trivia-test \
    --index wikipedia-dpr-100w \
    --input runs/run.ance.trivia-test.multi.bf.trec \
    --output runs/run.ance.trivia-test.multi.bf.json

$ python -m pyserini.eval.evaluate_dpr_retrieval \
    --retrieval runs/run.ance.trivia-test.multi.bf.json \
    --topk 20 100

Top20	accuracy: 0.8010
Top100	accuracy: 0.8522

Reproduction Log*