Skip to content
docTTTTTquery document expansion model
Branch: master
Clone or download
Latest commit f58c779 Dec 8, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information. Update Dec 9, 2019 Update Dec 6, 2019


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:

Why's the paper so short? Check out our proposal for micropublications!

Data and Trained Models

We make the following data available for download:

  • doc_query_pairs.train.tsv: Approximately 500,000 passage-query pairs used to train the model.
  • 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.
  • 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.
  • 80 predicted queries for each MS MARCO passage, using T5-base and top-k sampling.
  • 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.
  • trained T5 model used for generating the expansions.
  • 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] 283 KB 41e980d881317a4a323129d482e9f5e5 [GCS] [Dropbox] 140 KB 38a80559a561707ac2ec0f150ecd1e8a [GCS] [Dropbox]
collection.tar.gz 987 MB 87dd01826da3e2ad45447ba5af577628 [GCS] [Dropbox] 7.9 GB 8bb33ac317e76385d5047322db9b9c34 [GCS] [Dropbox] 133 MB d6c09a6606a5ed9f1a300c258e1930b2 [GCS] [Dropbox] 357 MB 881d3ca87c307b3eac05fae855c79014 [GCS] [Dropbox] 1.2 GB 21c7e625210b0ae872679bc36ed92d44 [GCS] [Dropbox]

Replicating Retrieval Results with Anserini

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
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,, collection.tar.gz, and 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).


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

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 \
    --collection_path=collection.tsv \
    --predictions=predicted_queries_topk.txt-1004000 \

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 ./ \
  -output ./ -hits 1000

Finally, we evaluate the results using the MS MARCO eval script:

python /path/to/anserini/src/main/python/msmarco/ ./ ./

The results should be:

MRR @10: 0.2767497271114737
QueriesRanked: 6980


T5 Inference: Predicting Queries from Passages

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:

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=" = {'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"

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.

T5 Training: Learning a New Prediction Model

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=" = @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"
You can’t perform that action at this time.