Skip to content
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
132 lines (98 sloc) 5.82 KB

Anserini: BM25 Baselines on MS MARCO (Document)

This page contains instructions for running BM25 baselines on the MS MARCO document ranking task. Note that there is a separate MS MARCO passage ranking task.

Data Prep

We're going to use msmarco-doc/ as the working directory. First, we need to download and extract the MS MARCO document dataset:

mkdir msmarco-doc
mkdir msmarco-doc/collection

wget -P msmarco-doc/collection

To confirm, msmarco-docs.trec.gz should have MD5 checksum of d4863e4f342982b51b9a8fc668b2d0c0.

There's no need to uncompress the file, as Anserini can directly index gzipped files. Build the index with the following command:

nohup sh target/appassembler/bin/IndexCollection -collection TrecCollection \
 -generator LuceneDocumentGenerator -threads 1 -input msmarco-doc/collection \
 -index lucene-index.msmarco-doc.pos+docvectors+rawdocs -storePositions -storeDocvectors -storeRawDocs \
 >& log.msmarco-doc.pos+docvectors+rawdocs &

On a modern desktop with an SSD, indexing takes around 40 minutes. The final log lines should look something like this:

2019-06-09 09:32:35,233 INFO  [main] index.IndexCollection ( - # Final Counter Values
2019-06-09 09:32:35,233 INFO  [main] index.IndexCollection ( - indexed:        3,213,835
2019-06-09 09:32:35,233 INFO  [main] index.IndexCollection ( - empty:                  0
2019-06-09 09:32:35,234 INFO  [main] index.IndexCollection ( - unindexed:              0
2019-06-09 09:32:35,234 INFO  [main] index.IndexCollection ( - unindexable:            0
2019-06-09 09:32:35,234 INFO  [main] index.IndexCollection ( - skipped:                0
2019-06-09 09:32:35,234 INFO  [main] index.IndexCollection ( - errors:                 0
2019-06-09 09:32:35,238 INFO  [main] index.IndexCollection ( - Total 3,213,835 documents indexed in 00:39:07

Retrieving and Evaluating the Dev set

Let's download the queries and qrels:

wget -P msmarco-doc
wget -P msmarco-doc
wget -P msmarco-doc

wget -P msmarco-doc
wget -P msmarco-doc
wget -P msmarco-doc

gunzip msmarco-doc/*.gz

Here are the sizes:

$ wc msmarco-doc/*.tsv
    5193   20772  108276 msmarco-doc/msmarco-docdev-qrels.tsv
    5193   35787  220304 msmarco-doc/msmarco-docdev-queries.tsv
  367013 1468052 7539008 msmarco-doc/msmarco-doctrain-qrels.tsv
  367013 2551279 15480364 msmarco-doc/msmarco-doctrain-queries.tsv
  744412 4075890 23347952 total

There are indeed lots of training queries! In this guide, to save time, we are only going to perform retrieval on the dev queries. This can be accomplished as follows:

target/appassembler/bin/SearchCollection -topicreader Tsv -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topics msmarco-doc/msmarco-docdev-queries.tsv -output msmarco-doc/ -bm25

On a modern desktop with an SSD, the run takes around 12 minutes. After the run completes, we can evaluate with trec_eval:

$ eval/trec_eval.9.0.4/trec_eval -c msmarco-doc/msmarco-docdev-qrels.tsv msmarco-doc/
runid                 	all	Anserini
num_q                 	all	5193
num_ret               	all	5191674
num_rel               	all	5193
num_rel_ret           	all	4599
map                   	all	0.2308

Let's compare to the baselines provided by Microsoft (note that to be fair, we restrict evaluation to top 100 hits per topic):

$ eval/trec_eval.9.0.4/trec_eval -c -mmap -M 100 msmarco-doc/msmarco-docdev-qrels.tsv msmarco-doc/msmarco-docdev-top100
map                   	all	0.2219

$ eval/trec_eval.9.0.4/trec_eval -c -mmap -M 100 msmarco-doc/msmarco-docdev-qrels.tsv msmarco-doc/
map                   	all	0.2301

We see that "out of the box" Anserini is already better!

BM25 Tuning

It is well known that BM25 parameter tuning is important. The above instructions use the Anserini (system-wide) default of k1=0.9, b=0.4.

Let's try to do better! We tuned BM25 using the queries found here: these are five different sets of 10k samples from the training queries (using the shuf command). Tuning was performed on each individual set (grid search, in tenth increments) and then we averaged parameter values across all five sets (this has the effect of regularization). Here, we optimized for average precision (AP). The tuned parameters using this approach are k1=3.44, b=0.87.

To perform a run with these parameters, issue the following command:

target/appassembler/bin/SearchCollection -topicreader Tsv -index lucene-index.msmarco-doc.pos+docvectors+rawdocs \
 -topics msmarco-doc/msmarco-docdev-queries.tsv -output -bm25 -k1 3.44 -b 0.87

Here's the comparison between the Anserini default and tuned parameters:

Setting AP Recall@1000
Default (k1=0.9, b=0.4) 0.2310 0.8856
Tuned (k1=3.44, b=0.87) 0.2788 0.9326

As expected, BM25 tuning makes a big difference!

You can’t perform that action at this time.