Sean MacAvaney, Andrew Yates, Arman Cohan, Nazli Goharian. CEDR: Contextualized Embeddings for Document Ranking. SIGIR 2019 (short).

Paper at: https://arxiv.org/abs/1904.07094

tl;dr We demonstrate the effectiveness of using BERT classification for document ranking ("Vanilla BERT") and show that BERT embeddings can be used by prior neural ranking architectures to further improve ranking performance ("CEDR-* models").

If you use this work, please cite as: bibtex

@InProceedings{macavaney2019ContextualWR,
  author = {MacAvaney, Sean and Yates, Andrew and Cohan, Arman and Goharian, Nazli},
  title = {CEDR: Contextualized Embeddings for Document Ranking},
  booktitle = {SIGIR},
  year = {2019}
}

This code is tested on Python 3.6. Install dependencies using the following command:

pip install -r requirements.txt

You will need to prepare files for training and evaluation. Many of these files are available in data/wt (TREC WebTrack) and data/robust (TREC Robust 2004).

qrels: a standard TREC-style query relevant file. Used for identifying relevant items for training pair generation and for validation (data/wt/qrels, data/robust/qrels).

train_pairs: a tab-deliminted file containing pairs used for training. The training process will only use query-document pairs found in this file. Samples are in data/{wt,robust}/*.pairs. File format:

[query_id]	[doc_id]

valid_run: a standard TREC-style run file for re-ranking items for validation. The .run files used for re-ranking are available in data/{wt,robust}/*.run. Note that these runs are using the default parameters, so they do not match the tuned results shown in Table 1.

datafiles: Files containing the text of queries and documents needed for training, validation, or testing. Should be in tab-delimited format as follows, where [type] is either query or doc, [id] is the identifer of the query or document (e.g., 132, clueweb12-0206wb-59-32292), and [text] is the textual content of the query or document (no tabs or newline characters, tokenization done by BertTokenizer).

[type]  [id]  [text]

Queries for WebTrack and Robust are available in data/wt/queries.tsv and data/robust/queries.tsv. Document text can be extracted from an index using extract_docs_from_index.py (be sure to use an index that has appropriate pre-processing). The script supports both Indri and Lucene (via Anserini) indices. See instructions below for help installing pyndri or Anserini.

Examples:

# Indri index
awk '{print $3}' data/robust/*.run | python extract_docs_from_index.py indri PATH_TO_INDRI_INDEX > data/robust/documents.tsv
# Lucene index (should be built with Anserini and the -storeTransformedDocs)
awk '{print $3}' data/robust/*.run | python extract_docs_from_index.py lucene PATH_TO_LUCENE_INDEX > data/robust/documents.tsv

To train a Vanilla BERT model, use the following command:

python train.py \
  --model vanilla_bert \
  --datafiles data/queries.tsv data/documents.tsv \
  --qrels data/qrels \
  --train_pairs data/train_pairs \
  --valid_run data/valid_run \
  --model_out_dir models/vbert

You can see the performance of Vanilla BERT by re-ranking a test run:

python rerank.py \
  --model vanilla_bert \
  --datafiles data/queries.tsv data/documents.tsv \
  --run data/test_run \
  --model_weights models/vbert/weights.p \
  --out_path models/vbert/test.run

To train a CEDR model, first train a Vanilla BERT model, and then use the following command:

python train.py \
  --model cedr_pacrr \ # or cedr_knrm / cedr_drmm
  --datafiles data/queries.tsv data/documents.tsv \
  --qrels data/qrels \
  --train_pairs data/train_pairs \
  --valid_run data/valid_run \
  --initial_bert_weights models/vbert/weights.p \
  --model_out_dir models/cedrpacrr

You can see the performance of CEDR by re-ranking a test run:

python rerank.py \
  --model cedr_pacrr \ # or cedr_knrm / cedr_drmm
  --datafiles data/queries.tsv data/documents.tsv \
  --run data/test_run \
  --model_weights models/cedrpacrr/weights.p \
  --out_path models/cedrpacrr/test.run

Note that this will calculate results using bin/trec_eval with P@20, whereas the nDCG@20 and ERR@20 results in Table 1 are calculated using bin/gdeval.pl.

These instructions are only needed if using the extract_docs_from_index.py script, and depend on the index from which you are extracting documents.

Here's what worked for me. Please refer to cvangysel/pyndri for futher assistance installing pyndri.

wget https://sourceforge.net/projects/lemur/files/lemur/indri-5.14/indri-5.14.tar.gz
tar xvfz indri-5.14.tar.gz
cd indri-5.14
./configure CXX="g++ -D_GLIBCXX_USE_CXX11_ABI=0"
make
sudo make install
pip install pyndri==0.4

Install pyjnius (refer to kivy/pyjnius for futher assistance with pyjnius.)

pip install pyjnius==1.1.4

Build Anserini (refer to castorini/anserini for further assistance with Anserini.)

wget https://github.com/castorini/anserini/archive/anserini-0.4.0.tar.gz
tar -xzvf anserini-0.4.0.tar.gz
cd anserini-anserini-0.4.0/
mvn clean package appassembler:assemble
mv target/anserini-0.4.0-fatjar.jar ~/cedr/bin/anserini.jar