The repo describes experiments with docTTTTTquery (sometimes written as docT5query or doc2query-T5), 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. The docTTTTTquery model 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.

This technique was introduced in November 2019 on MS MARCO passage ranking task. Results on the leaderboard 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:

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!

• Data and Trained Models: MS MARCO Passage Ranking Dataset
• Reproducing MS MARCO Passage Ranking Results with Anserini
• Predicting Queries from Passages: T5 Inference with PyTorch
• Predicting Queries from Passages: T5 Inference with TensorFlow
• Learning a New Prediction Model: T5 Training with TensorFlow
• Reproducing MS MARCO Document Ranking Results with Anserini
• Predicting Queries from Documents: T5 Inference with TensorFlow

The basic docTTTTTquery model is trained on the MS MARCO passage ranking dataset. We make the following data and models 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:

We provide instructions on how to reproduce our docTTTTTquery results for the MS MARCO passage ranking task 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 --recurse-submodules https://github.com/castorini/anserini.git
cd anserini
mvn clean package appassembler:assemble
cd tools/eval && tar xvfz trec_eval.9.0.4.tar.gz && cd trec_eval.9.0.4 && make && cd ../../..
cd tools/eval/ndeval && make && cd ../../..

For the purposes of this of this guide, we'll assume that anserini is cloned as a sub-directory of this repo, i.e., docTTTTTquery/anserini/.

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. The first three files can go into base directory of the repo docTTTTTquery/, but put the zip file in a separate sub-directory docTTTTTquery/passage-predictions. The zip file contains a lot of individual files, so this will keep your directory structure manageable.

Before appending the predicted queries to the passages, we need to concatenate them. The commands below create a file that contains 40 concatenated predictions per line and 8,841,823 lines, one for each passage in the corpus. We concatenate only the first 40 predictions as there is only a tiny gain in MRR@10 when using all 80 predictions (nevertheless, we provide 80 predictions in case researchers want to use this data for other purposes).

cd passage-predictions

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

As a sanity check:

$ wc predicted_queries_topk.txt-1004000
 8841823 2253863941 12517353325 predicted_queries_topk.txt-1004000

Go back to your repo base directory docTTTTTquery/. We can now append the predicted queries to the original MS MARCO passage collection:

tar xvf collection.tar.gz

python convert_msmarco_passage_to_anserini.py \
  --collection_path=collection.tsv \
  --predictions=passage-predictions/predicted_queries_topk.txt-1004000 \
  --output_folder=msmarco-passage-expanded

Now, create an index using Anserini on the expanded passages (we're assuming Anserini is cloned as a sub-directory):

sh anserini/target/appassembler/bin/IndexCollection \
  -collection JsonCollection -generator DefaultLuceneDocumentGenerator \
  -threads 9 -input msmarco-passage-expanded -index lucene-index-msmarco-passage-expanded

Once the expanded passages are indexed, we can retrieve 1000 passages per query for the MS MARCO dev set:

sh anserini/target/appassembler/bin/SearchMsmarco \
  -index lucene-index-msmarco-passage-expanded -queries queries.dev.small.tsv \
  -output run.msmarco-passage-expanded.dev.small.txt -hits 1000 -threads 8

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

python anserini/tools/eval/msmarco_eval.py qrels.dev.small.tsv run.msmarco-passage-expanded.dev.small.txt

The results should be:

#####################
MRR @10: 0.27680089370991834
QueriesRanked: 6980
#####################

Voilà!

We will use the excellent 🤗 Transformers library by Hugging Face to sample queries from our T5 model.

First, install the library:

pip install transformers

Then load the model checkpoint:

import torch
from transformers import T5Tokenizer, T5ForConditionalGeneration

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

tokenizer = T5Tokenizer.from_pretrained('castorini/doc2query-t5-base-msmarco')
model = T5ForConditionalGeneration.from_pretrained('castorini/doc2query-t5-base-msmarco')
model.to(device)

Sample 3 questions from a example document:

doc_text = 'The presence of communication amid scientific minds was equally important to the success of the Manhattan Project as scientific intellect was. The only cloud hanging over the impressive achievement of the atomic researchers and engineers is what their success truly meant; hundreds of thousands of innocent lives obliterated.'

input_ids = tokenizer.encode(doc_text, return_tensors='pt').to(device)
outputs = model.generate(
    input_ids=input_ids,
    max_length=64,
    do_sample=True,
    top_k=10,
    num_return_sequences=3)

for i in range(3):
    print(f'sample {i + 1}: {tokenizer.decode(outputs[i], skip_special_tokens=True)}')

The output should be similar to this:

sample 1: why was the manhattan project successful
sample 2: the manhattan project what it means
sample 3: what was the most important aspect of the manhattan project

For more information on how to use T5 with HuggingFace's transformers library, check their documentation.

