docTTTTTquery is the latest version of the doc2query family of document expansion models. 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. docTTTTTquery gets its name from the use of T5 as the expansion model.
The primary advantage of this approach is that expensive neural inference is pushed to indexing time, which means that "bag of words" queries against an inverted index built on the augmented document collection are only slightly slower (due to longer documents) — but the retrieval results are much better. Of course, these documents can be further reranked by another neural model in a multi-stage ranking architecture.
The results on the MS MARCO passage retrieval task show that docTTTTTquery is much more effective than doc2query and (almost) as effective as the best non-BERT ranking model, while increasing query latency (time to retrieve 1000 docs per query) only slightly compared to vanilla BM25:
MS MARCO Passage Ranking Leaderboard (Nov 30th 2019) | Eval MRR@10 | Latency |
---|---|---|
BM25 + BERT from (Nogueira et al., 2019) | 36.8 | 3500 ms |
FastText + Conv-KNRM (Single) (Hofstätter et al. SIGIR 2019) (best non-BERT) | 27.7 | - |
docTTTTTquery (this code) | 27.2 | 64 ms |
DeepCT (Dai and Callan, 2019) | 23.9 | 55 ms |
doc2query (Nogueira et al., 2019) | 21.8 | 61 ms |
BM25 | 18.6 | 55 ms |
For more details, check out our paper:
- Rodrigo Nogueira and Jimmy Lin. From doc2query to docTTTTTquery.
Why's the paper so short? Check out our proposal for micropublications!
We make the following data available for download:
doc_query_pairs.train.tsv
: Approximately 500,000 passage-query pairs used to train the model.queries.dev.small.tsv
: 6,980 queries from the MS MARCO dev set. In this tsv file, the first column is the query id and the second is the query text.qrels.dev.small.tsv
: 7,437 pairs of query relevant passage ids from the MS MARCO dev set. In this tsv file, the first column is the query id and the third column is the passage id. The other two columns (second and fourth) are not used.collection.tar.gz
: All passages (8,841,823) in the MS MARCO passage corpus. In this tsv file, the first column is the passage id and the second is the passage text.predicted_queries_topk_sampling.zip
: 80 predicted queries for each MS MARCO passage, using T5-base and top-k sampling.run.dev.small.tsv
: Approximately 6,980,000 pairs of dev set queries and retrieved passages using the passages expanded with docTTTTTquery + BM25. In this tsv file, the first column is the query id, the second column is the passage id, and the third column is the rank of the passage. There are 1000 passages per query in this file.t5-base.zip
: trained T5 model used for generating the expansions.t5-large.zip
: larger trained T5 model; we didn't find the output to be any better.
Download and verify the above files from the below table:
File | Size | MD5 | Download |
---|---|---|---|
doc_query_pairs.train.tsv |
197 MB | aa673014f93d43837ca4525b9a33422c |
[GCS] [Dropbox] |
queries.dev.small.tsv |
283 KB | 41e980d881317a4a323129d482e9f5e5 |
[GCS] [Dropbox] |
qrels.dev.small.tsv |
140 KB | 38a80559a561707ac2ec0f150ecd1e8a |
[GCS] [Dropbox] |
collection.tar.gz |
987 MB | 87dd01826da3e2ad45447ba5af577628 |
[GCS] [Dropbox] |
predicted_queries_topk_sampling.zip |
7.9 GB | 8bb33ac317e76385d5047322db9b9c34 |
[GCS] [Dropbox] |
run.dev.small.tsv |
133 MB | d6c09a6606a5ed9f1a300c258e1930b2 |
[GCS] [Dropbox] |
t5-base.zip |
357 MB | 881d3ca87c307b3eac05fae855c79014 |
[GCS] [Dropbox] |
t5-large.zip |
1.2 GB | 21c7e625210b0ae872679bc36ed92d44 |
[GCS] [Dropbox] |
We provide instructions on how to replicate our docTTTTTquery runs with the Anserini IR toolkit, using the predicted queries provided above.
First, install Anserini (see homepage for more details):
sudo apt-get install maven
git clone https://github.com/castorini/Anserini.git
cd Anserini
mvn clean package appassembler:assemble
tar xvfz eval/trec_eval.9.0.4.tar.gz -C eval/ && cd eval/trec_eval.9.0.4 && make
cd ../ndeval && make
Next, download queries.dev.small.tsv
, qrels.dev.small.tsv
, collection.tar.gz
, and predicted_queries_topk_sampling.zip
using one of the options above.
