Skip to content

Latest commit

 

History

History
148 lines (113 loc) · 6.24 KB

experiments-dkrr.md

File metadata and controls

148 lines (113 loc) · 6.24 KB
Raw

Pyserini: Reproducing DKRR Results

DKRR (Distilling Knowledge from Reader to Retriever) is a technique to learn retriever models described in the following paper:

Gautier Izacard and Edouard Grave. Distilling Knowledge from Reader to Retriever for Question Answering. arXiv:2012.04584, 2020.

We have incorporated this work into Pyserini. In particular, we have taken the pretrained nq_retriever and tqa_retriever models from the DKRR repo, used them to index English Wikipedia, and then incorporate into the dense retrieval framework in Pyserini.

This guide provides instructions to reproduce our results.

Natural Questions

Running DKRR retrieval on dpr-nq-dev and nq-test of the Natural Questions dataset:

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.dkrr-nq \
  --topics dpr-nq-dev \
  --encoded-queries dkrr-dpr-nq-retriever-dpr-nq-dev \
  --output runs/run.dpr-dkrr-nq.dev.trec \
  --query-prefix question: \
  --batch-size 512 --threads 16

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.dkrr-nq \
  --topics nq-test \
  --encoded-queries dkrr-dpr-nq-retriever-nq-test \
  --output runs/run.dpr-dkrr-nq.test.trec \
  --query-prefix question: \
  --batch-size 512 --threads 16

Alternatively, replace --encoded-queries ... with --encoder castorini/dkrr-dpr-nq-retriever for on-the-fly query encoding.

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

python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
  --topics dpr-nq-dev \
  --index wikipedia-dpr \
  --input runs/run.dpr-dkrr-nq.dev.trec \
  --output runs/run.dpr-dkrr-nq.dev.json

python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
  --topics nq-test \
  --index wikipedia-dpr \
  --input runs/run.dpr-dkrr-nq.test.trec \
  --output runs/run.dpr-dkrr-nq.test.json

Evaluating:

python -m pyserini.eval.evaluate_dpr_retrieval \
  --retrieval runs/run.dpr-dkrr-nq.dev.json \
  --topk 5 20 100 500 1000

python -m pyserini.eval.evaluate_dpr_retrieval \
  --retrieval runs/run.dpr-dkrr-nq.test.json \
  --topk 5 20 100 500 1000

The expected results are as follows, shown in the "ours" column:

Metric dpr-nq-dev (ours) dpr-nq-dev (orig) nq-test (ours)
Top-5 72.40 73.80
Top-20 82.36 82.4 84.27
Top-100 87.87 87.9 89.34
Top-500 90.37 92.24
Top-1000 91.30 93.43

For reference, reported results from the paper (Table 8) are shown in the "orig" column.

TriviaQA (TQA)

Running DKRR retrieval on dpr-trivia-dev and dpr-trivia-test of the TriviaQA dataset:

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.dkrr-tqa \
  --topics dpr-trivia-dev \
  --encoded-queries dkrr-dpr-tqa-retriever-dpr-tqa-dev \
  --output runs/run.dpr-dkrr-trivia.dev.trec \
  --query-prefix question: \
  --batch-size 512 --threads 16

python -m pyserini.search.faiss \
  --index wikipedia-dpr-100w.dkrr-tqa \
  --topics dpr-trivia-test \
  --encoded-queries dkrr-dpr-tqa-retriever-dpr-tqa-test \
  --output runs/run.dpr-dkrr-trivia.test.trec \
  --query-prefix question: \
  --batch-size 512 --threads 16

Alternatively, replace --encoded-queries ... with --encoder castorini/dkrr-dpr-tqa-retriever for on-the-fly query encoding.

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

python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
  --topics dpr-trivia-dev \
  --index wikipedia-dpr \
  --input runs/run.dpr-dkrr-trivia.dev.trec \
  --output runs/run.dpr-dkrr-trivia.dev.json

python -m pyserini.eval.convert_trec_run_to_dpr_retrieval_run \
  --topics dpr-trivia-test \
  --index wikipedia-dpr \
  --input runs/run.dpr-dkrr-trivia.test.trec \
  --output runs/run.dpr-dkrr-trivia.test.json

Evaluating:

python -m pyserini.eval.evaluate_dpr_retrieval \
  --retrieval runs/run.dpr-dkrr-trivia.dev.json \
  --topk 5 20 100 500 1000

python -m pyserini.eval.evaluate_dpr_retrieval \
  --retrieval runs/run.dpr-dkrr-trivia.test.json \
  --topk 5 20 100 500 1000

The expected results are as follows, shown in the "ours" column:

Metric dpr-trivia-dev (ours) dpr-trivia-dev (orig) dpr-trivia-test (ours)
Top-5 77.31 77.23
Top-20 83.63 83.5 83.74
Top-100 87.39 87.4 87.78
Top-500 89.77 89.87
Top-1000 90.35 90.63

For reference, reported results from the paper (Table 8) are shown in the "orig" column.

Hybrid sparse-dense retrieval with GAR-T5

Running hybrid sparse-dense retrieval with DKKR and GAR-T5 is detailed in experiments-gar-t5.md

Reproduction Log*