This page describes how to replicate the doc2query document expansion experiments in the following paper:
- Rodrigo Nogueira, Wei Yang, Jimmy Lin, Kyunghyun Cho. Document Expansion by Query Prediction. arxiv:1904.08375
The basic idea is to train a model, that when given an input document, generates questions that the document might answer (or more broadly, queries for which the document might be relevant). These predicted questions (or queries) are then appended to the original documents, which are then indexed as before.
For a complete "from scratch" replication (in particularly, training the seq2seq model), see this code repo. Here, we run through how to replicate the BM25+Doc2query condition with our copy of the predicted queries.
Note that docTTTTTquery is an improved version of the doc2query model.
To replicate our Doc2query results on the MS MARCO Passage Ranking Task, follow these instructions. Before going through this guide, it is recommended that you replicate our BM25 baselines first.
To start, grab the predicted queries:
wget https://www.dropbox.com/s/709q495d9hohcmh/pred-test_topk10.tar.gz -P collections/msmarco-passage
tar -xzvf collections/msmarco-passage/pred-test_topk10.tar.gz -C collections/msmarco-passage
To confirm, pred-test_topk10.tar.gz
should have an MD5 checksum of 241608d4d12a0bc595bed2aff0f56ea3
.
Check out the file:
$ wc collections/msmarco-passage/pred-test_topk10.txt
8841823 536425855 2962345659 collections/msmarco-passage/pred-test_topk10.txt
These are the predicted queries based on our seq2seq model, based on top k sampling with 10 samples for each document in the corpus. There are as many lines in the above file as there are documents; all 10 predicted queries are concatenated on a single line.
Now let's create a new document collection by concatenating the predicted queries to the original documents:
python src/main/python/msmarco/augment_collection_with_predictions.py \
--collection_path collections/msmarco-passage/collection.tsv --output_folder collections/msmarco-passage/collection_jsonl_expanded_topk10 \
--predictions collections/msmarco-passage/pred-test_topk10.txt --stride 1
We can then reindex the collection:
sh ./target/appassembler/bin/IndexCollection -collection JsonCollection \
-generator DefaultLuceneDocumentGenerator -threads 9 -input collections/msmarco-passage/collection_jsonl_expanded_topk10 \
-index indexes/msmarco-passage/lucene-index-msmarco-expanded-topk10 -storePositions -storeDocvectors -storeRaw
And run retrieval (same as above):
python ./src/main/python/msmarco/retrieve.py --hits 1000 --index indexes/msmarco-passage/lucene-index-msmarco-expanded-topk10 \
--qid_queries collections/msmarco-passage/queries.dev.small.tsv --output runs/run.msmarco-passage.dev.small.expanded-topk10.tsv
Alternatively, we can run the same script implemented in Java, which is a bit faster:
./target/appassembler/bin/SearchMsmarco -hits 1000 -threads 1 \
-index indexes/msmarco-passage/lucene-index-msmarco-expanded-topk10 -qid_queries collections/msmarco-passage/queries.dev.small.tsv \
-output runs/run.msmarco-passage.dev.small.expanded-topk10.tsv
Finally, to evaluate:
python ./src/main/python/msmarco/msmarco_eval.py \
collections/msmarco-passage/qrels.dev.small.tsv runs/run.msmarco-passage.dev.small.expanded-topk10.tsv
The output should be:
#####################
MRR @10: 0.2213412471005586
QueriesRanked: 6980
#####################
Note that these figures are slightly higher than the values reported in our arXiv paper (0.218) due to BM25 parameter tuning (see above) and an upgrade from Lucene 7.6 to Lucene 8.0 (experiments in the paper were run with Lucene 7.6).
One additional trick not explored in our arXiv paper is to weight the original document and predicted queries differently.
The augment_collection_with_predictions.py
script provides an option --original_copies
that duplicates the original text n times, which is an easy way to weight the original document by n.
For example --original_copies 2
would yield the following results:
#####################
MRR @10: 0.2287041774685029
QueriesRanked: 6980
#####################
So, this simple trick improves MRR by a bit over baseline Doc2query.
We will now describe how to reproduce the TREC CAR results of our model BM25+doc2query presented in the paper.
To start, download the TREC CAR dataset and the predicted queries:
mkdir collections/trec_car
mkdir indexes/trec_car
wget http://trec-car.cs.unh.edu/datareleases/v2.0/paragraphCorpus.v2.0.tar.xz -P collections/trec_car
wget https://storage.googleapis.com/neuralresearcher_data/doc2query/data/aligned5/pred-test_topk10.tar.gz -P collections/trec_car
tar -xf collections/trec_car/paragraphCorpus.v2.0.tar.xz -C collections/trec_car
tar -xf collections/trec_car/pred-test_topk10.tar.gz -C collections/trec_car
To confirm, paragraphCorpus.v2.0.tar.xz
should have an MD5 checksum of a404e9256d763ddcacc3da1e34de466a
and
pred-test_topk10.tar.gz
should have an MD5 checksum of b9f98b55e6260c64e830b34d80a7afd7
.
These are the predicted queries based on our seq2seq model, based on top k sampling with 10 samples for each document in the corpus. There are as many lines in the above file as there are documents; all 10 predicted queries are concatenated on a single line.
Now let's create a new document collection by concatenating the predicted queries to the original documents:
python src/main/python/trec_car/augment_collection_with_predictions.py \
--collection_path collections/trec_car/paragraphCorpus/dedup.articles-paragraphs.cbor \
--output_folder collections/trec_car/collection_jsonl_expanded_topk10 \
--predictions collections/trec_car/pred-test_topk10.txt --stride 1
This augmentation process might take 2-3 hours.
We can then index the expanded documents:
sh target/appassembler/bin/IndexCollection -collection JsonCollection \
-generator DefaultLuceneDocumentGenerator -threads 30 -input collections/trec_car/collection_jsonl_expanded_topk10 \
-index indexes/trec_car/lucene-index.car17v2.0
And retrieve the test queries:
sh target/appassembler/bin/SearchCollection -topicreader Car \
-index indexes/trec_car/lucene-index.car17v2.0 \
-topics src/main/resources/topics-and-qrels/topics.car17v2.0.benchmarkY1test.txt \
-output runs/run.car17v2.0.bm25.topics.car17v2.0.benchmarkY1test.txt -bm25
Evaluation is performed with trec_eval
:
eval/trec_eval.9.0.4/trec_eval -c -m map -c -m recip_rank \
src/main/resources/topics-and-qrels/qrels.car17v2.0.benchmarkY1test.txt \
runs/run.car17v2.0.bm25.topics.car17v2.0.benchmarkY1test.txt
With the above commands, you should be able to replicate the following results:
map all 0.1807
recip_rank all 0.2750
Note that this MAP is slightly higher than the arXiv paper (0.178) because we used TREC CAR corpus v2.0 in this experiment instead of corpus v1.5 used in the paper.
- Results replicated by @justram on 2019-08-09 (commit
5f098f
) - Results replicated by @ronakice on 2019-08-13 (commit
5b29d16
) - Results replicated by @edwinzhng on 2020-01-08 (commit
5cc923d
) - Results replicated by @HangCui0510 on 2020-04-23 (commit
0ae567d
) - Results replicated by @kelvin-jiang on 2020-05-25 (commit
b6e0367
)