Before appending the sampled queries to the passages, we need to concatenate them. The commands below create a file that contains 40 concatenated samples per line and 8,841,823 lines, one for each passage in the corpus. We concatenate only the first 40 samples as there is only a tiny gain in MRR@10 when using 80 samples (nevertheless, we provide 80 samples in case researchers want to use this data for other purposes).
unzip predicted_queries_topk_sampling.zip
for i in $(seq -f "%03g" 0 17); do
echo "Processing chunk $i"
paste -d" " predicted_queries_topk_sample0[0-3]?.txt${i}-1004000 \
> predicted_queries_topk.txt${i}-1004000
done
cat predicted_queries_topk.txt???-1004000 > predicted_queries_topk.txt-1004000
We can now append those queries to the original MS MARCO passage collection:
tar -xvf collection.tar.gz
python convert_collection_to_jsonl.py \
--collection_path=collection.tsv \
--predictions=predicted_queries_topk.txt-1004000 \
--output_folder=./docs
Now, create an index using Anserini on the expanded passages (replace /path/to/anserini/
with actual location of Anserini):
sh /path/to/anserini/target/appassembler/bin/IndexCollection \
-collection JsonCollection -generator LuceneDocumentGenerator \
-threads 9 -input ./docs -index ./lucene-index
Once the expanded passages are indexed, we can retrieve 1000 passages per query for the MS MARCO dev set:
sh /path/to/anserini/target/appassembler/bin/SearchMsmarco \
-index ./lucene-index -qid_queries ./queries.dev.small.tsv \
-output ./run.dev.small.tsv -hits 1000
Finally, we evaluate the results using the MS MARCO eval script:
python /path/to/anserini/src/main/python/msmarco/msmarco_eval.py ./qrels.dev.small.tsv ./run.dev.small.tsv
The results should be:
#####################
MRR @10: 0.2767497271114737
QueriesRanked: 6980
#####################
Voilà!
Next, we provide instructions on how to use our trained T5 models to predict queries from new passages. Note that T5 only works on TPUs (and consequently Google Cloud machines), so this installation must be performed on a Google Cloud instance.
To begin, install T5 (check the original T5 repository for the latest installation instructions):
pip install t5[gcp]
We first need to prepare an input file that contains one passage text per line. We achieve this by extracting the second column of collection.tsv
:
cut -f1 collection.tsv > input_docs.txt
We also need to split the file into smaller files (each with 1M lines) to avoid TensorFlow complaining that proto arrays can only be 2GB at the most:
split --suffix-length 2 --numeric-suffixes --lines 1000000 input_docs.txt input_docs.txt
We now upload the input docs to Google Cloud Storage:
gsutil cp input_docs.txt?? gs://your_bucket/data/
We also need to upload our trained t5-base model to GCS:
wget https://storage.googleapis.com/doctttttquery_git/t5-base.zip
unzip t5-base.zip
gsutil cp model.ckpt-1004000* gs://your_bucket/models/
We are now ready to predict queries from passages. Remember to replace your_tpu
, your_tpu_zone
, your_project_id
and your_bucket
with your values. Note that the command below will only sample one query per passage. If you want multiple samples, you will need to repeat this process multiple times (remember to replace decode_from_file.output_filename
with a new filename for each sample).
for ITER in {00..09}; do
t5_mesh_transformer \
--tpu="your_tpu" \
--gcp_project="your_project_id" \
--tpu_zone="your_tpu_zone" \
--model_dir="gs://your_bucket/models/" \
--gin_file="gs://t5-data/pretrained_models/base/operative_config.gin" \
--gin_file="infer.gin" \
--gin_file="sample_decode.gin" \
--gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
--gin_param="infer_checkpoint_step = 1004000" \
--gin_param="utils.run.sequence_length = {'inputs': 512, 'targets': 64}" \
--gin_param="Bitransformer.decode.max_decode_length = 64" \
--gin_param="decode_from_file.input_filename = 'gs://your_bucket/data/input_docs.txt$ITER'" \
--gin_param="decode_from_file.output_filename = 'gs://your_bucket/data/predicted_queries_topk_sample.txt$ITER'" \
--gin_param="tokens_per_batch = 131072" \
--gin_param="Bitransformer.decode.temperature = 1.0" \
--gin_param="Unitransformer.sample_autoregressive.sampling_keep_top_k = 10"
done
It should take approximately 8 hours to sample one query for each of the 8.8M passages, costing ~$20 USD (8 hours at $2.40 USD/hour) on a preemptible TPU.
Finally, we show how to learn a new prediction model.
The following command will train a T5-base model for 4k iterations to predict queries from passages.
We assume you put the tsv training file in gs://your_bucket/data/doc_query_pairs.train.tsv
(download from above).
Also, change your_tpu_name
, your_tpu_zone
, your_project_id
, and your_bucket
accordingly.
t5_mesh_transformer \
--tpu="your_tpu_name" \
--gcp_project="your_project_id" \
--tpu_zone="your_tpu_zone" \
--model_dir="gs://your_bucket/models/" \
--gin_param="init_checkpoint = 'gs://t5-data/pretrained_models/base/model.ckpt-999900'" \
--gin_file="dataset.gin" \
--gin_file="models/bi_v1.gin" \
--gin_file="gs://t5-data/pretrained_models/base/operative_config.gin" \
--gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
--gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
--gin_param="utils.run.train_dataset_fn = @t5.models.mesh_transformer.tsv_dataset_fn" \
--gin_param="tsv_dataset_fn.filename = 'gs://your_bucket/data/doc_query_pairs.train.tsv'" \
--gin_file="learning_rate_schedules/constant_0_001.gin" \
--gin_param="run.train_steps = 1004000" \
--gin_param="tokens_per_batch = 131072"