Next, we provide instructions on how to use our trained T5 models to predict queries for each of the 8.8M documents in the MS MARCO corpus. To speed up inference, we will use 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://git.uwaterloo.ca/jimmylin/doc2query-data/raw/master/T5-passage/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 output_filename with a new filename for each sample).

for ITER in {00..08}; 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="infer_checkpoint_step = 1004000" \
      --gin_param="utils.run.sequence_length = {'inputs': 512, 'targets': 64}" \
      --gin_param="Bitransformer.decode.max_decode_length = 64" \
      --gin_param="input_filename = 'gs://your_bucket/data/input_docs.txt$ITER'" \
      --gin_param="output_filename = 'gs://your_bucket/data/predicted_queries_topk_sample.txt$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.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" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology ='v3-8'

Here we detail how to reproduce docTTTTTquery runs for the MS MARCO document ranking task. The MS MARCO document ranking task is similar to the MS MARCO passage ranking task, but the corpus contains longer documents, which need to be split into shorter passages before being fed to docTTTTTquery.

Like the instructions for MS MARCO passage ranking task, we explain the process in reverse order (i.e., indexing, expansion, query prediction), since we believe there are more users interested in experimenting with the expanded index than expanding the document themselves.

Here are the relevant files to download:

The most straightforward way to use docTTTTTquery is to append the expanded queries to each document. First, download the original corpus (msmarco-docs.tsv.gz), the predicted queries (predicted_queries_doc.tar.gz), and a file mapping passages to their document ids (msmarco_doc_passage_ids.txt), using one of the options above. Put predicted_queries_doc.tar.gz in a sub-directory doc-predictions/.

Merge the predicted queries into a single file; there are 10 predicted queries per document. This can be accomplished as follows:

cd doc-predictions/

tar xvfz predicted_queries_doc.tar.gz

for i in $(seq -f "%03g" 0 9); do
    cat predicted_queries_doc_sample${i}.txt???-1004000 > predicted_queries_doc_sample${i}_all.txt
done

paste -d" " \
  predicted_queries_doc_sample000_all.txt \
  predicted_queries_doc_sample001_all.txt \
  predicted_queries_doc_sample002_all.txt \
  predicted_queries_doc_sample003_all.txt \
  predicted_queries_doc_sample004_all.txt \
  predicted_queries_doc_sample005_all.txt \
  predicted_queries_doc_sample006_all.txt \
  predicted_queries_doc_sample007_all.txt \
  predicted_queries_doc_sample008_all.txt \
  predicted_queries_doc_sample009_all.txt \
   > predicted_queries_doc_sample_all.txt

Sanity check:

$ md5sum predicted_queries_doc_sample_all.txt 
b01b2fbbb8d382684a80fbf51efbca93  predicted_queries_doc_sample_all.txt
$ wc predicted_queries_doc_sample_all.txt 
  20545677 1379262573 7672087649 predicted_queries_doc_sample_all.txt

We now append the queries to the original documents (this step takes approximately 10 minutes, the counter needs to get to 20545677):

python convert_msmarco_doc_to_anserini.py \
  --original_docs_path=msmarco-docs.tsv.gz \
  --doc_ids_path=msmarco_doc_passage_ids.txt \
  --predictions_path=doc-predictions/predicted_queries_doc_sample_all.txt \
  --output_docs_path=msmarco-doc-expanded/docs.json

Once we have the expanded documents (about 29 GB in size), the next step is to build an index with Anserini. As above, we'll assume that Anserini is cloned as a sub-directory of this repo, i.e., docTTTTTquery/anserini/. This step takes approximately 40 minutes:

sh anserini/target/appassembler/bin/IndexCollection -collection JsonCollection \
  -generator DefaultLuceneDocumentGenerator -threads 1 \
  -input msmarco-doc-expanded -index lucene-index-msmarco-doc-expanded

We can then retrieve the documents using the dev queries (this step takes approximately 10 minutes):

sh anserini/target/appassembler/bin/SearchCollection \
  -index lucene-index-msmarco-doc-expanded \
  -topicreader TsvString -topics anserini/src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -output run.msmarco-doc-expanded.dev.small.txt -bm25

