Skip to content

experiments-msmarco-irst.md

Latest commit

 

History

History
198 lines (148 loc) · 7.26 KB

experiments-msmarco-irst.md

File metadata and controls

198 lines (148 loc) · 7.26 KB

Pyserini: IRST on MS MARCO V1 Collections

❗ Code associated with these experiments was removed in commit a65b96. This page is preserved only for archival purposes.

This guide describes how to reproduce the IRST (Information Retrieval as Statistical Translation) experiments on the MS MARCO V1 collections, as described in the following paper:

Yuqi Liu, Chengcheng Hu, and Jimmy Lin. Another Look at Information Retrieval as Statistical Translation. Proceedings of the 45th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2022), July 2022.

Below, we discuss passage ranking and two document ranking conditions (full docs and segmented docs).

Passage Ranking

Here, we start directly from our pre-built indexes and already-trained IRST models. The IBM model we use is referenced in Boytsov et al. (2021). For training the model from scratch, consult the guide in FlexNeuART.

The following commands will reproduce the results in Table 1 of our paper:

IRST (Sum)

python -m pyserini.search.lucene.irst \
  --topics msmarco-passage-dev-subset \
  --index msmarco-v1-passage \
  --output runs/run.irst-sum.passage.dev.txt \
  --alpha 0.1

IRST (Max)

python -m pyserini.search.lucene.irst \
  --topics msmarco-passage-dev-subset \
  --index msmarco-v1-passage \
  --output runs/run.irst-max.passage.dev.txt \
  --alpha 0.3 \
  --max-sim

The option --topics specifies the different topics. The choices are:

  • MS MARCO V1 passage dev queries: msmarco-passage-dev-subset (per above)
  • TREC DL 2019 passage: dl19-passage
  • TREC DL 2020 passage: dl20

To evaluate results, use trec_eval. For MS MARCO V1 passage:

python -m pyserini.eval.trec_eval -c -M 10 -m ndcg_cut.10 -m map -m recip_rank \
  msmarco-passage-dev-subset runs/run.irst-sum.passage.dev.txt

For TREC DL 2019, note that we need to specify -l 2:

python -m pyserini.eval.trec_eval -c -m map -m ndcg_cut.10 -l 2 \
  dl19-passage runs/run.irst-sum.passage.dl19.txt

Similarly, for TREC DL 2020:

python -m pyserini.eval.trec_eval -c -m map -m ndcg_cut.10 -l 2 \
  dl20-passage runs/run.irst-sum.passage.dl20.txt

The results should match Table 1 from our paper, repeated below:

MS MARCO Dev TREC 2019 TREC 2020
MRR@10 nDCG@10 MAP nDCG@10 MAP
(1a) BM25 (k1= 0.82, b=0.68) 0.188 0.497 0.290 0.488 0.288
(2a) BM25 + IRST (Sum) 0.221 0.526 0.328 0.558 0.352
(2b) BM25 + IRST (Max) 0.215 0.537 0.329 0.547 0.336

The BM25 baseline is provided for reference.

Document Ranking

In the paper, we explore two different conditions for document ranking: full documents and segmented documents.

For full documents:

IRST (Sum)

python -m pyserini.search.lucene.irst \
  --topics msmarco-doc-dev \
  --index msmarco-v1-doc \
  --output runs/run.irst-sum.doc-full.dev.txt \
  --alpha 0.3 \
  --hits 1000

IRST (Max)

python -m pyserini.search.lucene.irst \
  --topics msmarco-doc-dev \
  --index msmarco-v1-doc \
  --output runs/run.irst-max.doc-full.dev.txt \
  --alpha 0.3 \
  --hits 1000 \
  --max-sim

For segmented documents:

IRST (Sum)

python -m pyserini.search.lucene.irst \
  --topics msmarco-doc-dev \
  --index msmarco-v1-doc-segmented \
  --output runs/run.irst-sum.doc-seg.dev.txt \
  --alpha 0.3 \
  --segments \
  --hits 10000

IRST (Max)

python -m pyserini.search.lucene.irst \
  --topics msmarco-doc-dev \
  --index msmarco-v1-doc-segmented \
  --output runs/run.irst-max.doc-seg.dev.txt \
  --alpha 0.3 \
  --hits 10000 \
  --segments \
  --max-sim

The option --topics specifies the different topics. The choices are:

  • MS MARCO V1 doc dev queries: msmarco-doc-dev (per above)
  • TREC DL 2019 passage: dl19-doc
  • TREC DL 2020 passage: dl20

To evaluate results, use trec_eval. For MS MARCO V1 doc:

python -m pyserini.eval.trec_eval -c -M 100 -m ndcg_cut.10 -m map -m recip_rank \
  msmarco-doc-dev runs/run.irst-sum.doc-full.dev.txt

For TREC DL 2019:

python -m pyserini.eval.trec_eval -c -M 100 -m map -m ndcg_cut.10 \
  dl19-doc runs/run.irst-sum.doc-full.dl19.txt

Similarly, for TREC DL 2020:

python -m pyserini.eval.trec_eval -c -M 100 -m map -m ndcg_cut.10 \
  dl20-doc runs/run.irst-sum.doc-full.dl20.txt

The results should match Table 2 from our paper, repeated below:

MS MARCO Dev TREC 2019 TREC 2020
MRR@100 nDCG@10 MAP nDCG@10 MAP
Document (Full)
(2a) BM25 (k1= 0.82, b=0.68) 0.249 0.510 0.241 0.528 0.378
(2b) BM25 + IRST (Sum) 0.302 0.549 0.252 0.556 0.383
(2c) BM25 + IRST (Max) 0.252 0.491 0.220 0.502 0.337
Document (Segmented)
(3a) BM25 (k1= 0.82, b=0.68) 0.269 0.529 0.240 0.531 0.362
(3b) BM25 + IRST (Sum) 0.296 0.560 0.271 0.534 0.376
(3c) BM25 + IRST (Max) 0.259 0.520 0.243 0.509 0.350

The BM25 baselines are provided for reference.

For the segmented documents collection, the above commands specify --hits 10000, which was the setting used in the SIGIR paper. Obviously, reducing the number of hits considered, e.g., --hits 1000, will speed up running times dramatically, but at the cost of a tiny degradation in effectiveness (in some cases). Many of the differences aren't even noticeable to three digits, so for reference, to contrast these two settings, we report scores to four digits:

MS MARCO Dev TREC 2019 TREC 2020
MRR@100 nDCG@10 MAP nDCG@10 MAP
Document (Segmented)
BM25 + IRST (Sum): --hits 10000 0.2961 0.5596 0.2711 0.5343 0.3759
BM25 + IRST (Max): --hits 10000 0.2589 0.5195 0.2425 0.5089 0.3496
BM25 + IRST (Sum): --hits 1000 0.2936 0.5549 0.2705 0.5343 0.3753
BM25 + IRST (Max): --hits 1000 0.2587 0.5187 0.2432 0.5064 0.3482

Reproduction Log*