Skip to content

Latest commit

 

History

History
75 lines (54 loc) · 3.62 KB

regressions-fever.md

File metadata and controls

75 lines (54 loc) · 3.62 KB
Raw

Anserini Regressions: FEVER Fact Verification

This page documents BM25 regression experiments for the FEVER fact verification task, which is integrated into Anserini's regression testing framework.

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.

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 fever

Indexing

Typical indexing command:

bin/run.sh io.anserini.index.IndexCollection \
  -threads 1 \
  -collection FeverParagraphCollection \
  -input /path/to/fever \
  -generator DefaultLuceneDocumentGenerator \
  -index indexes/lucene-index.fever-paragraph/ \
  -storePositions -storeDocvectors -storeRaw \
  >& logs/log.fever &

The directory /path/to/fever should be a directory containing the expanded document collection; see this page for how to prepare this collection.

For additional details, see explanation of common indexing options.

Retrieval

Topics and qrels are stored here, which is linked to the Anserini repo as a submodule. The regression experiments here evaluate on the 9999 claims as part of the dev set for the original FEVER paper. The original data can be found here.

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

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-index.fever-paragraph/ \
  -topics tools/topics-and-qrels/topics.fever.dev.txt \
  -topicReader TsvInt \
  -output runs/run.fever.bm25-default.topics.fever.dev.txt \
  -bm25 &

bin/run.sh io.anserini.search.SearchCollection \
  -index indexes/lucene-index.fever-paragraph/ \
  -topics tools/topics-and-qrels/topics.fever.dev.txt \
  -topicReader TsvInt \
  -output runs/run.fever.bm25-tuned.topics.fever.dev.txt \
  -bm25 -bm25.k1 0.9 -bm25.b 0.1 &

Evaluation can be performed using trec_eval:

bin/trec_eval -c -m recall.100 -c -m recall.1000 tools/topics-and-qrels/qrels.fever.dev.txt runs/run.fever.bm25-default.topics.fever.dev.txt

bin/trec_eval -c -m recall.100 -c -m recall.1000 tools/topics-and-qrels/qrels.fever.dev.txt runs/run.fever.bm25-tuned.topics.fever.dev.txt

Effectiveness

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

R@100 BM25 (Default) BM25 (Tuned)
FEVER Paper Development Dataset 0.8974 0.8988
R@1000 BM25 (Default) BM25 (Tuned)
FEVER Paper Development Dataset 0.9477 0.9481

The setting "default" refers the default BM25 settings of k1=0.9, b=0.4, while "tuned" refers to the tuned setting of k1=0.9, b=0.1.