And evaluate using trec_eval tool:

anserini/tools/eval/trec_eval.9.0.4/trec_eval -m map -m recall.1000 \
  anserini/src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt run.msmarco-doc-expanded.dev.small.txt

The output should be:

map                   	all	0.2886
recall_1000           	all	0.9259

In comparison, indexing with the original documents gives:

map                     all     0.2310
recall_1000             all     0.8856

Although per-document expansion is the most straightforward way to use docTTTTTquery, we have found that per passage expansion works even better. In this approach, we split the documents into passages and append the expanded queries to each passage. We then index the passages of this expanded corpus.

We will reuse the file predicted_queries_doc_sample_all.txt that contains all the predicted queries from last section. To start, append the queries to the passages:

python convert_msmarco_passages_doc_to_anserini.py \
  --original_docs_path=msmarco-docs.tsv.gz \
  --doc_ids_path=msmarco_doc_passage_ids.txt \
  --predictions_path=doc-predictions/predicted_queries_doc_sample_all.txt \
  --output_docs_path=msmarco-doc-expanded-passage/docs.json

This step takes several hours (the counter needs to get to 20545677). Upon completion, index the passages with Anserini:

sh anserini/target/appassembler/bin/IndexCollection -collection JsonCollection \
  -generator DefaultLuceneDocumentGenerator -threads 1 \
  -input msmarco-doc-expanded-passage -index lucene-index-msmarco-doc-expanded-passage

Then, we can retrieve the top 1k passages with dev queries:

sh anserini/target/appassembler/bin/SearchCollection \
  -index lucene-index-msmarco-doc-expanded-passage \
  -topicreader TsvString -topics anserini/src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -output run.msmarco-doc-expanded-passage.dev.small.txt \
  -bm25 -hits 10000 -selectMaxPassage -selectMaxPassage.delimiter "#" -selectMaxPassage.hits 1000

In a bit more detail, we retrieve the top 10k passages per query, but then use Anserini's -selectMaxPassage option to select only the best (highest-scoring) passage from each document, finally returning top 1k docid per query.

Evaluation:

anserini/tools/eval/trec_eval.9.0.4/trec_eval -m map -m recall.1000 \
  anserini/src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt \
  run.msmarco-doc-expanded-passage.dev.small.txt

The output should be:

map                   	all	0.3182
recall_1000           	all	0.9490

In comparison with per-passage expansion, we will use per passage without expansion as the baseline. In this method, we will not append the predicted queries to the passages.

We will first split the original documents into passages:

python convert_msmarco_passages_doc_to_anserini.py \
  --original_docs_path=msmarco-docs.tsv.gz \
  --doc_ids_path=msmarco_doc_passage_ids.txt \
  --output_docs_path=msmarco-doc-passage/docs.json \

It will also take several hours, and the generated file will be 27G. Same as what we did for per-passage expansion, we will use Anserini to index the file, retrieve the top1k passages from them for the dev queries and evaluate them.

sh anserini/target/appassembler/bin/IndexCollection -collection JsonCollection \
  -generator DefaultLuceneDocumentGenerator -threads 1 \
  -input msmarco-doc-passage -index lucene-index-msmarco-doc-passage

sh anserini/target/appassembler/bin/SearchCollection \
  -index lucene-index-msmarco-doc-passage \
  -topicreader TsvString -topics anserini/src/main/resources/topics-and-qrels/topics.msmarco-doc.dev.txt \
  -output run.msmarco-doc-passage.dev.small.txt \
  -bm25 -hits 10000 -selectMaxPassage -selectMaxPassage.delimiter "#" -selectMaxPassage.hits 1000

anserini/tools/eval/trec_eval.9.0.4/trec_eval -m map -m recall.1000 \
  anserini/src/main/resources/topics-and-qrels/qrels.msmarco-doc.dev.txt \
  run.msmarco-doc-passage.dev.small.txt

The result is:

map                   	all	0.2688
recall_1000           	all	0.9180

If you want to predict the queries yourself, please follow the instructions below.

We begin by downloading the corpus, which contains 3.2M documents.

wget http://msmarco.blob.core.windows.net/msmarcoranking/msmarco-docs.tsv.gz
gunzip msmarco-docs.tsv.gz

We split the corpus into files of 100k documents, which later can be processed in parallel.

split --suffix-length 2 --numeric-suffixes --lines 100000 msmarco-docs.tsv msmarco-docs.tsv

We now segment each document using a sliding window of 10 sentences and stride of 5 sentences:

for ITER in {00..32}; do
    python convert_msmarco_doc_to_t5_format.py \
        --corpus_path=msmarco-docs.tsv$ITER \
        --output_passage_texts_path=${OUTPUT_DIR}/passage_texts.txt$ITER \
        --output_passage_doc_ids_path=${OUTPUT_DIR}/msmarco_doc_passage_ids.txt$ITER
done

Note that we use spacy 2.1.6 to do so. Other versions generate different segments, which change retrieval results.

We are now ready to run inference. Since this is a costly step, we recommend using Google Cloud with TPUs to run it faster.

We will use the docTTTTTquery model trained on the MS MARCO passage ranking dataset, so you need to upload it to your Google Storage bucket.

wget https://git.uwaterloo.ca/jimmylin/doc2query-data/raw/master/T5-passage/t5-base.zip
unzip t5-base.zip
gsutil cp model.ckpt-1004000* gs://your_bucket/models/

Run the command below to sample one question per passage (note that you will need to start a TPU).

for ITER in {00..32}; 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="infer_checkpoint_step = 1004000" \
      --gin_param="utils.run.sequence_length = {'inputs': 512, 'targets': 64}" \
      --gin_param="Bitransformer.decode.max_decode_length = 64" \
      --gin_param="input_filename = './passage_texts.txt$ITER'" \
      --gin_param="output_filename = './predicted_queries_topk_sample.txt$ITER'" \
      --gin_param="tokens_per_batch = 131072" \
      --gin_param="Bitransformer.decode.temperature = 1.0" \
      --gin_param="Unitransformer.sample_autoregressive.sampling_keep_top_k = 10" \
      --gin_param="utils.tpu_mesh_shape.tpu_topology ='v3-8'
done

Here we provide instructions on how to reproduce our docTTTTTquery results for the MS MARCO V2 passage ranking task with the Anserini IR toolkit, using predicted queries. We opensource the predicted queries using the 🤗 Datasets library. Note that this is a very large dataset, so we ran the docTTTTTquery inference step across multiple TPUs. In fact, there is a signficant blow-up in the dataset size compared to MS MARCO v1, because of which we choose to only generate 20 queries per passage. Also, we use a different docTTTTTquery model trained on the MS MARCO v2 passage ranking dataset.

We use the metadata-augmented passage corpus which was shown to have better effectiveness.

First, we download the expanded queries dataset and expand this corpus using NUM_QUERIES queries per passage:

export NUM_QUERIES=20
python3 msmarco-v2/augment_corpus.py --hgf_d2q_dataset castorini/msmarco_v2_passage_doc2query-t5_expansions \
        --original_psg_path collections/msmarco_v2_passage_augmented \
        --output_psg_path collections/msmarco_v2_passage_augmented_d2q-t5_${NUM_QUERIES} \
        --num_workers 70 \
        --num_queries ${NUM_QUERIES} \
        --task passage \
        --cache_dir /path/to/cache/dir

The dataset is downloaded and processed in the cache directory after which the corpus is expanded too. So make sure you have enough storage space (around 300 GB for this entire task). If the dataset is not already cached, this script would take about 18 hours. If it is, you can expect it to finish in about 10 hours.

Upon completion, index the expanded passages with Anserini:

sh target/appassembler/bin/IndexCollection -collection MsMarcoV2PassageCollection \
 -generator DefaultLuceneDocumentGenerator -threads 70 \
 -input collections/msmarco_v2_passage_augmented_d2q-t5_${NUM_QUERIES} \
 -index indexes/msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES} \
 -optimize

Note that this index does not store any "extras" (positions, document vectors, raw documents, etc.) because we don't need any of these for BM25 retrieval.

Finally, we can perform runs on the dev queries (both sets):

target/appassembler/bin/SearchCollection -index indexes/msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES} \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-v2-passage.dev.txt \
 -output runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev.txt -bm25 -hits 1000

target/appassembler/bin/SearchCollection -index indexes/msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES} \
 -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-v2-passage.dev2.txt \
 -output runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev2.txt -bm25 -hits 1000

Evaluation:

$ tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m map -m recip_rank src/main/resources/topics-and-qrels/qrels.msmarco-v2-passage.dev.txt runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev.txt
map                     all     0.1160
recip_rank              all     0.1172

$ tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.100,1000 src/main/resources/topics-and-qrels/qrels.msmarco-v2-passage.dev.txt runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev.txt
recall_100              all     0.5039
recall_1000             all     0.7647

$ tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m map -m recip_rank src/main/resources/topics-and-qrels/qrels.msmarco-v2-passage.dev2.txt runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev2.txt
map                     all     0.1158
recip_rank              all     0.1170

$ tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.100,1000 src/main/resources/topics-and-qrels/qrels.msmarco-v2-passage.dev2.txt runs/run.msmarco-v2-passage-augmented-d2q-t5-${NUM_QUERIES}.dev2.txt
recall_100              all     0.5158
recall_1000             all     0.7659

This guide provide sinstructions on how to reproduce our docTTTTTquery results for the MS MARCO V2 document ranking task with the Anserini IR toolkit, using predicted queries. We opensource the predicted queries using the 🤗 Datasets library. Note that this is a very large dataset, so we ran the docTTTTTquery inference step across multiple TPUs. Also, we use a different docTTTTTquery model trained on the MS MARCO v2 passage ranking dataset.

We use the segmented document corpus which was shown to have better effectiveness.

First, we download the expanded queries dataset and expand this corpus using NUM_QUERIES queries per passage:

export NUM_QUERIES=10
python3 msmarco-v2/augment_corpus.py --hgf_d2q_dataset castorini/msmarco_v2_doc_segmented_doc2query-t5_expansions \
        --original_psg_path collections/msmarco_v2_doc_segmented \
        --output_psg_path collections/msmarco_v2_doc_segmented_d2q-t5_${NUM_QUERIES} \
        --num_workers 60 \
        --num_queries ${NUM_QUERIES} \
        --task segment \
        --cache_dir /path/to/cache/dir

The dataset is downloaded and processed in the cache directory after which the corpus is expanded too. So make sure you have enough storage space (around 300 GB for this entire task). If the dataset is not already cached, this script would take about 18 hours. If it is, you can expect it to finish in about 10 hours.

Upon completion, index the expanded document segments with Anserini:

sh target/appassembler/bin/IndexCollection -collection MsMarcoV2DocCollection \
 -generator DefaultLuceneDocumentGenerator -threads 60 \
 -input collections/msmarco_v2_doc_segmented_d2q-t5_${NUM_QUERIES} \
 -index indexes/msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES} \
 -optimize

Note that this index does not store any "extras" (positions, document vectors, raw documents, etc.) because we don't need any of these for BM25 retrieval.

Finally, we can perform runs on the dev queries (both sets):

target/appassembler/bin/SearchCollection -index /store/scratch/rpradeep/msmarco-v2/indexes/msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES} \
  -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-v2-doc.dev.txt \
  -output runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev.txt \
  -bm25 -hits 10000 -selectMaxPassage -selectMaxPassage.delimiter "#" -selectMaxPassage.hits 1000

target/appassembler/bin/SearchCollection -index /store/scratch/rpradeep/msmarco-v2/indexes/msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES} \
  -topicreader TsvInt -topics src/main/resources/topics-and-qrels/topics.msmarco-v2-doc.dev2.txt \
  -output runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev2.txt \
  -bm25 -hits 10000 -selectMaxPassage -selectMaxPassage.delimiter "#" -selectMaxPassage.hits 1000

Evaluation:

$ tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m map -m recip_rank src/main/resources/topics-and-qrels/qrels.msmarco-v2-doc.dev.txt runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev.txt
map                     all     0.2203
recip_rank              all     0.2226

$ tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.100,1000 src/main/resources/topics-and-qrels/qrels.msmarco-v2-doc.dev.txt runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev.txt
recall_100              all     0.7297
recall_1000             all     0.8982

$ tools/eval/trec_eval.9.0.4/trec_eval -c -M 100 -m map -m recip_rank src/main/resources/topics-and-qrels/qrels.msmarco-v2-doc.dev2.txt runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev2.txt
map                     all     0.2205
recip_rank              all     0.2234

$ tools/eval/trec_eval.9.0.4/trec_eval -c -m recall.100,1000 src/main/resources/topics-and-qrels/qrels.msmarco-v2-doc.dev2.txt runs/run.msmarco-v2-doc-segmented-d2q-t5-${NUM_QUERIES}.dev2.txt
recall_100              all     0.7316
recall_1000             all     0.8952